TTMS Blog
TTMS experts about the IT world, the latest technologies and the solutions we implement.
Posts by: Marcin Kapuściński
Limitations of AI in Legal Software: Risks of Incorrect Advice, Defective Court Filings and Missed Deadlines
An authentic case number supporting a fabricated proposition is one of the most dangerous forms of legal hallucination. In June 2026, Poland’s Supreme Administrative Court described a submission in which counsel cited three judgments that genuinely existed. The problem was that they concerned different legal issues and did not contain the propositions attributed to them. The submission appeared credible until someone checked the original sources. This example illustrates the practical limitations of AI in legal software. An error may affect legal arguments, the assessment of evidence, deadline calculations, advice given to a client or the content of a court filing. The lawyer who approves the advice, opinion or filing remains responsible for its final content. Any potential liability for damages is assessed by reference to the applicable professional standard of care, the scope of the engagement and the circumstances of the individual case. In this article, you will learn about: the most common AI errors in legal work, hallucinations involving legislation, judgments and case numbers, the procedural consequences of using incorrect AI-generated content in a court filing, a lawyer’s liability for advice prepared with the assistance of AI, principles for the safe implementation of AI software in a law firm. For a broader discussion of compliance, confidentiality, providers and risk classification, see AI for lawyers in Europe and the UK: key risks and limitations. Here, we focus on what may go wrong in a specific case governed by Polish law and how to design a process that helps identify errors before legal advice or a court filing is sent. An authentic case number and a fabricated proposition: lessons from the Supreme Administrative Court’s order in case I FZ 104/26 In its order of 23 June 2026 in case I FZ 104/26, the Supreme Administrative Court considered an appeal against an order refusing to stay the enforcement of a tax decision. The appellant was required to substantiate the conditions set out in Article 61 § 3 of the Law on Proceedings before Administrative Courts, particularly the risk of substantial damage or consequences that would be difficult to reverse. Counsel cited three judgments together with propositions purportedly drawn from them. The Court found that the judgments concerned different issues and that their written reasons did not contain the propositions attributed to them. The dates given for those judgments were also incorrect. The Supreme Administrative Court identified another issue: the submission was general in nature and did not contain specific data or documents that would have allowed the Court to assess the party’s financial position. The appeal was dismissed. The Court criticised counsel’s uncritical use of AI and emphasised that a client is entitled to expect professional service. The decision illustrates three practical consequences of a defective workflow: a legal argument may lose the support of the source on which it purports to rely, a submission may omit facts and evidence that are material to the outcome, the court may question the reliability of counsel’s work. A proper verification of a judgment includes checking the court, date, case number and type of decision, followed by reading the full written reasons. The lawyer should assess the factual context, legal basis, significance of the cited passage, available information about the decision’s finality and subsequent case law concerning the same issue. Seven AI limitations that may affect the outcome of a case 1. A genuine source may be cited in support of a fabricated proposition A model may provide a genuine case number, a provision of legislation or the title of a publication and then attribute content to that source which it does not contain. This type of error is more dangerous than an entirely fabricated case number because it may remain undetected during a cursory review. The system should take the user directly to the relevant passage in the source. Verification by a lawyer remains necessary. A Stanford RegLab study of US legal research tools combining database searches with generated answers found hallucinations in 17-33% of responses. The study covered products and questions from the US market, so it does not measure the quality of Polish systems. Connecting a model to a legal database reduces the risk. Further safeguards are required to verify the accuracy of its output. 2. The model may apply an outdated or inapplicable provision An answer may sound plausible while relying on a provision that has been amended, has not yet entered into force or applies in another jurisdiction. In a cross-border matter, the model may confuse the applicable law, jurisdiction and procedural rules. An AI system for lawyers should identify the applicable legal system and the date as of which the law has been verified for each conclusion. The user should be able to see the version of the legislation, the date on which an amendment entered into force and the relevant official source. A label stating that the information is “up to date” has limited value unless the interface also indicates when and on what basis its currency was verified. 3. The model analyses only the case materials made available to it The case file may be incomplete, an attachment may not have been read correctly, or a material fact may be contained in a table, scan or message that was not included in the analysis. The model formulates its answer on the basis of the materials it has received and has no knowledge of documents that were not submitted for analysis. As a result, a confident answer may be based on an incomplete picture of the case. Before starting the analysis, the system should display a list of the files used, any processing errors and the scope of any omitted material. In matters requiring findings of fact, it is useful to provide a separate summary of assertions, supporting evidence and missing materials. This allows the lawyer to identify conclusions reached without access to a key document. 4. Calculating procedural deadlines requires clearly defined rules The calculation of a deadline depends on factors including the type of proceedings, the method of service, the date of the relevant event, transitional provisions, public holidays and exceptions applicable to the specific procedural step. A language model may overlook one of these conditions or misread a date from a scanned document. Procedural deadlines should be calculated using a controlled, rules-based mechanism with clearly stated input data. A model may extract dates from documents and suggest their legal significance. The final calculation should identify the legal basis, show how the deadline was calculated and be approved by the lawyer handling the case. Deadlines whose expiry may produce procedural consequences require an independent calculation and a second verification. 5. Factual assertions in an AI-assisted court filing require appropriate evidence According to press reports discussing the judgment of the Regional Court in Wrocław of 27 November 2025 in case X GC 455/25, the claimant used ChatGPT when preparing the statement of claim. The claim was dismissed because the formal conditions of the request for proposals had not been satisfied. The Court also addressed the AI-generated content and the evidential value of the materials submitted. Printouts containing generated arguments, legal analysis and an assessment of the prospects of success did not constitute evidence of facts material to the determination of the case. They could be treated as part of the claimant’s legal argument. A legal AI system should clearly distinguish between factual assertions, legal grounds and evidence. A missing contract, item of correspondence, proof of service or financial document remains an evidential gap regardless of the quality of the generated reasoning. 6. The model may reinforce the client’s assessment of the case Clients often describe a dispute from one perspective and expect confirmation of their own assessment. A model may adopt the client’s assumptions and overlook the other party’s arguments. This creates a risk that the client will make a decision without understanding the weaknesses of the case. The analysis should consider the opposing party’s perspective, including missing facts, potential counterarguments, procedural obstacles, alternative legal characterisations and the level of uncertainty. In client-facing communications, users should be able to refer a question to a lawyer, particularly when it concerns a deadline, a legal claim, criminal liability, termination of employment or a significant financial decision. 7. A convincing style may conceal errors and uncertainty A model generates an answer word by word and may maintain the same professional tone when presenting both correct and incorrect conclusions. Assessing reliability on the basis of style is therefore unsafe. A confidence score expressed as a percentage may also create a false impression of precision if it has not been calibrated for the specific use case. A well-designed system identifies its sources, missing information, conditions that could change the answer and situations in which it cannot provide a definitive conclusion. It should also be able to decline to answer and refer the matter to a lawyer. Can a client claim that AI caused them to lose the case? A client may raise such an allegation. The assessment of liability will depend on how the legal service was performed and on the circumstances of the individual case. In a contractual relationship, the primary rules are those set out in the Polish Civil Code. Under Article 355 § 2 of the Civil Code, the professional nature of the debtor’s activity must be taken into account when assessing due care. Article 471 of the Civil Code sets out the rules governing liability for non-performance or improper performance of an obligation. An adverse outcome does not in itself establish that a legal services agreement was performed improperly. The assessment covers compliance with the applicable professional standard of care, the correct determination of the factual and legal position, and the appropriateness of the actions taken. In a contractual liability dispute, the client should demonstrate: the existence and terms of the obligation, non-performance or improper performance of the obligation, the loss suffered, an adequate causal link between the breach and the loss. Article 471 of the Civil Code establishes a presumption that non-performance or improper performance of an obligation results from circumstances for which the debtor is responsible. Whether the debtor can avoid liability must be assessed in light of the circumstances of the individual case. The assessment may be affected by how AI was used, particularly whether the sources were verified, the complete case file was considered, the current law was applied and control over the final content of the advice was retained. A system failure or an error attributable to the provider will be assessed together with the choice of tool, the contractual terms, the scope of testing performed and the lawyer’s method of verifying the output. Examples include: missing a deadline because the date of service was determined incorrectly, advising against pursuing a legal remedy on the basis of an outdated provision, filing a submission containing a quotation that does not appear in the cited judgment, recommending a settlement on the basis of an incomplete case file or an incorrect calculation of the financial consequences, a chatbot giving the client a definitive answer without referring the matter to a lawyer. These examples are illustrative. Any assessment of liability requires an examination of the scope of the engagement, the applicable standard of professional care, the loss suffered and the causal link. The conditions for disciplinary liability may be assessed separately. The Polish Law on the Bar and the Polish Act on Attorneys-at-Law also require advocates and attorneys-at-law to hold professional liability insurance. Compulsory professional liability insurance does not necessarily cover every loss connected with the use of AI. The insurer’s liability depends on the terms of the policy and the circumstances of the individual event. What procedural consequences may result from an incorrect AI-assisted court filing? The consequences depend on the type of defect and the applicable procedural rules. An error resulting from the use of AI is assessed in the same way as any other error in a court filing. Its significance depends on how it affects compliance with formal requirements, proof of the relevant facts, the legal and factual basis of the relief sought, and compliance with applicable deadlines. Depending on the type of defect, the consequences may include: the return of a filing if its formal defects have not been remedied, the rejection of a statement of claim, appeal or other means of challenge where the conditions specified in the applicable procedural rules are met, the court disregarding an application for evidence or the evidence itself, a finding that a material fact has not been proven, the court declining to accept an argument based on a source that does not support the proposition attributed to it, the dismissal of an application, appeal or claim because the required conditions have not been established, an order requiring the party to pay the costs of the proceedings or the imposition of a procedural sanction where provided for by the applicable rules. In case I FZ 104/26, the Supreme Administrative Court dismissed the appeal because the appellant had failed to substantiate the conditions for staying the enforcement of the decision under Article 61 § 3 of the Law on Proceedings before Administrative Courts. The general nature of the arguments and the absence of supporting documents were relevant to this assessment. The incorrect references to case law were an additional factor in the Court’s critical assessment of how the submission had been prepared. Any claim for damages against counsel is considered in separate proceedings. Disciplinary liability is assessed by the competent bodies of the relevant professional organisation. What follows from professional rules and the AI Act in 2026? On 15 June 2026, the Polish Bar Council announced the adoption of a resolution amending the Code of Ethics for Advocates and Dignity of the Profession by adding § 23e. According to the information published by the Polish Bar Council, technological tools should serve an auxiliary function. Their use must respect professional secrecy, the advocate’s independence and the advocate’s personal role in handling the case. The output produced by such a tool must be independently assessed and verified by the advocate. For attorneys-at-law, relevant points of reference include the recommendations on the use of AI published by the Polish National Bar Council of Attorneys-at-Law. These practical guidelines address professional responsibility, confidentiality, output verification and human oversight. At EU level, Article 4 of the AI Act has applied since 2 February 2025. It requires providers and deployers of AI systems to take measures supporting the development of AI literacy among their personnel. The measures selected should take account of the personnel’s technical knowledge, experience, education and training, as well as the context in which the system is used and the people in relation to whom it is intended to be used. The European Commission explains that the appropriate way to fulfil this obligation depends on the organisation’s role and the risks associated with the specific use of AI. Since 2 August 2026, the competent authorities have been responsible for supervising compliance with this obligation. The classification of an AI system depends on its intended purpose. High-risk systems may include solutions intended to be used by, or on behalf of, a judicial authority to assist that authority in researching and interpreting facts and law and in applying the law to a specific set of facts. The assessment covers the system’s actual function, the intended purpose specified by the provider and the way in which the deployer uses it. Law firm tools used to search documents, draft text or prepare summaries require an individual classification assessment. The fact that a system is used by a law firm or legal department does not in itself place it in the high-risk category. Following the amendments adopted in 2026, the obligations concerning systems listed in Annex III are due to apply from 2 December 2027. Depending on the system’s function, the type of data involved and the way in which it is used, the GDPR, rules protecting professional secrecy, the applicable procedural rules, and civil and disciplinary liability rules may also apply. How should an AI system for lawyers be designed to reduce risk? A properly designed implementation should make it easier to identify errors, limit their effect on the matter being handled and document completion of the required review. Risk Control built into the product or process Review record Fabricated proposition or quotation A link to the full source and the specific passage, a contextual preview, and mandatory approval before export Court, date, case number, type of decision, source, document version and approving reviewer Outdated law Jurisdiction and date metadata, version control for legislation, and notifications of amendments Date as of which the law was verified and the version of the provision used Incomplete case file A list of analysed files, OCR error notifications and an inventory of missing data Document inventory and file processing report Incorrect deadline calculation A mechanism based on defined rules, clearly stated input data and a second human review Legal basis, input data, calculation method and approving reviewer Overly definitive advice Questions about missing facts, escalation criteria and the ability to decline to provide an answer Reason for escalation and details of the follow-up action taken Disclosure of information protected by professional secrecy Case-level permissions, controls over access by the provider and its subprocessors, a defined processing location, data retention and deletion rules, and exclusion of client data from model training Access logs, provider configuration, retention period, information about subprocessors and incident records Changes in quality following a system update Testing on representative matters before deployment and after any change of model Test results, model version and the decision approving the new version for use The scope of documentation should be proportionate to the risk. Audit logs, including the history of prompts and outputs, may contain information protected by professional secrecy. The organisation should define which data is recorded, who is authorised to access it, how long it is retained, how it is deleted and how the logs are secured. The full history of interactions with the system need not be retained where a narrower set of information is sufficient to demonstrate that the required review was performed. A safe allocation of tasks between AI and the lawyer Tasks should be assigned according to the potential harm and how easily an error can be detected. Task Role of AI Required review Summarising a long document Preparing a working summary with references to the relevant pages Reviewing the passages material to the decision Comparing versions of a contract Identifying and organising changes Assessment of their legal significance by a lawyer Case law research and analysis Identifying potentially relevant judgments and extracting the relevant passages Reading the full judgment and assessing its factual and legal context Court filing Preparing a draft structure, editing the text and checking consistency Full verification of the facts, evidence, relief sought, legal grounds and attachments Procedural deadline Extracting dates and identifying potentially applicable calculation rules Determining the event that starts the time limit, the legal basis, the calculation method and the consequences of missing the deadline Substantive-law time limit Organising dates and identifying provisions requiring analysis Determining the nature of the time limit, when it begins and ends, and the consequences of its expiry Limitation period Organising events that may affect the running of the limitation period Assessing when the period begins, whether it has been suspended or interrupted, and when it expires Final advice to the client Preparing working materials and alternative analyses The lawyer’s personal assessment, approval and communication of the advice This approach is consistent with the practical direction set out by the CCBE in its guide for lawyers: lawyers remain responsible for their work, advice and representations, and generative AI output must be reviewed before it is used. A safe AI implementation in a law firm begins with process analysis The first step is to identify where AI-generated output may affect advice given to a client, a court filing, the assessment of a document or the calculation of a deadline. This provides the basis for defining the appropriate data sources, access permissions, verification rules and the people responsible for approving the output. If you are planning to use AI to analyse case files, work with documents or prepare contracts, explore the AI4Legal solution. We help law firms and legal departments design tools tailored to their workflows, security requirements and the scope of lawyers’ professional responsibilities. Sources Supreme Administrative Court, order of 23 June 2026, I FZ 104/26. Law on Proceedings before Administrative Courts, consolidated text, Journal of Laws of 2026, item 143. Polish Civil Code, consolidated text. Polish Law on the Bar, consolidated text. Polish Act on Attorneys-at-Law, consolidated text. Polish Bar Council, amendments to the professional ethics rules concerning AI, 15 June 2026. Polish National Bar Council of Attorneys-at-Law, recommendations on the use of AI. CCBE, Guide on the Use of Generative AI for Lawyers, 2 October 2025. European Commission, AI Act regulatory framework and guidance on AI literacy. Regulation (EU) 2026/1744 amending the timeline for the application of certain provisions of the AI Act. Stanford RegLab, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Dziennik Gazeta Prawna, discussion of the judgment of the Regional Court in Wrocław in case X GC 455/25, 7 January 2026. Dziennik Gazeta Prawna, interview concerning the grounds for dismissing the claim in case X GC 455/25. Law and sources current as of 1 September 2026. This material is provided for informational purposes and does not constitute legal advice. FAQ: limitations of AI in legal software Can a client claim damages if incorrect legal advice was produced using AI? Such a claim may be possible. Liability will depend on whether the legal service was performed with due professional care, whether the client suffered loss and whether there is a causal link between the breach and the loss. The assessment may also cover the lawyer’s professional standard of care, including how the AI-generated output was reviewed and verified. Will a court reject a filing solely because it was prepared using AI? There is no general rule under Polish law requiring a court to reject a filing for this reason. The court assesses compliance with formal requirements, applicable deadlines, legal arguments and evidence under the relevant procedural rules. Incorrect citations, insufficient evidence or a defective claim may, however, lead to the procedural consequences ordinarily associated with those deficiencies. Can AI manage procedural deadlines on its own? For critical deadlines, AI should operate as part of a broader workflow that includes controlled rules and human approval. A safer process uses verified calculation rules, clearly stated input dates, the relevant legal basis, human approval and an independent reminder. How can you verify whether a judgment actually supports a proposition generated by AI? Open the full judgment in an official or reliable legal database, locate the cited passage, read it in context and verify the date, judicial panel, type of ruling and applicable law. A correct case number and court name do not establish that the judgment supports the proposition attributed to it. Must every AI prompt and response be retained in the case file? There is no single general requirement to retain the full history of every interaction with an AI system. The appropriate scope of documentation should reflect the level of risk, the law firm’s internal policies, professional secrecy, data protection requirements and audit needs. In higher-stakes matters, it is advisable to retain the sources used, the system version, the scope of human review and the identity of the person who approved the output.
