Can you use ChatGPT, Claude or Copilot for GxP work?
In my reading, usually yes. Context of use decides: once output becomes a GxP record, an SOP, or a specification, normal review applies.
Regulatory status checked October 8, 2026. Annex 22 and the revised Annex 11 are drafts, and FDA’s AI guidance for drug manufacturing is announced but not published. Check the dates before you rely on any of it. The tables are my reading of current rules, not official guidance.
For code and coding agents, see Can you use GitHub Copilot, Codex or Claude Code for GxP code?.
In my reading, usually yes. What decides it is the context of use: what the output is used for and who relies on it, not the brand of the tool. Once AI output becomes a GxP record, goes into an SOP or specification, or is code in a validated system, a qualified person reviews and approves it and the system is validated for that use, the same as for work a person did. The official stance is still unsettled; here is what regulators have said.
Which tool is which: Microsoft 365 Copilot (Microsoft now calls it Microsoft Copilot) drafts documents, email and meeting notes over your Microsoft 365 data, while ChatGPT, Claude and an internal company GPT are chat assistants whose text you copy elsewhere or that feed connected apps and internal systems.
Context of use, in plain words
Context of use is what the AI output is used for, who relies on it, and how much a GxP record or decision depends on it.
FDA calls it context of use
FDA’s January 2025 AI draft guidance: “The COU defines the specific role and scope of the AI model used to address a question of interest.” The description “should describe in detail what will be modeled and how model outputs will be used.”
In its example, an AI model checks vial fill volume. Because fill volume is independently verified on a representative sample from every batch, “the AI-based model will not be the sole determinant for the release of product”. In my reading, that is the difference between helping a person and deciding.
Other texts say “intended use”
The EU’s draft Annex 22 asks that “The intended use of a model and the specific tasks it is designed to assist or automate should be described in detail”.
FDA’s Computer Software Assurance (CSA) guidance calls itself “a risk-based approach for establishing and maintaining confidence that software is fit for its intended use.”
EMA, in one line
EMA’s workshop report on Annex 22, dated June 30, 2026 and published October 2, 2026: “Technology labels should not determine acceptability; intended use, impact, uncertainty, complexity and control strategy should.”
One tool, several rows
CSA says “software may have one or more intended uses depending on the individual features, functions, and operations of that software.”
Say a developer uses Claude Code in the morning on an internal dashboard with no GxP data, and in the afternoon on a change to a validated LIMS report. Same tool, two rows.
A caveat, in my reading. FDA’s AI draft is about AI that produces information for regulatory decisions, and it “does not address” AI used “for operational efficiencies”. CSA is written for medical device manufacturers. I borrow their words, not their scope.
The short answer, by context of use
These cards are my reading of current rules, not official guidance. Each one is a context of use: what the output is used for, who relies on it, and its GxP impact. One tool can land on different rows. Your quality unit decides how your company applies this.
Table A: Writing, analysis and records
Row 1 GxP impact: none Everyday work with no GxP content
Say a team wants an email, meeting notes, a summary of a public paper or guidance, or to learn a concept.
- Who relies on it The author only
- What’s usually needed The company’s AI use policy. The enterprise or business plan, not a personal account. The company’s data rules. The enterprise agreement can address confidentiality; it does not make the use GxP ready
- Anchor in the sources Vendor data terms; CSA’s “email or accounting applications” example (device scope, borrowed): software for “general business processes or operations not specific to production or the quality management system, such as email or accounting applications” is generally not treated as part of production or the quality system
Row 2 GxP impact: indirect A first draft of a GxP document that a person finishes in the controlled system
Say a team wants an SOP or specification draft, training slides, or a change control rationale.
- Who relies on it The author, then the reviewers and the quality unit who approve the final document
- GxP impact Indirect. The approved document is the GxP record; the chat is not
- What’s usually needed The author checks every fact against the source. Normal review, and quality unit approval where the rule requires it. The final version lives in the validated document system. Record AI assistance if your SOP asks for it
- Anchor in the sources 21 CFR 211.22(c) and 211.100(a); draft Annex 22, human in the loop for non-critical use; Purolea letter
Row 3 GxP impact: indirect Questions over GxP content to support a person’s decision
Say a team is summarizing deviations, searching SOPs, or comparing batch documents.
