Copilot, ChatGPT and Gemini Need Your Own Controls for GxP Work

Microsoft 365 Copilot, ChatGPT Enterprise and Gemini for Google Workspace can each draft part of a standard operating procedure (SOP). The records an inspector will ask about that draft are the company's to keep. We read each vendor's own documentation against six questions an inspector could put to an AI-drafted SOP. Of the 18 pairs of question and product, settings answer 8 in part, and none in full. The rest of each answer has to come from controls the company adds.
That division of work follows from how the rules are written. GxP, the good-practice rules for manufacturing, laboratory and clinical work, applies to the record. So do FDA's Part 11 and its EU counterpart, Annex 11, which govern records and signatures kept on computers, and none of them asks who or what drafted the text. FDA's first warning letter about AI, sent in April 2026, went to the firm that used the tool. It told PurOlea Cosmetics Lab: "If you use AI as an aid in document creation, you must review the AI generated documents to ensure they were accurate and actually compliant with CGMP."
Microsoft and Google place GxP compliance with the customer
Two of the three vendors say in their own GxP documents that the regulated company carries this work. Microsoft's GxP page lists Microsoft 365 among its services in scope and notes that "There's no GxP certification for cloud service providers". Google's GxP whitepaper says "the responsibility for GxP compliance ultimately lies with our life science customers". Neither document mentions Copilot or Gemini, and Google's dates from January 2023. We found no OpenAI page on GxP, Part 11 or validation.
Their position follows from what validation is: documented evidence that a computer system does what the company relies on it to do. What the company relies on it for is the context of use, FDA's term for an AI tool's role and scope, which validation practice calls intended use. A vendor can't validate its product for a use it has never seen. Where a use falls under GxP, a question covered separately, the company supplies the validation and the records.
Documented settings answer 8 of 18 inspector questions in part, mostly on logs
An inspector who finds an AI-drafted paragraph in an SOP will trace it from the model that wrote it to the signature that approved it. The table gives what each vendor's pages say at each step.
| Inspector's question | Microsoft 365 Copilot | ChatGPT Enterprise | Gemini for Workspace |
|---|---|---|---|
| 1. Which model version drafted it? | Provider logged; model name and version "Not available" | Model logged; files kept 30 days | No model field |
| 2. Who reviewed it, and what did they check? | Prompts users to review | Not documented | Warns output may be inaccurate |
| 3. Does the record show the AI's part? | Log names the resource used | Not documented | Log says "assisted by Gemini" |
| 4. What happens when the model changes? | Not every change announced | Admin enables new models | New model reaches users automatically |
| 5. Can the audit trail be reproduced? | 180-day log; chats kept in mailbox | 30-day log; export it | 6-month log; Google Vault for chats |
| 6. Is an e-signature linked to the content? | Customer's responsibility | No statement found | Customer's responsibility |
Eight of the 18 cells have a documented setting that answers part of the question, mostly by keeping logs. In six, the vendor's pages leave the question to the customer, and in the other four we found nothing on the vendor's public pages. Each card below quotes the vendor's page word for word.
6 × 3 inspector questions checked against three assistants22 vendor pages read on 18 November 202625 of 25 quotes found word for word on the vendor's page
Open a group to see each vendor's own words. The highlight marks the phrase the coding rests on, and the grey note is our reading of it.
8 of the 18 question and vendor pairs, in question order.
Audit logs for Copilot and AI applicationsModelVersion is the version of the model used. Not available in Microsoft 365 Copilot scenarios.
Audit logs for Copilot and AI applicationsAudit logs for Microsoft Copilot and Microsoft Copilot Chat include the model provider and model name when a request is processed.
Our readingEvery Microsoft 365 Copilot audit record names the model's provider. A model name appears for Copilot Chat, and the same page says it may be missing when the user leaves the choice on Auto. No record gives a model version.
Microsoft's pageOpenAI Admin API ReferenceModel-authored message; `model` is always present (e.g. `gpt-5-2-pro`)
OpenAI Admin API ReferenceFiles expire after 30 days by default.
Our readingEvery assistant message in the compliance log names the model that wrote it. The log files expire after 30 days unless the company downloads and keeps them.
OpenAI's pageAudit logs for Copilot and AI applicationsID is the identifier of the resource that was being used during the Copilot interaction.
Our readingThe audit log can tie a Copilot interaction to the file that was open. That link lives in the log, which keeps entries for 180 days unless the company sets a longer policy.
Microsoft's page
Every card is in the data (JSON), and every quote with its URL is in the CSV below.