Read
What Can GPT DO in 2026 That It Couldn’t Do in 2025?
In 2025, ChatGPT could search the web, analyse documents, create great-looking visuals and help write code. Even then, the pace of development was staggering: over the course of a single year, OpenAI released five major GPT versions, from GPT-4.5 to GPT-5.2. Some of us, users of one of the world’s most popular language models, were already struggling to identify what else could be significantly improved. This article is for everyone else, especially those who still wanted: to delegate an entire task to ChatGPT to let it work freely across applications and on a computer to have it deliver a finished result without being guided through every stage of the process. These are precisely the missing elements that began to appear in 2026. Below, I have collected more than a dozen new ChatGPT capabilities, some of which I had not even heard of until recently, that best illustrate how much the way we work with this tool has changed. Not all the capabilities described below belong to the GPT-5.6 model itself. Some are part of ChatGPT Work, Codex, the desktop app or the Responses API, which provide the environments and tools through which the model completes its work. Readers interested in GPT’s evolution over previous years can also read The Evolution of AI: From GPT-1 to GPT-4o. Users could already choose models designed for more demanding reasoning tasks in 2025. GPT-5.6 provides greater control over this process: the Max level allocates more resources to analysis, while Pro mode allows the model to do additional work before presenting a single final answer. The two settings can be combined for tasks in which the importance of the decision justifies a longer wait and higher token usage, such as selecting a technology, assessing risk or comparing several business scenarios. I have written more about how to manage this “slider” and match GPT modes to specific tasks in the article GPT-5.6 from OpenAI: What Has Changed? Pricing, Capabilities and Business Applications. 1. ChatGPT Can Operate Your Computer for You Behind this slightly clickbait-style heading is a practical situation in which you want to check something that a website does not describe explicitly, meaning that a language model cannot simply read the answer from the page. Normally, you would need to open the website and navigate its interface yourself. With the Computer Use skill, ChatGPT can do this for you: launch a browser or application, click the appropriate controls, enter data and observe the results of each action. You define the scope of the task, approve access to the website or application and retain control over actions that require confirmation. The first capabilities of this kind appeared in 2025 with Operator and later ChatGPT agent. In 2026, they were extended to applications running on macOS and Windows computers, local browser sessions and broader use with ChatGPT Work and Codex. I recently tested this with a very simple example. I wanted to find out whether the free version of a particular VOD service interrupted a specific film with advertisements, something that would have taken me at least several minutes to check manually. I delegated the task to GPT, which, after receiving my permission, opened the website, started the film, scrubbed through the entire video and reassured me that it contained no advertising breaks. The whole process took a minute and a half and ran entirely in the background. Since August 2026, ChatGPT Work has also been able to continue a task on a supported website that requires the user to sign in. When it encounters a login screen, it hands control to the user through a secure form for entering a username, password and 2FA code. According to OpenAI, the model neither sees nor stores these details. Once the user has signed in, ChatGPT returns to the task, and the session can remain active for subsequent instructions. You can therefore ask ChatGPT to find an invoice in a customer portal, check a bill or compare plans that become visible only after signing in. More information is available in the official ChatGPT browser documentation. 2. ChatGPT Can Combine Information from Several Business Applications A client meeting begins in an hour, and you need to review the latest email arrangements, important Slack messages, the current proposal stored on the company drive and the status of the sales opportunity in your CRM. Gathering this information manually means opening one tab after another, reconstructing the project history and dealing with unnecessary stress before the meeting. ChatGPT can search approved sources, select the information related to the client and prepare a one-page brief covering the current situation, open issues, risks and suggested next steps. With the appropriate permissions, it can also save a note in the CRM or prepare tasks for approval. In 2025, ChatGPT could already search and combine information from connected applications. In 2026, ChatGPT Work can gather information from several sources, turn it into a finished deliverable and then perform actions for which it has received permission. Access to each application and the ability to save changes depend on the installed integrations, granted permissions and workspace settings. 3. You Can Create an Agent for a Specific Business Process Imagine an agent assigned to support your most important clients. A new request can automatically trigger its work. The agent collects previous correspondence, checks similar cases and finds the appropriate procedure in the company’s documentation. It then prepares a response and, if it has the necessary permissions, creates a task for the appropriate team. The message to the client remains pending until an employee approves it. In 2025, a similar assistant could be built as a custom GPT equipped with instructions, knowledge and integrations. Workspace Agents, introduced in 2026, expand this concept into a saved process that includes data sources, actions in applications, required approvals and the format of the finished deliverable. An agent can be shared with a team and started manually, according to a schedule or through a signal sent by a company system. During configuration, you define which applications and sources the agent can use, which actions it is allowed to perform and when it should stop and ask a person to make a decision. This allows the same process to follow a consistent set of rules every time a new request arrives. 4. ChatGPT Can Learn a Process by Observing Your Work Some procedures are difficult to describe. You may know exactly where to click and what to change, yet writing down every step would take longer than completing the task again. A good example is a monthly report prepared in a legacy system: you select the appropriate filters, export the data, organise the spreadsheet and save the file in the correct folder. Record & Replay, introduced in 2026, allows you to complete the process once with recording enabled. ChatGPT or Codex observes the required actions and uses them to create a skill describing the workflow, its variable data and the method for checking the result. You can review it, make corrections, add your own rules and then use it when preparing the next report. This is particularly useful when process knowledge exists mainly in employees’ heads and a proper set of instructions has never been created. I personally cannot wait to try this feature while editing a video in DaVinci Resolve. I will show ChatGPT how I remove unsuccessful takes, organise the audio track and conceal visible cuts with a short zoom-in. I am curious to see how much of this repetitive work can be turned into a reusable skill for future recordings. 5. ChatGPT Can Help You Resume Interrupted Work Human memory is unreliable. It was probably Tuesday, after a conversation with your manager, when you made changes to an important spreadsheet. You cannot remember its name or location, or even whether it was open in the desktop version of Excel or in a browser. A standard file search is not much help. This is where Computer History can help. You can ask ChatGPT, for example: “Find the spreadsheet I edited after my conversation with Krzysztof.” Your activity history can connect the sequence of applications you opened, the actions you performed and the available context, and then identify the material you probably have in mind. If ChatGPT has access to the appropriate source, it can also open the file directly. You can ask what you were working on before a break, request a summary of the previous day’s work or identify activities that occur regularly and could be turned into a skill or automation. You choose which applications and websites are included in the history, and you can pause data collection at any time. Computer History records information including clicks, typed text, keyboard shortcuts, application changes and context provided by macOS. Temporary event data is deleted after no more than 48 hours, while local memories created from that data remain until the user deletes them. The feature is currently available in the ChatGPT app for macOS, requires Memories to be enabled and is turned off by default. In Business and Enterprise workspaces, an administrator must first grant access, after which each employee can decide whether to enable it. 6. ChatGPT Can Design, Test and Publish a Web Application How many good ideas have become stuck in the IT department’s backlog because something more urgent always came up? Suppose the administration team needs a simple tool for reporting faults, with a form, a list of cases, their statuses and a separate view for the person responsible for repairs. With Sites, you can describe the tool in your own words and provide sample data, company materials or a screenshot of a similar system. GPT-5.6 will prepare a working prototype, and you can ask it to simplify the form, add filters or adapt the design to your brand identity. ChatGPT or Codex can then test the application in a browser, fix any identified issues and publish the finished version at a specified address. Websites and simple applications were being created with GPT long before 2026. This year’s change concerns the entire process: GPT-5.6 is better at designing usable interfaces, while Sites provides a single place to launch a prototype, make changes, save subsequent versions and publish the finished application. The same process can be used to create calculators, dashboards and small internal tools. Before they are deployed in a real business environment, their security, permissions, data storage and regulatory compliance still need to be reviewed. 7. ChatGPT Can Work Directly in Excel and Google Sheets Consider a simple example. You open a spreadsheet containing campaign results and notice that some formulas are outdated, several cells are empty and the cost per lead has suddenly increased in two campaigns. Instead of checking everything one item at a time, you ask ChatGPT to analyse the data, complete the calculations, identify unusual results and prepare a chart for the monthly summary. ChatGPT performs this work directly in Excel or Google Sheets. You can see the changes it makes, review the formulas and immediately request another adjustment, such as changing the chart range, adding a comparison with the previous month or highlighting campaigns that exceed the target cost. In 2025, a similar analysis usually required uploading a spreadsheet to ChatGPT and then downloading the modified file. ChatGPT for Excel and Google Sheets, introduced in May 2026, allows you to work on the spreadsheet where the data is stored and where you will later use it. This makes the conversation with the model part of the document workflow instead of a separate stage outside it. 8. ChatGPT Can Create and Refine Content in Its Final Format The board meeting is one day away, and the information needed for the presentation is scattered across meeting minutes, a spreadsheet of results and several project documents. There is also a company template that must be followed. ChatGPT Work can collect these materials, organise them into a coherent story, prepare the slides, add charts and identify figures or conclusions that are not supported by the source data. You receive an editable PowerPoint or Google Slides presentation that can retain the structure, visual style and brand elements of the provided template. In 2025, GPT could already generate editable PowerPoint, Word and Excel files. In 2026, ChatGPT Work combined document creation with previews, targeted revisions, saved templates and final material checks. The entire process can therefore take place within one conversation, from gathering the sources to preparing subsequent versions of the document. 9. Codex Can Manage Long-Running Work as a Persistent Goal You have a large audit of a company website ahead of you. There are hundreds of pages to review, along with outdated information, broken links and differences between language versions. Finally, all the issues must be grouped and converted into a list of specific changes. Codex could already perform tasks in the background in 2025. Goal Mode, introduced in 2026, allows you to delegate much longer assignments and manage them across successive stages. You can monitor progress, change the direction of the work when necessary and access the active project from your phone. Several such assignments can run in parallel while you return to them whenever your input or decision is required. 10. You Can Talk to ChatGPT While It Works on Your Screen Voice messages have become part of everyday communication, while assistants such as Siri and Google Assistant have accustomed us to controlling technology by voice (“Hey Siri”, “OK Google”). We increasingly expect the same from the tools we use at work, especially when a dashboard, document or application is already open and pointing something out is easier than describing it in detail. Needless to say, OpenAI has kept pace with this trend. Imagine that, hypothetically, your screen displays a sales dashboard showing a clear drop in performance in one region. You only need to say: “Find out what happened.” ChatGPT can see the report, compare the relevant periods and, if it has access to the source data, investigate the cause and prepare a comment for the results presentation. Voice conversations and screen sharing were already available in 2025. In 2026, voice also became a way to direct work performed by ChatGPT Work. Instructions such as “compare these two periods”, “check the data in the application” or “improve this section” refer directly to the material visible on the screen. Part of the task can then continue in the background. 