- Who relies on it A qualified person who makes the GxP decision
- GxP impact Indirect, but GxP data goes in and reliance can creep
- What’s usually needed The tool is approved for that data. Permissions are right. A documented risk assessment says how far people may rely on it. The qualified person stays responsible for the decision. Reassess when people stop checking
- Anchor in the sources Draft Annex 22 section 3.3; Annex 11 (2011) section 1; FDA draft’s fill volume example
Row 4 GxP impact: direct Output written into a GxP record, or driving a GxP decision
Suppose the work is classifying deviation severity, supporting batch disposition, or an agent that writes to the QMS.
- Who relies on it The GxP process itself and everyone downstream
- GxP impact Direct (critical)
- What’s usually needed A specific GxP computerized system with a written intended use, validation, audit trail, access control and change control for model versions. Not a job for a general chat assistant. The draft Annex 22 says generative AI and LLMs should not be used in critical GMP applications
- Anchor in the sources Annex 11 (2011) principle; 21 CFR 211.68(b); 21 CFR 11.10 and FDA’s 2003 enforcement-discretion guidance; draft Annex 22 scope; EMA workshop report, dated June 30, 2026 and published October 2, 2026
Row 5 GxP impact: regulatory Responses to MHRA Compliance Teams following GxP inspections
Say a team drafts an inspection response.
- Who relies on it The inspectorate
- What’s usually needed Accurate, verifiable, technically reviewed and signed off by someone with authority and accountability
- Anchor in the sources MHRA Inspectorate blog, June 29, 2026
For anything in rows 3, 4, 7 or 8, the classification worksheet helps set how much evidence you need.
Rows 6 to 8 are in Can you use GitHub Copilot, Codex or Claude Code for GxP code?.
The line is the output
What changes is where the output goes, so here is what the existing rules say for each place.
Documents and records
A qualified person reviews and approves, and the final version lives in the controlled system. Under 21 CFR 211.22(c), “The quality control unit shall have the responsibility for approving or rejecting all procedures or specifications impacting on the identity, strength, quality, and purity of the drug product.” Under 211.100(a), written procedures, “including any changes, shall be drafted, reviewed, and approved by the appropriate organizational units and reviewed and approved by the quality control unit.” In my reading, who drafted them, a person or a model, does not change that. 211.68(b) adds that “Input to and output from the computer or related system of formulas or other records or data shall be checked for accuracy.”
Part 11 “applies to records in electronic form that are created, modified, maintained, archived, retrieved, or transmitted, under any records requirements set forth in agency regulations.” FDA calls those underlying requirements, such as Part 211, predicate rules. In my reading, a chat window is usually not where the final GxP record lives. The approved document in your controlled document system is.
Systems you rely on
The 2011 EU Annex 11, still current, “applies to all forms of computerised systems used as part of a GMP regulated activities.” Its section 1 says “decisions on the extent of validation and data integrity controls should be based on a justified and documented risk assessment of the computerised system.” Part 11’s text calls for “Validation of systems to ensure accuracy, reliability, consistent intended performance, and the ability to discern invalid or altered records.” FDA’s 2003 guidance says it intends to exercise enforcement discretion for the specific Part 11 validation requirement in 11.10(a) and the specific audit trail requirements in 11.10(e) and 11.10(k)(2), while applicable predicate-rule requirements still apply. What that means for an assistant is in the validation section.
Data going in
Can you put GxP data or confidential information into Copilot, ChatGPT or Claude? In my reading, keep three questions separate.
- Is the tool approved for that data class? Your company’s data classification decides, not the vendor.
- Is it validated for the use? A tool approved for confidential data can still be unvalidated for any GxP use.
- What can it see? Microsoft says Copilot “only surfaces organizational data to which individual users have at least view permissions”, so clean up permissions before rollout.
Confidentiality is a contract and settings question. For the cited business and enterprise products, the vendor pages say customer data is not used to train models by default, subject to any opt-in feedback or model-improvement settings; the exact lines are in the vendor cards. Personal plans can differ. OpenAI says for Pro and Plus, “Conversations may be used to improve models unless you turn off training in ChatGPT data controls.” Anthropic says it will “train new models using data from Free, Pro, and Max accounts when this setting is on (including when you use Claude Code from these accounts).”
What regulators have and haven’t said (as of October 2026)
Full status map: where every AI guidance document stands as of fall 2026.