Even the best-covered question, which model version drafted a paragraph, needs a record the company keeps. OpenAI's Admin API reference says "model is always present" on each AI-written message in its compliance log, whose files expire after 30 days. Microsoft's Copilot audit log records the model's provider, but its model name and version fields both read "Not available in Microsoft 365 Copilot scenarios." Copilot Chat logs a model name, though not when the user leaves the choice on Auto. Google lists no model among its Gemini log fields. To answer the question next year, the company has to export or record the model itself.
Only the company's records can show who checked the AI's text
The two vendors that address accuracy point the user to the check. Microsoft says its product prompts "users to review, fact-check, and adjust content based on subject-matter expertise." Google cautions that Gemini Apps "may sometimes produce inaccurate" information. None of the three vendors' pages describes a record of that check, so an SOP saved from an assistant carries only the name of the employee who saved it.
An inspector will look for that record. FDA's letter to PurOlea says AI output "must be reviewed and cleared by an authorized human representative" of the quality unit, the group that approves procedures. The review has to be substantive, a qualified person's documented check of the text against its sources. Its record belongs in the company's document system, next to the text it covers.
The company's change control has to catch a new model
A new model can change what an assistant writes. Annex 11 says changes to a computer system "should only be made in a controlled manner in accordance with a defined procedure." Microsoft's page on model changes says "not every model change is communicated through a dedicated announcement to customers." Google told Workspace customers that users "will receive the new 3 Pro model update automatically." OpenAI launched recent models switched off in its Enterprise workspaces, where "An administrator must enable each model before members can select it." Retirements still follow OpenAI's dates. With notice this varied, the company's change control has to check which model is in use on its own schedule.
Logs and signatures have to last as long as the record
Part 11 ties the life of the audit trail, the time-stamped log of who created or changed a record, to the record itself. 21 CFR 11.10(e) requires the trail to be kept "for a period at least as long as that required for the subject electronic records." For batch records, 21 CFR 211.180 sets that at least 1 year after the batch expires. The standard log periods are shorter. Microsoft Purview keeps standard audit records for 180 days, OpenAI keeps compliance log files for 30 days, and Google keeps Gemini log events for 6 months. Each vendor documents a way to keep more, through retention policies, exports or Google Vault rules, so the gap closes once an administrator sets it up.
Signatures sit outside the assistants altogether. 21 CFR 11.70 requires electronic signatures to be "linked to their respective electronic records to ensure that the signatures cannot be excised, copied, or otherwise transferred". None of these assistants signs anything, so the signature belongs in the company's document system. There, under 11.50, it states its meaning, such as review or approval, and that meaning holds only if the review behind it is on file.
An inventory of AI use already in GxP work comes first
The FDA and EMA principles of good AI practice of January 2026 ask for "proportionate validation, risk mitigation, and oversight based on the context of use and determined model risk." Controls can only be proportionate to uses the company knows about. The first control is therefore an inventory of where staff already use general assistants on GxP work, including personal accounts and tools nobody approved. Each use then gets a risk level by what the AI does with the record. Uses that add text to a record get a written procedure and a documented review. The questions to ask an AI vendor then apply to the tools the company keeps. With those controls in place, an inspector's question about an AI-drafted paragraph is answered from the company's own files, whichever assistant drafted it.
For a steering committee, the open questions and decisions are these:
- Where do staff use AI assistants on GxP work today, including personal accounts?
- Does each use stay informational, add text to a record, or inform a decision?
- Which tools do we approve for which uses, and in which written procedure?
- Who reviews AI text before it enters a record, and where is that review kept?
- Which admin settings must be on for model logging and log retention?
- How will we learn that a vendor changed the model, and who assesses it?
Data and method
On 18 November 2026 we read and saved 22 pages that Microsoft, OpenAI and Google publish on GxP and on their assistants' logs, models and data. We quoted the passage that bears on each question for each product and coded the pair in one of three groups. All 25 quotes appear word for word on the saved pages. The coding is our reading, and a vendor may answer more in contracts or admin consoles we can't see. OpenAI's main site and help center refused automated access, so its cells rest on its other pages.
Downloads: vendor quotes, URLs and coding (CSV).
Frequently asked questions
Is Copilot or ChatGPT validated for GxP?
No vendor validates an assistant for your use, because validation depends on how your company uses a system. Microsoft and Google say in their own GxP documents that GxP compliance rests with the customer.
Who is accountable if the AI gets something wrong?
The company is, through its quality unit, just as for any document it approves. FDA's April 2026 letter to PurOlea cited the firm for using AI-drafted procedures without review.
What do we tell the auditor?
Show where AI is used, the procedure that governs each use, and the review record for any AI text in a GxP document. Keep the assistant's logs for as long as the records they relate to.