11. A Recurring Report Can Be Generated Without Repeating the Prompt Imagine that every Monday in your company begins with the preparation of a sales report. Each time, someone has to retrieve the latest results, compare them with the plan, identify the largest variances and turn them into a short summary for management. ChatGPT could already run simple tasks at a specified time in 2025. In 2026, an entire repeatable process can be placed on a schedule. Once configured, the task accesses the specified available sources, performs the analysis and leaves the completed report for your review. The same approach can be used to prepare a morning briefing, review new requests, summarise changes in documents or monitor competitors regularly. Each run has its own status and history, allowing you to review the result, refine the instructions and improve the process based on subsequent reports. 12. You Can Combine Company Skills and Tools in a Single Plugin Imagine that several teams regularly prepare materials for clients. They should always use current service descriptions, approved case studies, company templates and the same communication guidelines. They also need access to the content management system, the asset library and the tool in which documents are submitted for approval. An Agent Plugin allows these elements to be combined into a single package available in the company workspace. Once installed, ChatGPT can use the appropriate instructions, sources and permitted actions when preparing a presentation, proposal, service description or website content. When the company template or approval process changes, the plugin can be updated once instead of modifying the configuration of every agent separately. In 2025, a similar solution required custom GPTs, instructions and integrations to be connected and configured separately. Agent Plugins, introduced in 2026, allow them to be distributed as a single installable package. A plugin is not a separate agent and does not start work by itself. It extends ChatGPT, Codex or a Workspace Agent with the company knowledge, rules and tools needed to complete a task. 13. GPT-5.6 Can Divide a Task Among Several Agents You are considering entering a foreign market, which requires an analysis from several perspectives. You need to research the competition, compare prices, estimate costs, assess demand for the services offered and prepare an initial list of potential partners. GPT-5.6 can divide these areas among several subagents. Each subagent handles its own part of the analysis, and the work can proceed in parallel. The main agent collects the results, compares the findings and combines them into a single recommendation with sources and a list of issues requiring further verification. Multi-agent solutions were already being developed in 2025, but they required the system to be designed independently. In GPT-5.6, the mechanism for dividing work has been built into the Responses API and is currently available as a beta feature. It works best when a task can be divided into independent parts assessed according to shared criteria. Parallel work by several agents can reduce the time required for the analysis, although it usually increases token usage. 14. GPT-5.6 Can Combine Multiple Operations into a Single Program Consider a hypothetical scenario: a CRM migration is only a few weeks away, and before it begins, 800 records must be checked, missing information located and data requiring correction identified. GPT-5.6 can write a JavaScript program that retrieves successive records through an available tool, checks the required fields, compares the data against defined criteria and creates a list of missing information. Loops, conditions and result processing are handled within the runtime environment, so the model does not need to analyse each of the 800 records separately. Tool calling was already possible in 2025, but successive operations usually required the results to be sent back to the model or the necessary logic to be written within the application. Programmatic Tool Calling allows GPT-5.6 to prepare code that combines multiple predictable operations. This can reduce processing time and token usage. The company decides which tools and data the program can access and which actions require approval. The feature operates through the API, so it must be implemented within a company application or process. It is not an option available directly in the ChatGPT window. 15. ChatGPT Can Turn Data into an Interactive Visualisation Before a meeting, you receive a table containing the results of an employee survey. A standard chart shows the average scores, but it does not allow you to explore how responses differ between departments, locations and seniority groups. ChatGPT can prepare an interactive visualisation in which you select the criteria you are interested in, change the data range and observe how these choices affect the results. In 2025, GPT could already create charts and analyse data. Visualize, introduced in August 2026, makes it possible to build interactive diagrams, maps, timelines, simulations and information exploration tools directly within a conversation. The finished visualisation can be filtered, adjusted and used to investigate additional questions that arise during the analysis. GPT and ChatGPT Features in 2026: Availability by Plan and Device The table below summarises the availability of the features described in this article. Information accurate as of 31 August 2026. Feature Free / Go Plus / Pro Business / Enterprise / Edu Where is it available? Key limitations GPT-5.6 Yes, GPT-5.6 Luna is the default model Yes, with access to the Sol, Terra and Luna family Yes, depending on workspace settings Web, desktop app, Codex and iOS; also through the API Usage limits and the choice of model variant depend on the plan and environment. ChatGPT Work Yes, with lower usage limits Yes Yes, with additional administrator controls Web and desktop app; some capabilities are also available on mobile It shares usage limits with Codex. Tool availability depends on the plan, region and workspace settings. Computer Use Full availability has not been clearly confirmed Yes, in supported regions Yes, if permitted by workspace settings Desktop app on macOS and Windows Work running on a computer can be monitored remotely from iOS. Linux does not yet support this feature. Browser and website interaction Limited availability Yes Yes, although some capabilities may be disabled by an administrator Web, desktop app and partially on mobile Signing in to websites through the cloud browser is available on Plus and Pro, but is not currently available on Enterprise and Edu. Plugins, skills and application integrations Depends on the individual plugin Yes Yes, after administrator approval Web, desktop, iOS and Android; also Codex CLI Plugins do not work in the Codex IDE extension. The availability of a specific integration may depend on the plan and region. Workspace Agents No No Yes, on Business, Enterprise and Edu plans Company ChatGPT workspace An administrator must enable agents and grant permissions to create, publish and connect them to applications. Computer History No Pro only Yes, on Business and Enterprise ChatGPT desktop app on macOS only The feature is disabled by default, requires Memories and needs administrator approval in a company workspace. It is available in the EEA, Switzerland and the United Kingdom. Creating documents, spreadsheets and presentations Yes, with lower usage limits Yes Yes ChatGPT Work on the web and in the desktop app The scope of editing depends on the format, available plugins and permissions for the source files. ChatGPT for Excel and Google Sheets Availability has not been confirmed Yes Yes Directly within a supported spreadsheet The feature uses ChatGPT Work limits. It is not provided through API-key authentication alone. Sites No Yes Yes ChatGPT on the web and in the desktop app The feature remains in public beta. Limits depend on the plan, and full Sites management is not described as a mobile app capability. Long-running tasks and Goal Mode The available scope depends on plan limits Yes Yes ChatGPT Work on the web, the desktop app, Codex CLI and the IDE extension Tasks that work with local files require the application to remain running and the computer to be available. Voice control for tasks Not in the ChatGPT Work form described here Yes Yes Desktop app; on iOS through Remote after pairing with a computer Voice conversation limits depend on the plan. Tasks started by voice also use the Codex allowance. Scheduled tasks and automations Availability depends on the account Yes Yes, if enabled by an administrator Web and desktop app; event triggers are also available on iOS and Android Triggers can respond to supported events in Gmail, Slack and GitHub. Subagents and Ultra mode Full access has not been confirmed Yes, depending on the available model and reasoning level Yes ChatGPT Work, the desktop app and Codex CLI In most modes, the user must explicitly request delegation. Ultra can launch subagents automatically and consumes more of the usage allowance. Interactive visualisations Depends on whether the feature has reached the account Yes, on supported accounts Yes, if the workspace allows the plugin to be used Web; currently rolling out on desktop and mobile Visualisations are not rendered in Codex CLI or the IDE extension. Programmatic Tool Calling Not available as a ChatGPT interface feature Available through paid API usage Available through the API Responses API This feature is intended for developers building their own applications. It allows GPT-5.6 to coordinate tools through executed JavaScript code. Persisted Reasoning and Pro and Max modes Not available as separate settings for Free users Partially available in the interface; full configuration through the API Depends on workspace settings; full configuration through the API Responses API and selected ChatGPT and Codex surfaces GPT-5.6 can retain compatible reasoning elements between successive calls. Pro mode can improve quality at the cost of additional time and token usage. The availability of OpenAI features changes with subsequent updates. It may depend on the plan, region, operating system, application version, selected model and permissions granted by the administrator of a company workspace. How Do You Turn GPT’s Capabilities into a Solution That Works for Your Business? These fifteen examples are enough to illustrate the scale of the change. Companies still need to make the most important decisions: which processes should be delegated to GPT, which data it may access, which actions it is allowed to perform and when human approval is required. At TTMS, we begin with these questions. We analyse the process, its objective, the systems involved and the acceptable level of risk. We then select the appropriate solution, whether ChatGPT Work, Codex or an integration through the OpenAI API, design the permissions and connect GPT to the company’s applications. We test the finished solution using real cases and measure its quality and costs. Following the launch, we provide monitoring, maintenance and further development. We have delivered projects of this kind. For Stäubli, we integrated ChatGPT with Adobe Experience Manager, while for Takeda, we created an AI solution supporting document analysis in Salesforce. This experience is backed by more than 800 TTMS specialists and an approach to AI management and security confirmed by ISO/IEC 42001 and ISO/IEC 27001 certifications. Would you like to find out which process could deliver real value to your company with GPT? Discuss it with the TTMS experts. Are all new GPT features available to every user? Access to individual features may depend on the subscription plan, region, operating system, and workspace settings. Some capabilities are initially released to selected users or require administrator approval. In a business environment, availability also depends on user permissions and access to specific applications. Before planning a process, it is worth checking which features are supported in your organisation’s current setup. How can you identify a process that is suitable for GPT? A good candidate is a process that is performed regularly, uses digital data, and follows rules that can be clearly described. A practical starting point is a time-consuming task such as gathering information from several sources, preparing reports, comparing documents, or completing missing data. The process should also have a clearly defined outcome, making it possible to evaluate the quality of GPT’s work, the time saved, and the number of errors. Can GPT-5.6 work securely with company data? Yes, provided that access to the data is properly designed and controlled. The organisation should define which sources the model can use, which actions it is allowed to perform, and when human approval is required. Data storage policies, permission management, monitoring, and regulatory compliance also need to be considered. For processes involving confidential information or personal data, security should be built into the implementation from the beginning. How can a company measure the benefits of implementing GPT? The most reliable approach is to compare the process before and after implementation. Useful metrics include task completion time, operating costs, the number of errors, the percentage of cases requiring manual correction, and the quality of the final output. It is also worth checking whether employees actually use the solution and whether it makes their everyday tasks easier to complete. A pilot involving a limited number of cases can help measure these results before the implementation is expanded.