FDA. No rules or guidance written specifically for AI assistants or coding agents. The January 2025 AI draft guidance is marked “Not for Implementation” and “does not address” AI used “for operational efficiencies”. For drug manufacturing, its footnote 21 says “the responsibilities of the quality control unit described in 21 CFR 211.22 and 211.68 are applicable.” CDER’s guidance agenda (July 2026 version) lists two planned AI drafts: “Computer Software Assurance for AI-Based Systems in Drug Manufacturing and Clinical Investigations; Supplemental Guidance” and “AI and ML Quality Considerations in Pharmaceutical Manufacturing”. I found neither published as of October 8, 2026, and CDER says it “is not bound by this list”.
EU. The 2011 Annex 11 is in effect today. The Annex 22 consultation draft says generative AI and LLMs should not be used in critical GMP applications and states that, if they are used in non-critical GMP applications, qualified and trained personnel should ensure the outputs are suitable for the intended use. EMA’s workshop report, dated June 30, 2026 and published October 2, 2026, records that the drafting group discussed a broader scope covering generative AI and LLMs under a documented, risk-based control strategy, and that “concerns remain about the elements needed for such a control strategy.” No revised Annex 22 text is published on the Commission’s EudraLex page; the consultation page still links the 2025 draft. The revised Annex 11 is also still a draft.
UK. The MHRA Inspectorate blog of June 29, 2026, on submissions to MHRA Compliance Teams following GxP inspections: “our concern isn’t whether you use AI; it’s whether your submissions are accurate, verifiable, and prepared under appropriate oversight.” It sets out current thinking.
The one warning letter. The only posted FDA drug CGMP warning letter I found that says a firm used AI to create GMP documents went to Purolea Cosmetics Lab on April 2, 2026. The firm used AI agents to create drug product specifications, procedures and master production or control records, and it did not review the AI-generated documents as 21 CFR 211.22(c) requires. FDA wrote: “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.” In my reading, the AI-specific problem was reliance on AI-generated documents without quality unit review, not the mere use of AI. One letter is not a trend. The full story is here.
Does Microsoft 365 Copilot, ChatGPT or Claude need to be validated?
You validate a context of use, not a product. In my reading, an assistant used for drafts that a person checks and approves in the controlled system is usually handled through policy, training and normal document review. A context of use you rely on for a GxP record or decision needs its own intended use, testing and change control. See how much validation does your AI need and the most valuable AI call is not this one.
What the vendors actually promise
All pages below were read on October 8, 2026. Quotes are exact.
Microsoft Learn, GxP (FDA 21 CFR Part 11)
- What it says “There is no GxP or FDA 21 CFR Part 11 certification for cloud service providers”
- What it doesn’t say That any Azure or Copilot use is GxP ready
- What stays with you “you should determine the GxP requirements that apply to the computerized system based on its intended use”
- What it says “There’s no GxP certification for cloud service providers”; lists Azure, Microsoft 365 and Microsoft Dynamics 365 as in-scope cloud platforms and services, and points to Qualification Guidelines for Office 365
- What it doesn’t say Copilot is not named
- What stays with you The same intended use sentence
Microsoft, Data, Privacy, and Security for Microsoft Copilot
- What it says “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs”; “The responses that generative AI produces aren’t guaranteed to be 100% factual.”
- What it doesn’t say Nothing on GxP, Part 11 or Annex 11
- What stays with you Permission cleanup before rollout; review of output; assessment when models change
- What it says “We do not train our models on your data by default”; ChatGPT Enterprise, Edu and for Healthcare have each “successfully completed a SOC 2 Type 2 audit”
- What it doesn’t say No GxP, Part 11 or Annex 11 statement
- What stays with you Context of use, validation where relied on, records kept in your own system
- What it says “By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models.”
- What it doesn’t say No GxP statement
- What stays with you Feedback settings; data classification
- What it says “It is Customer’s responsibility to evaluate whether Outputs are appropriate for Customer’s use case, including where human review is appropriate, before using or sharing Outputs.”
- What it doesn’t say No GxP statement
- What stays with you Human review and fitness for the context of use, in the vendor’s own words
Anthropic, life sciences guide (PDF)
- What it says “Software that supports regulated development and manufacturing must meet Good Practice (GxP) expectations and, for electronic records and signatures, 21 CFR Part 11. In practice that means validated systems, controlled changes, and complete audit trails.”