Read
15 ChatGPT Integrations with Business Apps in 2026
How can ChatGPT integrations with business applications simplify everyday work in 2026? Here is a simple example: a client emails us asking for a project status update. At this point, we face half an hour of clicking between Google Drive, Slack, Asana and the CRM system. What if ChatGPT could collect information from all these sources in a single conversation and immediately prepare a summary, response or plan for the next steps? In this article: we examine 15 ChatGPT integrations with popular business applications that can make the scenario described above part of a company’s everyday workflow, we explain the differences between apps, integrations, plugins, GPTs and MCP servers, we present specific business use cases and highlight what should be checked before implementation, including permission scopes, data security, availability and costs. How do business application integrations extend ChatGPT’s capabilities? In the client enquiry scenario described above, the right set of integrations could work as follows: ChatGPT would find documents in Google Drive, summarise conversations in Slack, check task statuses in Asana and analyse the client’s data in the CRM system. Individual integrations may be available in ChatGPT as apps, connectors or MCP-based solutions. They make it possible to use data and selected functions from external services without leaving the conversation, and then prepare an up-to-date summary, a response for the client or a plan for the next steps. The available capabilities depend on the specific solution. Some integrations are more “passive” and are used mainly for searching and reading data. More “active” integrations support creating, updating and sending data. GPTs, apps, connectors, plugins and MCP – how do these concepts differ? The terminology surrounding ChatGPT extensions includes several related concepts. In our previous article, we described the ecosystem of the most useful ChatGPT plugins. In this comparison, we use the term “ChatGPT integrations” as an umbrella term for the different ways of connecting ChatGPT to business applications, their data and their functions. Concept Proposed definition ChatGPT integration An umbrella term for connecting ChatGPT to an external application, its data or its functions. An integration may be implemented as an app, connector, plugin, MCP server or GPT Action. App A function of an external service available directly in ChatGPT, sometimes with an interactive interface. Connector A ready-made connection that gives ChatGPT access to the data or functions of a specific service. Depending on the solution, it may support search, synchronisation or actions. MCP server A layer that gives ChatGPT access to selected tools, data and operations from an external system in accordance with the Model Context Protocol standard. Plugin An installable package that extends ChatGPT or Codex and may include instructions, skills, an MCP connection and an optional interface. These concepts describe different elements of the same ecosystem and are not always completely separate. Integration is the umbrella term for connecting ChatGPT to an external service. It may be available as an app, use a connector or MCP server, while a plugin may combine several of these elements into a ready-made workflow. How do you connect an app to ChatGPT step by step? Define the task the integration should perform. Open the app or plugin directory in ChatGPT. Select the appropriate service and start the connection process. Sign in to the external application and approve the required permissions. Open a new conversation and select the connected app. Test the integration using a limited dataset before deploying it across the entire team. How did we select 15 ChatGPT integrations with business applications? This comparison covers integrations that support recurring business processes and are available directly in ChatGPT or through documented MCP-based solutions. We considered five criteria: Frequency of use: the tool stores data or supports tasks performed by teams every day. Value of context: the connection gives ChatGPT access to information that significantly improves the quality of its output. Scope of actions: the integration supports searching, analysing, creating or updating data. Access control: the provider describes authentication, user permissions or administrative controls. Usefulness across multiple roles: the solution can support sales, marketing, operations, IT, product development or knowledge management. The availability of individual features depends on the ChatGPT plan, the external service plan, the country, workspace settings and administrator decisions. The catalogue and permission scopes should be checked immediately before implementation. 15 ChatGPT integrations with business applications in 2026 1. Google Drive integration with ChatGPT – searching and analysing company documents The Google Drive integration with ChatGPT enables users to work with materials stored in Drive, Docs, Sheets and Slides. Users can search for files, combine information from several documents, analyse spreadsheets and use existing materials as sources for a new report, brief or presentation. It provides the greatest value to teams with well-organised folders and consistent document naming conventions. ChatGPT can then locate the correct versions of proposals, reports, meeting notes and project materials more quickly. Best use case: preparing a project summary based on documents, a results spreadsheet and a status presentation. Example prompt: “Find materials in Google Drive related to Project X from the last 30 days and prepare a summary of decisions, risks and next steps.” The video shows how to connect Google Drive to ChatGPT, create an SEO-optimised blog post and save it as a document in Google Drive. It also highlights the importance of detailed prompts for improving the quality of generated content. 2. SharePoint integration with ChatGPT – access to organisational knowledge, procedures and files SharePoint is a natural source of information for organisations using Microsoft 365. It stores documents, intranet pages, procedures, policies and project materials. The SharePoint integration with ChatGPT enables users to find these resources and use them when preparing responses or documents. It is particularly useful in larger organisations where knowledge is distributed across sites, document libraries and teams. The SharePoint permission structure continues to determine which information is available to each employee. Best use case: finding current policies, instructions, templates and project documentation. Example prompt: “Based on the current procedures in SharePoint, prepare an onboarding checklist for a new supplier.” 3. Box integration with ChatGPT – secure analysis of company documents Box combines content management with access controls and is often used by organisations working with confidential documents. The Box integration with ChatGPT can retrieve data on demand or synchronise selected content. On-demand access retrieves the required information while a prompt is being processed, while synchronisation indexes approved resources in advance and speeds up searches across large repositories. The choice of access mode should take into account data classification, retention requirements and the expected response time. Best use case: analysing contracts, project materials, client documentation and approved company resources. Example prompt: “Find the current versions of documents for Client X in Box and identify discrepancies in the project scope.” 4. Gmail integration with ChatGPT – summarising correspondence and preparing replies The Gmail integration with ChatGPT enables users to search for messages, summarise long threads and prepare draft replies based on their email history. Gmail in ChatGPT is useful in sales, customer service, recruitment and day-to-day coordination when important decisions are distributed across multiple messages. To help the Gmail connector return an accurate result, specify the relevant period, senders, subject and expected outcome. ChatGPT can then find the appropriate messages and turn them into a summary, list of decisions or ready-to-use draft reply. Best use case: summarising an email thread, preparing a follow-up and identifying the commitments made by each party. Example prompt: “Summarise the correspondence with Company X from the last two weeks. List the agreed actions, deadlines and questions that still require a response.” The video shows how to connect Gmail to ChatGPT step by step using the official app. Once the Gmail integration with ChatGPT has been configured, users can search for messages, summarise long threads, find important information and prepare draft replies directly within the conversation. 5. Outlook Email integration with ChatGPT – analysing messages in Microsoft 365 The Outlook Email integration with ChatGPT enables users to find messages, analyse long email threads and prepare replies that take the conversation history into account. Outlook in ChatGPT is particularly useful for organisations using Microsoft 365. If ChatGPT is also connected to SharePoint and Microsoft Teams, it can combine email discussions with documents and team conversations. The Outlook Email integration operates only within sources approved by the organisation and available to the individual user. Best use case: preparing a client response based on email history and current project materials. Example prompt: “Find the latest email thread about renewing the contract with Company X and prepare a draft reply that addresses the outstanding issues.” 6. Slack integration with ChatGPT – summarising team conversations, decisions and actions The Slack integration with ChatGPT gives the model access to context from messages, files, channels and team member profiles. Slack in ChatGPT helps reconstruct the history of decisions, prepare project status updates and identify recurring problems in team conversations. The Slack MCP server also supports selected actions, such as sending messages and creating or viewing Canvas documents. The Slack integration with ChatGPT only uses channels available to the authenticated user and operates according to the rules configured by the administrator. Best use case: preparing a weekly status update covering decisions, blockers, owners and open questions. Example prompt: “Review the project channel from Monday onwards and prepare a status update covering completed actions, risks, decisions and tasks for the coming week.” 7. Microsoft Teams integration with ChatGPT – analysing conversations, meetings and tasks The Microsoft Teams integration with ChatGPT enables users to search and analyse messages from individual chats, group conversations and channels available to them. Microsoft Teams in ChatGPT can also work with Microsoft Planner plans and tasks. When the relevant actions are enabled, it can create chats and channels, as well as send messages and replies. On the Enterprise plan, the integration can also retrieve transcripts from scheduled meetings if the user has the appropriate permissions. Files shared in Teams channels are usually stored in SharePoint, so analysing them requires an additional connection between ChatGPT and SharePoint. Best use case: finding decisions in team conversations and turning them into summaries, tasks and status materials. Example prompt: “Review the conversations in the project channel from the last five days and prepare a list of decisions, open questions, responsible individuals and deadlines.” 8. Notion integration with ChatGPT – creating and updating company knowledge The Notion integration with ChatGPT enables users to read, create and update content on Notion pages directly from a conversation. Notion in ChatGPT can support product documentation, campaign plans, knowledge bases, feature specifications and implementation checklists. The Notion MCP server operates within the permissions of the signed-in user. A person with broad access to the workspace gives the integration an equally broad scope of data and operations, so it is worth beginning the implementation with clearly limited use cases and accounts with appropriately assigned roles. Best use case: transforming notes and analysis results into structured pages, databases and action plans. Example prompt: “Create a feature specification in Notion based on these notes. Add objectives, requirements, acceptance criteria, risks and open questions.” The video shows how to connect Notion to ChatGPT and work with content stored in a workspace. The Notion integration with ChatGPT enables users to search for information and create or update pages directly from a conversation. 9. Atlassian Rovo integration with ChatGPT – working with Jira, Confluence and Bitbucket The Atlassian Rovo integration with ChatGPT connects the model to Jira, Jira Service Management, Confluence and Bitbucket. Jira and Confluence content can be searched and summarised in ChatGPT, while users can also create and update tasks, tickets and pages using natural language commands. The Atlassian Rovo MCP server supports software development, ticket management, change management and documentation processes. OAuth 2.1 authentication preserves existing user roles and permissions, while actions affecting data should be subject to approval and monitoring. Best use case: creating tickets from meeting notes, updating statuses and connecting Confluence documentation with Jira tasks. Example prompt: “Based on this specification, create five Jira tasks with descriptions, acceptance criteria and priorities. Show me the proposed tasks before saving them.” 10. Asana integration with ChatGPT – creating tasks and managing projects The Asana integration with ChatGPT provides information about projects and portfolios, and allows users to create and assign tasks, set up new projects and monitor progress. Asana in ChatGPT can turn decisions made during a conversation into a structured plan saved directly in the work management system. The Asana integration with ChatGPT is useful for planning campaigns, implementations, product launches and cross-departmental initiatives. The integration produces the most accurate results when projects, owners and custom fields have clear and consistent names. Best use case: creating a project plan and turning decisions into assigned tasks. Example prompt: “Create a plan in Asana for launching a new product page. Divide the work into stages, tasks, dependencies and responsible team members. Show me the proposed structure for approval before saving it.” 11. HubSpot integration with ChatGPT – CRM analysis and record updates The HubSpot integration with ChatGPT provides information about contacts, companies, sales opportunities, tickets and customer interaction history. HubSpot in ChatGPT can analyse the sales funnel, campaign results and customer activity, as well as create and update selected records and log activities. The HubSpot integration with ChatGPT is one of the most extensive solutions available to sales and marketing teams. The quality of its results depends on the completeness of CRM data, consistently defined funnel stages and correctly assigned permissions. Best use case: preparing an account brief, analysing the pipeline, updating an opportunity and creating a follow-up. Example prompt: “Analyse the sales opportunities in HubSpot that have had no activity for 14 days. Identify the priorities and prepare a plan for the next contact.” 12. Salesforce Agentforce Sales integration with ChatGPT – opportunity analysis and CRM management The Salesforce Agentforce Sales integration with ChatGPT combines information about customers, sales opportunities and the pipeline with analysis and planning capabilities. Salesforce in ChatGPT allows sales representatives to prioritise opportunities, prepare account plans, update records and run Agentforce actions directly from a conversation. The Agentforce Sales app for ChatGPT is currently available through the Open Beta programme to eligible customers using the required Agentforce add-ons. Before implementation, organisations should verify their Salesforce edition, access requirements and regional availability. Best use case: preparing a sales representative for a meeting, prioritising opportunities and updating the CRM after a client conversation. Example prompt: “Show me five Salesforce opportunities that require attention this week. Include their value, stage, most recent activity, risk and recommended next step.” 13. GitHub integration with ChatGPT – analysing code, issues and project changes The GitHub integration with ChatGPT gives the model access to context from repositories, code, issues, proposed changes and automated test results. GitHub in ChatGPT can help analyse code changes, organise issues, prepare documentation and identify dependencies between project components. Administrators can specify which repositories the GitHub integration with ChatGPT can access and which operations it can perform. This makes it possible to test the integration on a small number of selected projects before gradually making it available to additional teams. Best use case: analysing proposed code changes, organising issues, reviewing automated test results and preparing change documentation. Example prompt: “Review the open pull requests in the mobile application repository. Identify risks, missing tests and issues blocking the release.” The video shows how to connect GitHub to ChatGPT and give the integration access to selected repositories. The GitHub integration with ChatGPT enables users to explore project structures, analyse code and documentation, and summarise changes, commits and pull requests directly within a conversation. 14. Canva integration with ChatGPT – creating and editing visual content The Canva integration with ChatGPT enables users to search and summarise existing materials, as well as create, edit and display designs directly within a conversation. Canva in ChatGPT is useful for preparing presentations, social media posts, documents and other visual materials. Designs created through the Canva app for ChatGPT remain editable in Canva, allowing the team to continue refining their content and appearance. The best results can be achieved by specifying the intended audience, objective, format, source materials and brand requirements. Best use case: presentations, social media content, sales documents and visual summaries. Example prompt: “Create a presentation in Canva for the management team based on this report. Use eight slides, concise conclusions and one chart on the results slide.” 15. Adobe integration with ChatGPT – editing photos, videos, graphics and PDF documents Adobe for ChatGPT is a package that brings together features from Adobe applications, including Photoshop, Premiere, Firefly, Express and Acrobat. It supports photo editing, consistent batch processing, preparation of social media formats, video shortening, work with PDF documents and searches across Creative Cloud assets. The solution supports workflows intended to produce a finished file. For example, a workflow may begin with a set of employee photos, include lighting correction and consistent cropping, and finish with the export of materials ready for publication. Best use case: repeatable photo editing, adapting content for different channels, working with PDF documents and quickly creating designs from templates. Example prompt: “Standardise the lighting and colours in these photos, apply consistent cropping and prepare versions for employee profiles on the company website.” 15 ChatGPT integrations with business applications – comparison table No. ChatGPT integration Area Best use case Primary type of work 1 Google Drive Documents and knowledge Analysing files from Drive, Docs, Sheets and Slides Search, reading and analysis 2 Microsoft SharePoint Organisational knowledge Working with controlled Microsoft 365 resources Search, reading and analysis 3 Box Content management Secure work with company files and folders On-demand access or synchronisation 4 Gmail Email Summarising email conversations and preparing replies Search, analysis and drafting 5 Outlook Email Microsoft 365 email Analysing email in a business environment Search, analysis and drafting 6 Slack Team communication Finding decisions and summarising channels and messages Search, reading and actions 7 Microsoft Teams Collaboration Analysing conversations, meetings and team context Search and summarisation 8 Notion Knowledge and documentation Creating and updating pages, databases and plans Real-time reading and writing 9 Atlassian Rovo Projects and IT Working with Jira, Confluence, Jira Service Management and Bitbucket Search, creation and updates 10 Asana Work management Managing project portfolios and creating tasks Analysis and project actions 11 HubSpot CRM, marketing and sales Analysing customers, the sales funnel and contact history Analysis, record creation and updates 12 Salesforce Agentforce Sales Enterprise sales Prioritising opportunities, planning accounts and updating the CRM Analysis and sales actions 13 GitHub Software development Working with repositories, issues, proposed changes and automated tests Search, analysis and issue organisation 14 Canva Design and communication Creating editable presentations and marketing materials Search, generation and editing 15 Adobe Creative work and documents Photos, videos, social media content, PDF files and Creative Cloud assets Search, generation, editing and export Security of ChatGPT integrations in a business environment Secure integration of ChatGPT with company systems requires appropriate permission management, separation of read and write operations, selection of the right data access method, approval of actions and operation logging. Data processing terms, OAuth scopes, retention and data residency requirements should also be reviewed for every connected service. In ChatGPT Business, Enterprise and Edu plans, data retrieved through integrations is not used to train OpenAI models. When is it worth building a custom ChatGPT integration? Ready-made ChatGPT integrations cover popular business applications and common use cases. A custom integration becomes justified when critical data is stored in an internal system, the process requires specific logic or the organisation needs greater control over its architecture and information flows. The most common reasons include: a private API, legacy system, internal database or on-premises solution; a workflow involving several systems and rules specific to the organisation; requirements concerning data residency, auditability and approval of operations; the need to combine RAG-based search, business logic and actions performed in external systems; a regulated environment requiring risk assessment, documentation and controlled implementation; a scale at which a custom integration simplifies access and cost management. Such a solution may use a dedicated integration, MCP server, GPT Actions, API layer or an architecture combining several approaches. The starting point should be a specific process, a clearly identified data owner and the expected business outcome. ChatGPT integrations as part of a secure enterprise AI ecosystem ChatGPT integrations provide the greatest value when the connection supports a real process, respects user roles and produces an output that is ready to use. For one organisation, this may mean faster knowledge retrieval. For another, it may involve CRM updates, document automation or a controlled process spanning several systems. Transition Technologies MS designs and implements AI solutions for business tailored to an organisation’s data, architecture, security requirements and operating model. The scope of a project may include API and MCP integrations, RAG solutions, action automation and a model for managing access, risk and accountability. Our approach to AI has been confirmed by ISO/IEC 42001 certification for our Artificial Intelligence Management System (AIMS). TTMS was the first company in Poland to obtain accredited certification for compliance with this standard and is among the first organisations in Europe operating within its framework. This means that we deliver AI projects according to structured principles covering security, accountability, documentation and risk management. TTMS also develops proprietary AI products that support specific business processes: AI4Content analyses documents and creates structured reports; AI4Knowledge helps employees use company knowledge more effectively; AI4E-learning transforms source materials into editable online training courses; AI4Localisation supports the translation and adaptation of content for different markets; AI4Legal automates document analysis and selected legal processes; AML Track supports customer verification, risk monitoring and compliance with AML obligations; AI4Hire structures application analysis and supports the initial assessment of candidates; QATANA uses AI to create test cases and manage the software testing process. This expertise allows us to combine integration, product and regulatory experience. We can help organisations establish a single connection to a company data source or design a solution spanning several systems, access controls and end-to-end process automation. FAQ: Frequently asked questions about ChatGPT integrations Can ChatGPT use multiple connected apps in a single task? Yes. Supported ChatGPT environments can use several approved sources within a single task. For example, a workflow could collect project decisions from Slack, retrieve a report from Google Drive and prepare an action plan in Asana. Availability depends on the ChatGPT plan, the interface or mode being used and the workspace configuration. The prompt should clearly identify the required sources, expected result and point at which ChatGPT should request approval. The organisation should also define which types of data may be combined in a single output. Does an integration give ChatGPT access to all of a user’s data? The scope of access depends on the permissions granted to the integration and the user’s role in the source system. Many integrations respect existing permissions for folders, channels, repositories and CRM records. An administrator account may therefore expose significantly more data than an employee account assigned to a specific team. During configuration, review the OAuth scopes, user roles and options for restricting access to selected resources. A pilot should ideally use an account with permissions corresponding to the intended user role. Can ChatGPT send messages and modify data in external applications? Selected apps, integrations and MCP servers support write actions such as sending messages, creating tasks, updating CRM records or adding pages. The available actions vary by provider, subscription plan and integration version. Some tools show the proposed change and request confirmation before completing it. Administrators may also restrict an integration to read-only access or allow only selected operations. Actions affecting customers, financial data, publications or regulated processes should always have clearly defined human approval requirements. Do I need a paid ChatGPT plan to use integrations? Not always. A limited selection of apps may also be available on the free ChatGPT plan, although search and analysis features may have lower usage limits. Broader access, including data synchronisation and custom MCP-based integrations, usually requires a paid plan such as Plus, Pro, Business, Enterprise or Edu. Availability may also depend on the user’s region, administrator settings and subscription to the external service. The current requirements for a specific integration should be checked directly in the ChatGPT app or plugin directory. Can ChatGPT be connected to a company’s internal system? Yes. An organisation can build a custom MCP server, dedicated integration or API connection that gives ChatGPT access to selected data and actions. A private system may remain behind a firewall or operate on-premises if the architecture uses a secure tunnel and controlled authentication. The project should define tool schemas, roles, logging, action approvals, error handling and protection against prompt injection. Before production deployment, the integration should be tested using valid requests, edge cases and tasks that it is expected to refuse. How do you connect an app to ChatGPT? First, define the task the integration should perform and the data required to complete it. Then open the app or plugin directory in ChatGPT, select the appropriate service and start the configuration process. Sign in to the external application, carefully review the requested permissions and approve only the access that is necessary. Once configuration is complete, open a new conversation, select the connected app and test it using a limited dataset. In a business environment, it is best to begin with a pilot for a small group of users before making the integration available to the wider organisation. Can a company administrator restrict access to apps in ChatGPT? Yes. A workspace administrator can decide which apps and plugins are available within the organisation, who may use them and which actions they can perform. For example, the administrator may allow read-only access while blocking message sending or CRM record updates. In managed workspaces, access can also be assigned according to user roles and groups. Integrations continue to respect permissions in the source system, so users should not gain access through ChatGPT to information they cannot view in the connected application. Does ChatGPT store copies of data retrieved from connected systems? It depends on how the integration works. With on-demand access, data is retrieved when a specific request is processed and is not indexed in advance. Integrations that use synchronisation may create an indexed copy of selected content to speed up searches and improve response quality. Disconnecting an app prevents further access, while its synchronised index is scheduled for deletion from OpenAI systems, typically within 30 days. Information previously used in conversations may remain in chat history, so removing it may also require deleting the relevant conversations and saved memories.
Read
Top 7 AI QA Tools for Pharma in 2026
In the pharmaceutical industry, test results become part of quality documentation and must be reproducible during an audit. A complete record includes links to requirements, execution details, change history, approvals and audit evidence. When selecting an AI-powered quality assurance tool for pharma, organisations should therefore consider test automation, traceability, data integrity and compliance with GxP requirements. The ranking opens with QATANA, a platform designed for comprehensive test process management. It combines AI capabilities, manual and automated testing, role-based access, audit logs and on-premise deployment. These features address the key needs of pharmaceutical QA teams by accelerating testing, maintaining control over data and supporting complete test documentation. The ranking covers seven solutions addressing different layers of the quality assurance process. Their capabilities include test management and traceability, end-to-end test execution, digital validation, visual testing and device labs. This comparison will help you select a tool suited to a specific system and validation model. Top 7 AI QA Tools for Pharma – Comparison at a Glance Rank Tool Main category Deployment model Best use case in pharma 1 QATANA AI-assisted test management On-premise Controlled testing lifecycle, auditability, and manual and Playwright tests managed in one environment 2 Tricentis Tosca + qTest + Vera Enterprise automation and digital validation Cloud, on-premise or hybrid, depending on the component Large CSV programmes, formal approvals and complex application environments 3 Opkey Enterprise application automation and continuous validation Cloud or on-premise Veeva, TrackWise, Oracle, SAP, Workday and other frequently updated systems 4 Leapwork No-code test automation and continuous validation Cloud, on-premise or hybrid Regression testing of business processes across web, desktop, Salesforce, SAP and Oracle systems 5 Applitools Visual AI and regulated content control Public cloud, private cloud or on-premise Product websites, portals, applications, eIFUs, PDF documents and mandatory safety communications 6 ACCELQ Full-stack no-code automation Public cloud, private cloud, on-premise or hybrid Omnichannel processes covering web, mobile, API, desktop and enterprise applications 7 TestGrid CoTester Agentic testing and device infrastructure Cloud, private cloud or on-premise device lab Mobile applications, patient portals and testing on real devices and browsers What Makes an AI QA Tool Ready for Pharmaceutical Applications? In a regulated environment, test case generation speed is one of several important selection criteria. The tool should support a controlled process in which requirements, risks, test cases, executions, defects and approvals form a consistent chain. Data integrity, decision traceability and the retention of evidence for the required period are equally important. Compliance with GxP, EU GMP Annex 11 and 21 CFR Part 11 depends on how the system is used, configured and maintained within a specific organisation. The assessment should cover procedures, roles, data, supplier qualification and risk analysis. AI capabilities can support this process, while computerised system validation confirms that the solution is fit for its intended use. We assessed AI-powered QA tools for pharmaceutical applications across six areas: Fit for pharmaceutical environments: capabilities, documentation and use cases in pharma, biotech, healthcare or life sciences. Traceability and audit evidence: links between requirements, tests and results, change history, roles, approvals, reports and data exports. Control over AI: review of AI-generated content, execution predictability, change management and human involvement in decision-making. Security and deployment: on-premises deployment, private cloud, data residency, communication with the AI model and access control. Technology coverage: manual, web, mobile, API, desktop, ERP, CRM and legacy application testing, as well as document and device testing. Operational scalability: integrations with Jira, CI/CD and automation frameworks, reporting, licensing models and the effort required to maintain tests. 1. QATANA QATANA combines AI capabilities with features that are particularly important in regulated QA environments. The platform centralises test cases, executions, defects, reporting, and the results of manual and automated tests. This enables teams to manage the testing process and documentation within a single controlled environment. AI generates draft test cases from tickets and requirements and helps select regression suites based on the scope of a given release. In a pharmaceutical environment, generated proposals should be reviewed by the people responsible for requirements, quality and risk assessment. QATANA supports this operating model because every AI-generated item remains an editable test artefact, while the team retains responsibility for its final assessment and approval. QATANA is particularly well suited to pharmaceutical companies that need a central test management system, want to keep data within their own environment and combine manual testing with Playwright automation. During a proof of concept, organisations should verify specific requirements for electronic signatures, retention, artefact versioning and export formats defined in their applicable SOPs. QATANA: for pharma: key facts Tool provider: Transition Technologies MS (TTMS) Website: ttms.com/ai-software-test-management-tool/ Solution type: AI-assisted test lifecycle management platform Key AI capabilities: Draft test case generation, intelligent regression selection, and analysis of ticket data and release information Best use case in pharma: Controlled test management for GxP and non-GxP applications, patient and HCP portals, internal systems and successive software releases Deployment model: On-premises, with the option to configure integration with the organisation’s selected AI model Integrations: Jira, Playwright, AI models and ticketing systems, as well as test artefact import and export Pricing: Custom pricing with a scalable multi-user licensing model What to verify before selection: Signatures and approvals required by SOPs, retention policies, versioning, evidence package exports and AI model governance rules 2. Tricentis Tosca, qTest and Vera Tricentis combines three complementary solutions: Tosca for test automation, qTest for test management and Vera for digital validation and process approvals. The platform supports more than 160 technologies and enables end-to-end testing of processes spanning ERP and CRM systems, web applications, APIs and data layers. The integration of Tosca, qTest and Vera supports requirements management, electronic signatures, formal approvals and the collection of evidence required for CSV. The solution can operate in the cloud, on-premises or in a hybrid model, with the latest agentic capabilities developed primarily for cloud environments. Tricentis for pharma: key facts Tool provider: Tricentis Website: www.tricentis.com Solution type: Ecosystem for enterprise automation, test management and digital validation Key AI capabilities: Agentic test creation from natural language, Tosca Copilot, portfolio and results analysis, and model-based test automation Best use case in pharma: Large CSV programmes, complex end-to-end processes, formal approvals and organisations using multiple enterprise applications Deployment model: Cloud, on-premises or hybrid, depending on the product and required AI capability Integrations: Tosca, qTest and Vera within one process, as well as popular enterprise applications, CI/CD pipelines, APIs, user interfaces and data layers Pricing: Custom pricing based on the selected products, number of users and execution scale What to verify before selection: Required licence scope, availability of AI capabilities in the selected deployment model, data flows and completeness of the validation package 3. Opkey Opkey automates testing for enterprise applications commonly used in life sciences, including Veeva Vault, TrackWise, Oracle, SAP, Salesforce and ServiceNow. Its AI engine analyses the impact of updates, generates test scenarios and automatically repairs tests following interface changes. The platform supports processes spanning multiple systems and provides pre-built libraries of business processes. For GxP applications, it offers IQ, OQ and PQ protocols, electronic signatures, traceability and automated collection of validation evidence. It is particularly well suited to organisations automating the validation of changes in widely used business applications. Opkey for pharma: key facts Tool provider: Opkey Website: www.opkey.com Solution type: No-code test automation and continuous validation for enterprise applications Key AI capabilities: Change impact analysis, test generation, self-healing, root cause analysis and intelligent regression scope selection Best use case in pharma: Validation of updates to Veeva, TrackWise, Oracle, SAP, Workday and Salesforce, as well as processes spanning multiple applications Deployment model: Cloud or on-premises, adapted to the customer’s infrastructure Integrations: Veeva, TrackWise, Oracle, SAP, Workday, Salesforce, Jira, Azure DevOps, qTest, Jenkins, ServiceNow and GitHub Pricing: Custom pricing; demo and test coverage assessment available What to verify before selection: Alignment of pre-built tests with the system configuration, validation protocol content, control over self-healing and maintenance costs following updates 4. Leapwork Leapwork enables visual, no-code automation for web, desktop, ERP, CRM and legacy applications. Its AI capabilities support test generation from natural language, requirements analysis and self-healing while maintaining deterministic scenario execution. The platform’s suitability for GxP environments is demonstrated by its implementation at NecstGen, where 110 workflows were automated and 270 functions within laboratory and quality systems were covered by a compliant process. Leapwork supports cloud, on-premises and hybrid deployment. When selecting a deployment model, organisations should verify the availability of AI capabilities, data processing location and the method used to approve changes proposed by the model. Leapwork for pharma: key facts Tool provider: Leapwork Website: www.leapwork.com Solution type: No-code test automation and continuous validation platform Key AI capabilities: Natural language test creation, self-healing, knowledge building from requirements and documentation, and coverage generation with traceability to source materials Best use case in pharma: Regression automation for web and desktop systems, Salesforce, SAP, Oracle and applications used by quality and operational teams Deployment model: Cloud, on-premises or hybrid Integrations: Playwright, Selenium, Cucumber, GitHub, CI/CD pipelines, test management systems, SAP, Oracle, Salesforce and Microsoft technologies Pricing: Annual subscription with custom pricing based on architecture and execution scale What to verify before selection: Availability status of AI capabilities, data processing location, human approval mechanisms and the ability to freeze a validated configuration 5. Applitools Applitools uses Visual AI to detect visual defects that conventional functional tests may overlook. It compares websites, application screens and PDF documents against approved baselines, identifying issues such as obscured messages, insufficient contrast and incorrect content placement. In pharma, it helps control risk information, instructions for use, regulatory messages and approved product content across devices, markets and language versions. Version history, screenshots, detected differences and approvals create an evidence set that supports QA and compliance teams. Applitools is particularly effective as a visual validation layer supporting functional testing and CSV processes. Applitools for pharma: key facts Tool provider: Applitools Website: www.applitools.com Solution type: Visual AI, visual, functional and cross-browser testing Key AI capabilities: Deterministic visual comparison, detection of significant changes, difference grouping, visual element-based self-healing and root cause analysis Best use case in pharma: Control of approved content, warnings, eIFUs, PDFs, product portals, patient applications and digital accessibility Deployment model: Public cloud, private cloud or on-premises Integrations: Playwright, Cypress, Selenium, Appium, more than 50 frameworks, Jira and popular CI/CD tools Pricing: Free trial; Starter and Enterprise plans priced individually What to verify before selection: Baseline approval rules, retention of screenshots and detected differences, language version support, audit package exports and the scope of accessibility testing 6. ACCELQ ACCELQ is a no-code platform for testing web, mobile, API and desktop applications, as well as enterprise systems. Its AI capabilities support scenario design, change impact analysis, self-healing and automation maintenance. The platform can test processes spanning multiple systems, such as portals, APIs, Salesforce, SAP and Oracle. SaaS, private cloud, on-premises and hybrid deployment models allow organisations to align the architecture with their data processing policies. ACCELQ for pharma: key facts Tool provider: ACCELQ Website: www.accelq.com Solution type: Unified no-code platform for test management and full-stack automation Key AI capabilities: Scenario generation, process modelling, change impact analysis, self-healing and AI-assisted automation maintenance Best use case in pharma: End-to-end processes spanning web, mobile, API, desktop, backend, Salesforce, SAP, Oracle and other enterprise applications Deployment model: Public cloud, private cloud, on-premises or hybrid Integrations: Jira, Azure DevOps, Jenkins, GitHub, GitLab, TeamCity, Bamboo, Salesforce, SAP, Oracle and Workday Pricing: Annual subscription with custom pricing; a 14-day free trial is available What to verify before selection: Validation documentation package, signatures and approvals, complete AI data flow, and the cost of private cloud or on-premises deployment 7. TestGrid CoTester TestGrid combines the CoTester agent with a cloud of real devices and browsers and a private device lab. AI generates tests from requirements or an application URL, updates them following interface changes and allows users to approve each scenario before execution. The platform supports web, mobile, API, visual and performance testing, as well as existing Selenium, Appium, Cypress and Playwright test suites. In pharma, it can support the testing of patient portals, therapeutic applications and solutions used in clinical trials on real devices. For on-premises deployments, organisations should note that the AI capabilities require a connection to TestGrid’s hosted infrastructure. TestGrid CoTester for pharma: key facts Tool provider: TestGrid Website: www.testgrid.io Solution type: Agentic testing, test management, and cloud or on-premises device lab Key AI capabilities: Test generation from requirements, conversational editing, AgentRx self-healing, error summarisation and results analysis Best use case in pharma: Mobile and web applications, patient portals, field solutions and testing on real devices and browsers Deployment model: Cloud, private cloud or on-premises device lab; AI capabilities may require an outbound connection Integrations: Jira, Jenkins, GitHub Actions, GitLab, Azure DevOps, Selenium, Appium, Cypress and Playwright Pricing: Starter plan from USD 199 per user per month, based on pricing available in August 2026; Growth and on-premises plans are priced individually What to verify before selection: Scope of data sent to the hosted AI service, data residency, log immutability, retention policies and the ability to operate without external connectivity AI-generated image. The people depicted are fictional. Before Implementing an AI QA Tool in Pharma: 9 Questions to Ask the Vendor Before selecting a tool, conduct a proof of concept under conditions that closely reflect the actual testing process. This allows you to assess test creation and execution speed, documentation completeness and the ability to reconstruct the entire process during an audit. Before selecting and implementing an AI QA tool in a pharmaceutical company, ask the vendor the following questions: Does the platform connect requirements, test cases, test executions and reported defects? Which activities and changes are recorded in the audit logs? Can roles and permissions be configured according to the organisation’s procedures? Can AI-generated test cases be reviewed, edited and approved before use? Are manual and automated test results available in one consistent view? How does on-premises deployment work, and how can the platform connect to an AI model selected by the organisation? Does the platform integrate with the organisation’s existing tools, such as Jira, Playwright and CI/CD pipelines? Can test data, reports and other artefacts be imported and exported in the required formats and scope? How do licensing, deployment, integrations, training and ongoing support affect the total cost of the solution? Best AI QA Tool for Pharma: Final Recommendation The final decision should reflect the intended use, risk assessment and a proof of concept conducted on a representative process. The solution provider also plays an important role, as its experience affects implementation quality, change management and the audit readiness of the testing process. TTMS, the provider of QATANA, has worked in the pharmaceutical industry since 2011, involving more than 400 specialists in over 100 projects and services. The company combines QA engineering with expertise in quality management and computerised system validation in line with GAMP 5 and EU GMP Annex 11. These capabilities are supported by the TTMS Integrated Management System, which includes ISO 9001 and ISO 27001. This enables TTMS to support the entire implementation lifecycle, from requirements definition and tool configuration to validation, maintenance and controlled change management. FAQ Is a “21 CFR Part 11 compliant” claim sufficient when selecting an AI QA tool for pharma? No. Such a claim usually describes the available features or the way the product has been designed, while compliance is assessed for a specific intended use and implementation. The organisation must determine which electronic records and signatures fall within scope, configure roles, permissions, audit trails, retention policies and procedures, and then demonstrate that the system is fit for its intended use. Integrations with Jira, CI/CD pipelines, code repositories and other systems are also important because data flows may extend beyond the QA tool itself. Vendor documentation can facilitate validation, but responsibility remains with the pharmaceutical company. A proof of concept should therefore include the reconstruction of a complete evidence chain from the original requirement to the approved test result. How should AI-generated test cases be validated in pharma? An AI-generated test case should be treated as a draft requiring expert review. A person familiar with the requirement and its associated risk should verify the preconditions, test data, steps, expected results, negative scenarios and traceability to the source requirement. The system should record the source, model version, generation date, approver and all subsequent changes. Functionality with a greater potential impact on product quality, patient safety or data integrity requires more rigorous review and independent approval. AI performance should be evaluated against a controlled reference set using measures such as coverage completeness, the number of rejected suggestions and errors identified during review. This approach preserves the time-saving benefits of AI while keeping accountability with qualified personnel. Are self-healing tests safe in a validated GxP environment? They can be used when the mechanism operates in a controlled manner and maintains a complete record of every change. Automatically correcting a technical locator can reduce false failures, provided that the repair does not alter the meaning of a step, the acceptance criterion or the scope of the test. A well-configured system displays the proposed change, its rationale, and the previous and new values, and requires approval for significant modifications. The organisation should define in its SOPs which repairs may be accepted automatically, which require review and when a test must be reapproved. False positives, false negatives and the effects of self-healing engine updates should also be reviewed periodically. Execution repeatability and decision traceability are more important than the number of tests repaired without tester involvement. Can production data from pharmaceutical systems be sent to an external AI model? The preferred starting point is to use synthetic, anonymised or masked data limited to the minimum required for testing. Sending production data requires a legal basis, information classification assessment, vendor agreement, transfer controls, retention rules, processing location controls and clear policies regarding model training. Patient data, clinical trial information, safety data and confidential product information require particular protection. An on-premises deployment may still rely on a hosted AI service if an agent or interface communicates with an external model. The architecture should therefore show separately where tests are stored, where automation is executed and where AI processing takes place. Access to sensitive data should be granted only after the vendor provides a clear and verifiable description of the complete data flow. What should be done after an update to the AI model used by a QA tool? A model update should be managed as a controlled change, with the scope of assessment determined by risk. The first step is to identify which functions use the model and whether the update could affect test generation, regression selection, self-healing, defect classification or reporting. A previously approved reference test set should then be executed, and the results compared with those produced by the earlier version. Any differences should be assessed, documented and approved before the updated model is used more broadly. The change record should include the model version, date, scope, assessment results, accepted limitations and the person responsible for the decision. When a vendor updates the model without offering the option to freeze a version, the agreement should define advance notification, a testing window and a rollback procedure. Effective model version control is essential for maintaining the validated state.
Read
Astra, the Future GPT-6? OpenAI’s New Model Explained
Solving mathematical problems that scientists had wrestled with for years – could there be a better demonstration of what a new AI model can do? OpenAI has typically previewed new versions of its large language models with benchmark results, meaning scores from standardised tests designed to measure a model’s capabilities. I have to admit that seeing GPT tackle genuine research problems makes a much stronger impression on me. What will you learn about OpenAI Astra? What Astra is and why it is being discussed as a potential GPT-6, 10 results in mathematics and theoretical computer science presented by OpenAI, How Astra analyses problems, tests hypotheses and changes its approach, The differences between a conversational model, an AI agent and a system capable of managing an entire project, What Astra could mean for science, business and the future of AI models, Critical responses to the model’s achievements, Cybersecurity risks associated with autonomous AI agents, Which important questions OpenAI has yet to answer. What is OpenAI Astra, and could it become GPT-6? OpenAI describes Astra, the prototype’s working name, as “our next major model”, although the company has disclosed very few details so far. Its task was to develop arguments independently, test hypotheses, recognise unproductive approaches and find new paths towards a solution. The results of its work can then undergo formal and independent verification. We do not know how Astra is built, how much information it can analyse at once or how it organises its work on a complex task, although we can speculate about the last of these. OpenAI has also not disclosed whether Astra is a single model, a team of collaborating AI agents or a more extensive system equipped with mechanisms for coordinating their work and retaining previous results. The prototype may be connected to a model previously described by OpenAI as capable of operating autonomously over very long periods. Such a system can make repeated attempts, analyse intermediate results and maintain its direction of work for many hours, potentially even days. According to media reports, Sam Altman has already presented Astra to US politicians and regulators. The term “GPT-6 Astra” should therefore be treated as media shorthand. Astra could eventually be released as GPT-6, another version of GPT-5 or a separate family of models. For now, all of these possibilities remain open. Why could Astra’s 10 results matter more than another benchmark record? OpenAI presented ten results concerning problems that had remained open for at least a decade and, in most cases, considerably longer. The problems come from eight fields: high-dimensional geometry, coding theory, group theory, operator algebras, computational complexity theory, quantum computing, lattice geometry and post-quantum cryptography, extremal combinatorics. In simple terms, the process worked as follows: GPT generated mathematical arguments. Once the results had been obtained, researchers worked with the model to develop them into scientific papers. The system then translated the arguments into Lean 4, allowing a computer to check every step of the proofs. For readers interested in the technical details, here are the relevant links: the complete collection of papers, the Lean formalisation repository and reconstructions of how the solutions were developed. Independent verification of all the claims by the scientific community is only beginning. Mathematicians can now review the papers, check the definitions, run the formalised proofs and look for potential gaps. I discuss this in more detail in one of the final sections. 10 new results from Astra in mathematics and theoretical computer science A quick warning: this section is about to become fairly technical. These subjects are new, abstract and extraordinarily difficult for me as well, so I have tried to explain each result in the simplest possible terms. Here is how GPT Astra approached the individual problems. 1. Sphere packing in high-dimensional spaces The sphere-packing problem asks how densely identical spheres can be arranged, much like coins on a table or balls in a box. Mathematicians also study this question in spaces with hundreds or thousands of dimensions because it has applications in areas such as information theory and data encoding. Astra used an established mathematical method to determine more precisely how densely spheres can be packed in spaces with a very large number of dimensions. According to the authors, this is the first improvement since 1978 to the value used in the formula describing how quickly the possible packing density decreases as the number of dimensions increases. The difference becomes more significant as the number of dimensions grows and enables a more precise estimate of the maximum packing density. Put simply, Astra’s calculations improve our understanding of how many spheres can fit inside such a “high-dimensional box”. 2. Binary and spherical codes: new bounds on the number of error-resistant codes A binary code is a set of sequences made up of zeros and ones. These sequences must differ from one another sufficiently for a system to detect and correct transmission errors. This can be compared to positioning transmitters at safe distances from one another so that their signals remain easy to distinguish. Astra determined more precisely how many codes can be placed sufficiently far apart for a system to continue distinguishing between them and correcting errors. This enables mathematicians to estimate more accurately how many codes with the required level of error resistance can fit within a given space. The model tested its initial idea on a simple example consisting of eight digits and discovered that it produced an incorrect result. It therefore abandoned that approach and reformulated the problem. This case demonstrates Astra’s ability to test its own assumptions and redesign its solution when the original direction proves unsuccessful. OpenAI’s published materials support four important conclusions. 3. The first explicit example of a non-sofic group A group is a mathematical way of describing symmetries and operations that can be performed in sequence, much like a set of moves used to rotate a Rubik’s Cube. Sofic groups can be approximated with arbitrary precision using simpler structures based on a finite number of elements. For decades, mathematicians wondered whether this property applied to every group. Astra identified a specific example of a group that cannot be approximated with arbitrary precision using simpler models composed of a finite number of elements. The result demonstrates that these simplified models cannot represent every mathematical group. The solution combined several distant areas of mathematics, demonstrating the model’s ability to bring together tools that had not previously formed an obvious path towards a proof. 4. Disproving Connes’ rigidity conjecture The von Neumann algebra associated with a group can be compared to its highly complex mathematical “fingerprint”. Connes’ conjecture proposed that, for a certain class of particularly rigid groups, this fingerprint uniquely identifies the group in question. Astra constructed infinitely many different groups with exactly the same mathematical “fingerprint”. In doing so, it disproved Connes’ conjecture and answered a later question posed by mathematician Sorin Popa. Astra used a mechanism resembling the carrying operation in binary addition. This made it possible to construct many different groups with the same mathematical “fingerprint”. Put simply, Astra demonstrated that a single mathematical “fingerprint” can belong to infinitely many different groups. 5. The matrix permanent: the minimum number of operations required for its computation The permanent of a matrix is calculated in a similar way to the determinant, except that all terms are added with a positive sign. This seemingly minor change makes the permanent one of the most important examples of a problem with extremely high computational complexity. Astra determined the minimum number of basic operations required to calculate the permanent. It proved that no solution within this class can be simplified below a certain level of complexity. This can be compared to determining the minimum number of components required to build any machine capable of performing a particular task. Such a proof must cover every possible construction that meets the specified conditions, which makes it exceptionally difficult to develop. The result provides a more precise lower bound on the number of operations needed to solve this problem. It also brings mathematicians closer to answering a fundamental question: which problems can be solved efficiently, and which will always require an enormous amount of computation? 6. Quantum games: why does the probability of a perfect win decrease so rapidly? Imagine a game in which two players answer a referee’s questions separately, while their shared goal is to complete every round successfully. In the classical version, each additional round rapidly reduces the probability of a perfect win, much like repeatedly tossing a coin reduces the chance of getting heads every time. In the quantum version, the players’ results can be correlated even when they do not communicate during the game. They can also analyse several rounds as a single combined problem. Astra proved that even such quantum correlations cannot prevent the probability of winning every repeated round from decreasing very rapidly. The problem had remained open since at least 2004. The key to the solution was a method for transforming quantum states without changing the probabilities of their possible outcomes. The result advances the theory of interactive proofs, quantum information theory and methods for increasing the reliability of protocols. 7. The Closest Vector Problem: even an approximate solution remains difficult A lattice can be imagined as a regular grid of points, similar to street intersections in a perfectly planned city, extending across many dimensions. The Closest Vector Problem (CVP) involves finding the point on this grid that lies closest to a selected location. It is highly relevant to geometry, coding theory and post-quantum cryptography. Astra connected CVP with the well-known 3SAT logic problem and demonstrated that finding even a solution that merely approximates the optimal one is extremely difficult. This difficulty increases with the number of dimensions in the lattice. The model represented the logical puzzle as a system of points and distances between them. This can be compared to encoding a complex logic puzzle in a spatial arrangement of points so that solving one problem also provides a solution to the other. The result deepens our understanding of the theoretical difficulty of lattice-based mathematical problems. Assessing the security of specific cryptographic algorithms requires a separate analysis of their variants, parameters and methods of data generation. 8. Proving Ehrhart’s conjecture on the volume of high-dimensional shapes A high-dimensional convex body can be imagined as a solid placed on a regular lattice of points, with its centre of gravity being the only lattice point located inside it. Ehrhart’s conjecture specified the maximum possible volume of such a body, and Astra proved it for any number of dimensions: vol(K) ≤ (n+1)n / n! The main difficulty was connecting the number of individual lattice points with the volume of the entire body. The first approach provided only part of the information required. Astra therefore reformulated the problem in the language of another branch of geometry and began searching for a solution using its tools. The model combined several advanced methods for describing the body’s shape, its boundaries and the distribution of points. Put simply, Astra translated the geometric puzzle into a different mathematical language in which it became possible to determine the exact volume bound. 9. Multicolour Ramsey numbers A complete graph can be imagined as a group of people in which every pair is connected by a line, with each line assigned one of k colours. Mathematicians ask how large such a network must become before it inevitably contains three people whose connecting lines are all the same colour. This minimum size is denoted by Rk(3). Astra developed new colouring methods which, when combined with previous results, established the growth rate of this number: Rk(3) = kΘ(k). The result does not provide an exact value for every number of colours, but it reveals the correct scale of growth. In doing so, it resolves Erdős Problem No. 183. Astra expanded the network in stages according to the same rule. This made it possible to construct increasingly large configurations without creating a triangle whose edges were all the same colour. 10. Two counterexamples in extremal graph theory An extremal number determines how many connections a network can contain before a specified forbidden configuration inevitably appears. Astra disproved two conjectures proposed by Erdős and his collaborators concerning how this value could be predicted. In the first case, it constructed a family of graphs in which forbidding each member individually still allowed approximately n4/3 edges, while applying all the restrictions simultaneously reduced the maximum number to O(n21/16). This shows that several forbidden structures can constrain a graph far more strongly together than when each is considered separately. In the second case, Astra found a network divided into two groups in which every small section contained few connections, while the complete construction could be considerably denser than the conjecture predicted: ex(n, H) ≥ cn3/2+ε. The two results resolve Erdős Problems No. 146 and 180. They also show that the simple structure of small sections of a network does not always allow us to predict how dense the entire construction can become. A critical perspective: how do experts assess Astra’s mathematical achievements? After the initial excitement, important reservations began to emerge. Mathematicians pointed out that at least two of Astra’s results rely heavily on earlier work, raising questions about their novelty. OpenAI has since changed the way it describes the experiment. It now increasingly refers to “making meaningful progress”, rather than solely to “solving longstanding problems”. Interestingly, a researcher affiliated with Anthropic reported that the Claude Fable model had reproduced solutions to five of the ten problems tackled by “GPT-6” within 24 hours, although these results have yet to be fully verified. This does not undermine Astra’s capabilities, but it makes it more difficult to determine whether we are witnessing a breakthrough driven by the exceptional abilities of one model or broader progress across AI models as a whole. Above all, there is still no reliable, independent and fair comparison conducted using the same problems, prompts, computational budgets and rules governing access to tools. What do the results reveal about how Astra works? OpenAI’s published materials support three important conclusions. 1. The model can abandon dead ends The published reconstructions show Astra trying different approaches, identifying obstacles, reformulating problems and returning to earlier stages of its work when necessary. This resembles genuine research more closely than an extended answer generated in a single pass. In the binary-codes problem, the first recurrence was rejected after the model found a small counterexample. When working on Ehrhart’s inequality, Astra spent considerable time developing an approach based on symmetrisation before reformulating the problem in terms of toric geometry. In the proof concerning quantum games, it recognised that the classical argument lost control after conditioning on rare events and began searching for a representation that preserved quantum probabilities. The published document does not reveal the model’s complete internal reasoning process. It is a narrative produced by a model that reviewed the original reasoning traces and the final papers. 2. GPT Astra combines discovery with automated verification Lean checks the correctness of a formal proof step by step. The repository contains separate files for all ten results, along with instructions for performing additional checks of the formalised proofs. Computer verification does not replace assessment by independent mathematicians. Researchers must still establish, among other things: whether the formal theorem corresponds precisely to the original problem, whether the definitions introduce any unintended simplifications, whether the result is genuinely new, how significant it is for the relevant field, whether the manuscript correctly connects the formalisation with the informal argument. A computer can confirm that a written proof is logically correct under the adopted definitions and assumptions. It does not automatically confirm that the authors formalised precisely the version of the problem that mathematicians intended to address. The research was published on 1 August 2026, so full independent verification by the mathematical community will take time. Thomas Bloom of the University of Manchester nevertheless described the results as “big news” and rated the significance of the presented constructions particularly highly. 3. The cost of Astra’s results and the importance of additional computing power OpenAI claims that, based on the API pricing for GPT-5.6 Sol, the tokens required to find all ten solutions would have cost approximately $2,000. This figure is, of course, neither the actual cost of developing Astra nor the full cost of the project. It does not include model training, infrastructure, researchers’ work, problem selection, validation or all the unsuccessful attempts. It is simply the cost of the tokens used to find the published solutions, calculated according to current API pricing. The average comes to approximately $200 per published result, but we do not know: the total number of problems presented to the model, the success rate, how the costs were distributed across the problems, how long the system operated, how many agents were involved, how many runs were conducted in parallel. Noam Brown, an OpenAI researcher involved in the work on Astra, acknowledged that the system had also been tested unsuccessfully on other major mathematical challenges, including the Millennium Prize Problems. He added that OpenAI had not allocated an especially large amount of computing power to each problem. The company therefore believes that Astra could achieve better results if given more time and resources to search for solutions. From GPT-5.6 to Astra: how AI is moving from answering questions to managing projects GPT-5.6 already includes several features that point towards the direction described above. The model can independently select tools, analyse the results it obtains and use them to plan its next actions. Ultra mode uses four agents by default, while OpenAI has also tested configurations involving sixteen agents. The company also offers a multi-agent mode in the Responses API in beta. Astra may develop this architecture towards much longer and more coherent periods of autonomous operation. The most important difference would be its ability to manage an entire project over many hours or days. The system would need to remember what it had already tried, which ideas it had rejected, what results it had obtained and how the individual tasks related to one another. From the user’s perspective, the change could be very tangible. Instead of guiding the model through a sequence of prompts, the user gives it an objective, a set of available tools, a defined scope of permissions, a budget and completion criteria. The user then returns to a finished result accompanied by a record of the attempts, tests and decisions made along the way. This progression can be presented as three successive units of work: A conversational model generates an answer. An agent completes a task using tools. A multi-agent system manages a project in which tasks are created and modified as the work progresses. Only the technical documentation will show whether Astra genuinely operates at the third level as a coherent system. The mathematical demonstration is, however, the first strong indication that this direction is becoming more than a promise. OpenAI, Google DeepMind and Anthropic: the race to develop long-horizon AI models Google DeepMind, Anthropic and OpenAI are developing AI systems capable of independently handling increasingly long and complex tasks. Aletheia, Google’s mathematical agent based on Gemini Deep Think, can generate solutions, verify their correctness and revisit them when it detects an error. When analysing 700 Erdős problems, it solved four questions that had previously remained open. Anthropic, meanwhile, is focusing on coordinating the work of multiple agents. According to the company, Claude Opus 4.8 can divide a large project into smaller parts and assign them to hundreds of subagents working in parallel. This allows it to carry out tasks such as migrations involving hundreds of thousands of lines of code. Claude Science, another environment being developed by the company, is intended to make it possible to trace and verify the successive stages of research work. All these projects point in the same direction: models are expected to work towards a single objective for longer, monitor their own results and revise earlier decisions. Astra stands out for producing results at the frontier of contemporary knowledge and for formally encoding some of its proofs, allowing their correctness to be checked by a computer. How could Astra change the AI model and agentic tool market? 1. Benchmarks may lose their role as the primary evidence of AI model quality Competition will increasingly focus on the final outcome: a new hypothesis, a discovered vulnerability, a completed system migration, a developed scientific model, a working application, a result that can be verified automatically. Astra was presented through its scientific results because conventional benchmarks do a poor job of communicating the difference between a model that answers a question and a system that manages an entire project. Benchmarks will remain necessary for comparing models under controlled conditions. Their market significance may, however, decline in favour of evaluations that measure project completeness, operational continuity and the quality of the final result. 2. The cost of a completed task may matter more than the price per token For business customers, the following factors will become increasingly important: the cost of completing the project, the time required to obtain the result, the probability of success, the number of human interventions, the cost of validation, the ability to resume work after an error. Approximately $2,000 for the tokens that produced ten published results is a powerful economic signal, even with all the reservations concerning how the results were selected. Perhaps future pricing models will include offers such as the “cost of a successfully completed migration”. This would, however, require transparent information about the number of failed attempts, the additional human work involved and the cost of verifying the result. 3. Astra could affect platforms that coordinate AI agents If models begin dividing work among agents, retaining a record of their progress and combining the resulting outputs themselves, some external platforms used to coordinate AI agents may lose part of their current advantage. The value of these platforms will primarily come from integration with company systems, access and cost controls, visibility into agent activity and the ability to audit the entire workflow afterwards. Dividing a task among several agents may become a standard model capability. The implementation layer will derive its value from embedding agents securely within organisational processes. 4. Keeping AI agent workflows under full control When a task lasts several days, the system must regularly save its progress and check whether subsequent actions are still leading towards the defined objective. OpenAI acknowledges that long-running work requires monitoring the model’s complete sequence of actions. Each individual step may appear safe even though, taken together, the steps lead to an unacceptable outcome. The company described a case in which a model attempted to bypass a safeguard by splitting a digital access key into separate parts and reassembling them only when the program was launched. Each step, considered in isolation, could appear relatively harmless, while the objective of the complete sequence was unacceptable. In another experiment, the model was instructed to deliver its results through Slack. Instructions found in the code it was analysing suggested publishing the proposed changes directly to GitHub instead. The model spent approximately an hour searching for a way around the test environment’s restrictions before finding a vulnerability and publishing the result there. Earlier models had abandoned the task under similar circumstances. The case shows that a longer operating horizon increases both the model’s effectiveness and the amount of time it can devote to searching for ways around safeguards. In response, OpenAI developed a system that analyses the model’s entire sequence of actions as its work progresses. The system can stop the session, notify the user and ask them to decide whether the work should continue. Further details are available in OpenAI’s report on the safety of long-horizon models. 5. Research fields in which Astra could accelerate progress The most immediate impact is likely to appear in fields with: precisely defined problems, extensive available literature, formal or automated verification tools, the ability to conduct computational experiments, unambiguous criteria for measuring progress. Mathematics is an ideal testing ground because a proof can be verified. Similar conditions exist in software development, chip design, some areas of chemical research, bioinformatics and cybersecurity. Economics, strategy, law, management and social research will remain much more challenging because correctness cannot be reduced to a machine-verifiable certificate. In these fields, a model may produce an impressively coherent project that is still based on flawed assumptions or a poorly defined objective. 6. Long-horizon models will require more computing power An important capability of a model will be the option to allocate more computing power and more attempts to particularly difficult problems. This will give an advantage to laboratories with: extensive computing resources, efficient communication between agents, effective context management, automated detection of dead ends, the ability to run multiple attempts and select the best result. The next stage of competition may concern more than model size. It may also depend on how effectively models use time and computing power when working on a specific task. The same model could operate as a relatively inexpensive assistant for everyday questions and as a costly research system when the user increases the budget for time, agents and parallel attempts. The section likely to age quickly: what do we still not know about Astra? OpenAI has not disclosed basic information about Astra, including its architecture, size, method of agent collaboration, memory mechanism or capabilities beyond mathematics. We also do not know its price, release date or whether OpenAI plans to make the model available through ChatGPT or the API. The published results do not demonstrate that Astra selected the problems independently, operated without supervision or can manage an entire research process. Nor do we know whether it can achieve similar results in other fields. There is therefore no basis for describing Astra as a system that matches human capabilities across a broad range of intellectual tasks. We also do not know the total number of failures. OpenAI published selected successes, while Noam Brown confirmed that the system had attempted to solve other major problems without success. Without knowing the total number of attempts, it is impossible to calculate Astra’s actual success rate or the expected cost of obtaining one valuable result. Why is Astra not yet an autonomous scientist? The published papers show a system solving problems selected and presented by humans. An autonomous scientist would also need to: select research directions independently, assess which questions are important, determine whether a result is genuinely new, design subsequent experiments, decide when sufficient evidence has been collected, place the result within the broader context of the field. Astra completed the most technically demanding part of this process: it developed new arguments and brought them to a form that could be formally verified. This is a major achievement, but it does not encompass the full scope of scientific work. Can Astra succeed beyond mathematics and controlled environments? The ten published papers demonstrate what Astra was able to achieve in a carefully selected environment. Mathematics offers clearly defined problems, extensive literature, precise language and formal verification tools. The real test will be whether this capability can be transferred to projects in which the objective changes as the work progresses, tools fail, data is incomplete and the correctness of the result requires human judgement. If Astra can maintain a coherent process over many hours or days, delegate subtasks, retain the results of previous attempts and return to a problem after detecting an error, the change will be more significant than another increase in benchmark scores. Models such as Astra demonstrate how rapidly the capabilities of artificial intelligence are advancing. In business, their value depends on selecting the right process, ensuring data quality, integrating AI with company systems and maintaining control over its operation. TTMS helps organisations design and implement solutions tailored to specific operational needs. Explore TTMS AI solutions for business and implementation examples. How autonomous was Astra when solving mathematical problems? OpenAI states that the mathematical arguments were generated by the system, while humans contributed to preparing the manuscripts, formalising the results and verifying their correctness. The papers list OpenAI as the author, and the company has not attributed individual proofs to specific employees. This creates an interesting precedent: the organisation assumes responsibility for the publications while crediting the model with producing the arguments. However, it remains unclear who selected the problems, prepared the prompts, initiated subsequent attempts and decided which results were suitable for publication. Without this information, it is difficult to determine Astra’s precise level of autonomy or distinguish the capabilities of the model itself from the work of the wider research team. Can artificial intelligence be the author of a scientific paper? Authorship involves responsibility for the research method, the evidence presented, the conclusions and any potential errors. An AI system cannot formally accept such responsibility, so researchers should remain the authors of scientific publications. The model’s contribution should be described clearly in the methodology, including how it was used and which elements of its work were verified by humans. How can researchers verify whether AI has made a genuinely new discovery? A correct result is not necessarily a new one. Researchers must compare it with the existing literature, previously unpublished work and known variants of the same problem. One particular challenge is determining whether the model developed a new solution or reproduced a relationship contained in its training data. Novelty should therefore be assessed separately from the correctness of the proof itself. Can a result produced by a closed AI model be reproduced? Reproducing an experiment is difficult when researchers do not know the model’s architecture, training data or exact settings. Recording the prompts, system version, tools used, intermediate results and human interventions can make the process more transparent. The final result should also be verifiable using a method independent of the model that generated it. Without this documentation, other scientists may be able to verify the result itself, but not the full process that led to it. Could AI agents increase the risk of errors and unreliable scientific publications? An AI agent can generate large numbers of convincing hypotheses, proofs and interpretations of data in a short time. This scale can accelerate research, but it can also spread flawed assumptions more quickly. Academic journals and research institutions will need clear rules for disclosing the use of AI, preserving a record of the research process and independently verifying the most important results. The transparency of the process will become as important as the quality of the final publication. How should a research team prepare to work with AI agents? A good starting point is to select tasks with results that can be verified unambiguously. The team should determine which data and tools the agent can access, which actions require human approval and who is responsible for accepting the final result. It should also establish procedures for recording each stage of the work, reporting errors and stopping an experiment when necessary. This preparation allows researchers to benefit from the speed of AI while maintaining control over the quality of the research.
Read
GPT-Powered AI Agents: How to Match Autonomy to the Process?
Until recently, enterprise automation followed a simple division: systems performed tasks defined by rules, while cases requiring interpretation were passed to people. GPT-powered AI agents expand the range of processes that can be supported through automation. They can work with documents, incomplete data and the language used by customers or employees, making them suitable for processes that were previously difficult to automate. For large organisations, this raises a practical question about AI agent autonomy: where does expert support end, and where does independent action within a process begin? In some situations, the agent’s role is to gather information and prepare a recommendation. In others, it prepares an action for approval. There are also areas where it can independently carry out repetitive steps when the organisation has defined the rules, permissions, limits and exception-handling paths. GPT-powered AI agents can already support teams with ticket handling, document analysis, decision preparation, data updates and multi-step tasks. The key implementation question is: which decisions and actions should remain with people, and which can an agent perform within agreed rules? An AI agent in the enterprise is a process participant, not just a chatbot In practice, a GPT-powered agent needs five elements: access to reliable sources of knowledge, a clearly defined business objective, tools and integrations with enterprise systems, permissions aligned with its role, rules that define the boundaries of its actions. A language model can interpret the content of a document, a customer message or an incident description effectively. It does not, however, replace a business process. Workflows, permissions, validations and decision history are what make an agent operate predictably, even when it handles hundreds or thousands of cases each month. Three levels of AI agent autonomy In a large organisation, it is worth designing agents across three levels. This allows autonomy to grow alongside process maturity and trust in the solution. Operating level Agent’s role Example tasks Human role Level 1: Advisory agent Analyses information and prepares a recommendation. Case summary, risk identification, proposed response, ticket prioritisation. Makes the decision and carries out the action. Level 2: Agent preparing an action for approval Completes the next steps in a process, stopping before actions with significant consequences. Creates an application, updates data, prepares a communication, submits an instruction for approval. Reviews and approves specified steps. Level 3: Agent performing tasks automatically Independently carries out tasks in line with the process policy. Case classification, status updates, sending standard information, creating a task in a system. Handles exceptions, monitors quality and updates process rules. The level of autonomy does not need to apply to the entire agent. The same agent may independently classify tickets, prepare a response that requires approval and transfer unusual cases to an expert. In practice, an organisation therefore designs autonomy for individual decisions and actions, rather than choosing a single operating model for the whole solution. What determines whether an AI agent can complete a task independently? A useful starting point is to assess two factors: the impact of the action on the organisation and whether it can be reversed. The greater the business, legal, financial or reputational consequences of a decision, the more important human approval becomes. Nature of the action Recommended model Low impact, simple rules, easy to reverse Automatic execution with a record in the process history. Medium impact, data from several sources, possible exceptions The agent prepares the action and an authorised person approves it. High financial, legal or customer impact The agent presents analysis, options and justification. The decision remains with a person. Unclear rules, incomplete data or conflicting information Automatic escalation to an expert, together with the context and collected data. This principle is particularly useful in organisations operating across multiple countries, with complex permission structures and a large number of systems. Just as important as the list of tasks is knowing what the agent must not do and when it should hand a case over to a person. 7 questions to ask before giving an AI agent permission to act What action should the agent perform? Describe it specifically, for example: “create a service ticket”, “update contact details” or “prepare a response to a complaint”. What data will it work with? Identify the sources, data owners, update frequency and access rules. What business rules must it follow? These may include financial limits, contractual terms, SLA levels, compliance requirements or communication policies. What exceptions should stop the process? The agent needs a clear escalation path for unusual or incomplete cases, or those requiring specialist assessment. Can the action be reversed? The ease of correction affects the appropriate level of autonomy, the scope of testing and the need for additional approval. Who is accountable for the decision? The process owner, approver and technical team should all have clearly assigned roles. How will the organisation establish why the agent took a particular action? The case history should show the input data, rules, sources used, recommendation and process outcome. This is why AI agent projects often begin with bringing the process itself into order. The organisation gains more than a new AI capability: it also gains better visibility of responsibilities, exceptions and how work actually flows. Where can GPT-powered AI agents add value in a large enterprise? Customer service and back-office teams An agent can read a customer message, identify its subject, retrieve data from a CRM or case-management system, prepare a response in line with company policy and route it to the appropriate queue. For standard cases, it can also update a status, create a task for the team or send the customer a confirmation. Full autonomy works well for low-risk actions, such as providing information about the status of a ticket. Complaints, individual commercial terms or cases requiring interpretation of a contract should be passed to an employee together with the agent’s analysis. Finance, procurement and document workflows An AI agent can read a document, check whether the data is complete, compare it with a purchase order and flag discrepancies that require clarification. It can also prepare a case summary, collect missing information and initiate the appropriate approval workflow. Decision thresholds are particularly important in this area. The agent can process a document automatically when it meets all conditions, while cases that exceed a defined amount, contain discrepancies or concern a new supplier can be submitted for approval. IT, administration and ticket management In an IT environment, an agent can classify tickets, create an incident summary, search for similar cases in the knowledge base, propose actions in line with a runbook and update the user on progress. In administrative processes, it can prepare an application, complete data in a form and remind the requester about missing documents. For actions involving configuration changes, access permissions or production systems, an approval-based model is advisable. The agent reduces the time needed to prepare a decision, while the administrator retains control over the change. Sales and commercial information management An agent can prepare a briefing before a meeting by bringing together information from the CRM, proposals, correspondence and notes, then highlighting open points and suggested next steps. After the meeting, it can create a summary, propose data updates and prepare tasks for the team. These are extensions of scenarios already familiar from everyday work with generative AI. Read more about what the current generation of models helps teams achieve in our article: GPT-5.6 from OpenAI: capabilities and business applications. Why does an AI agent need a workflow? An AI agent can interpret information and suggest next steps, but the process should define the sequence of actions, required validations and the people responsible for approval. In a large organisation, this is what determines the repeatability and scalability of the solution. A process automation platform can act as a control layer: it triggers a task, provides the agent with the necessary context, receives the result, records the history and routes the case to the next stage. The agent then becomes part of a controlled workflow rather than operating as a separate tool outside the core process. This approach is relevant to document workflows, request handling, HR processes, procurement and administration. See how WEBCON BPS can support the digitalisation and control of business processes, and how TTMS delivers process automation. Four forms of human oversight of an AI agent Human-in-the-loop is a model of control embedded in the process—from reviewing recommendations to handling exceptions and making decisions with greater impact. In a mature solution, people can play several different roles. Approving an action when the agent has prepared a specific instruction, communication or system change. Selecting an option when the agent has presented several possible solutions and their consequences. Handling an exception when a case falls outside the agent’s rules, available data or permissions. Overseeing process quality by analysing errors, rejected recommendations, completion times and changing business needs. The most effective implementations use all four forms. The team does not manually review every standard operation, yet retains full control over actions with greater significance and over the direction in which the process evolves. It is also worth observing whether human approval genuinely improves process safety or simply moves a bottleneck elsewhere. If an approver nearly always accepts the agent’s proposals without changes and the cases are easy to reverse, the organisation can consider automating the selected step. If recommendations often require correction or the approver needs to return to source data, this indicates that the process rules, quality of knowledge or scope of the agent’s permissions need attention. When can an AI agent act automatically? Automation delivers the most value when a task is frequent, has a repeatable structure, relies on available data and leads to a clearly defined outcome. It is also important to ensure that execution can be verified and corrected when data or rules change. Good candidates include ticket classification, routing requests to the appropriate queue, completing data from approved sources, creating standard tasks, updating statuses and sending communications based on approved templates. Combining GPT models with an enterprise knowledge layer, integrations and security rules provides a significant advantage. This allows the solution to work with information available to a specific role, rather than with an unstructured collection of documents and conversations. When should an AI agent primarily provide advice? An advisory role is especially valuable in cases that require contextual assessment, interpretation of company policy, negotiation, an individual approach to a customer or decisions with significant financial and legal consequences. In these situations, the agent can gather facts, summarise documents, identify missing information, compare options and prepare the rationale for a recommendation. The person gains time for business judgement, while the decision remains grounded in the knowledge, experience and accountability appropriate to the role. This model is particularly useful for managers, compliance specialists, legal teams, strategic procurement, finance teams and teams responsible for key accounts. FAQ What is the difference between an AI agent and a chatbot? A chatbot primarily responds to questions in a conversation. An AI agent can also use approved tools, retrieve information from enterprise systems, follow workflow rules and complete defined process steps. Its value comes from combining language understanding with access to business context, permissions and a controlled process. Should every AI agent have human approval before taking action? No. The appropriate level of oversight depends on the impact and reversibility of the action. Low-risk, repeatable activities such as categorising tickets or sending a standard confirmation can be automated under defined rules. Actions affecting customers, contracts, finances, compliance or production systems should usually include approval or escalation to an authorised person. Can one AI agent operate at different levels of autonomy? Yes. Autonomy should be designed for individual actions rather than assigned to an entire solution. The same agent may classify a request automatically, prepare a response for approval and escalate an unusual case to an expert. This makes it possible to automate safely without treating every task in the same way. What information does an AI agent need to work reliably in an enterprise? An agent needs access to reliable and current knowledge sources, a clearly defined objective, appropriate permissions and rules for handling exceptions. It should also receive only the context relevant to the task and role. Workflows, validations and an auditable history of actions help ensure that its output can be reviewed and used consistently. How can a company start implementing GPT-powered AI agents? Start with one clearly defined process step that has measurable volume, repeatable inputs and a known outcome. Set the boundaries of the agent’s permissions, test it with standard and exceptional cases, and measure the effect on process time, quality and escalations. Once the team has evidence that the solution works reliably, its scope and autonomy can be expanded gradually.
Read