- What it doesn’t say It discusses GxP but does not say Claude is validated or GxP compliant. It is a vendor guide, not guidance
- What stays with you Validation, change control and audit trails
Microsoft Learn does describe a Veeva QualityDocs connector that lets Microsoft 365 Copilot surface “SOPs, work instructions, policies, CAPAs, batch records, and other GxP documents”. That page describes a connector, not a GxP qualification of Copilot.
So the plans answer where does my data go and who can see it. They do not answer is this context of use validated. Models also change: Microsoft says “The AI models that power Microsoft Copilot are regularly updated and enhanced.”
Is ChatGPT safe, then? Ask: safe for which data, and for which context of use.
What pharma companies have said in public
Examples from the Pharma AI Tracker, in each entry’s own words:
What none of them says: which context of use applies, or whether anything was validated. In my reading of the routing layer, none of the 37 platform and assistant entries carried a GxP label. Where the Tracker labels something GxP, it is an application, such as Takeda’s deviation investigation assistant, not the company-wide assistant. That is about disclosure, not what companies run. More internal GenAI platforms are in the Tracker hub.
When the context of use changes
Say a tool assessed for drafting emails is not assessed for batch record entries, and a coding agent approved for internal tools is not approved for validated code. In my reading, treat the new context of use as a change, with its own assessment. That is boundary creep.
Connectors and agents move a request down the tables fast: Copilot agents, GPTs with actions, an internal GPT wired to the QMS, a coding agent with production credentials. See the agent edge. For an internal GPT behind a gateway, see what a CSV validator should look for in an LLM gateway, a gateway from scratch, and an ALCOA+ audit trail for LLM calls (Part 11’s text calls for “secure, computer-generated, time-stamped audit trails”, read with FDA’s enforcement-discretion policy described above).
Eight questions before saying yes
My checklist, not a regulator’s:
- What is the context of use: what will the output be used for, and does it become or change a GxP record, an SOP, a specification or code in a validated system?
- Who relies on it, and who reviews and approves it? Is that review written into an SOP?
- Is the user on the company’s enterprise or business plan, not a personal account?
- What data goes in, and is the tool approved for that data class?
- Does the tool see more than the user should (permissions, connectors, credentials)?
- Where does the final version live, and is that system validated?
- For code: does it go through the same review, testing and change control as code a person wrote, and who checks the tests the agent wrote?
- What happens when the vendor changes the model or adds a feature?
FAQ
Can ChatGPT be used in GMP?
Yes, depending on the context of use. Drafts that a qualified person checks, with the final record in your controlled system, are usually fine; critical GMP decisions are not a job for it. See the cards by context of use.
Does Microsoft 365 Copilot need to be validated?
You validate a context of use, not a product. See the validation section.
Is Microsoft Copilot, ChatGPT or Claude GxP compliant?
No product is GxP compliant on its own. Microsoft says “There is no GxP or FDA 21 CFR Part 11 certification for cloud service providers”, and none of the other vendor pages I read claims GxP compliance. See the vendor cards.
Can I paste a deviation or batch record into ChatGPT, Claude or Copilot?
In my reading, only if your company has approved that tool and plan for that data. That is a confidentiality question, not a validation one. See data going in.
What does "context of use" mean for AI in GxP?
What the output is used for, who relies on it, and how much a GxP decision depends on it. FDA’s draft AI guidance: “The COU defines the specific role and scope of the AI model used to address a question of interest.” See context of use, in plain words.
Has FDA or the EU issued rules on AI assistants?
Not specifically, and the draft texts are still moving. See what regulators have said.
Can AI write our SOPs?
It can draft them. FDA’s April 2, 2026 letter to Purolea cited the missing review: “If you use AI as an aid in document creation, you must review the AI generated documents”, under 21 CFR 211.22(c). See one AI warning letter is not a trend.
Can we use AI to draft an inspection response?
For responses to MHRA Compliance Teams after GxP inspections, the MHRA says its concern “isn’t whether you use AI” but whether submissions are accurate, verifiable and prepared under appropriate oversight, and signed off by someone with authority and accountability. See the MHRA Inspectorate blog.
Sources
Read on October 8, 2026 unless noted. How Tracker entries are sourced: method.
- FDA, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, draft, January 2025.
- FDA, Computer Software Assurance for Production and Quality Management System Software, February 3, 2026.
- FDA, CDER guidance agenda, July 2026 version.
- FDA, Part 11, Electronic Records; Electronic Signatures, Scope and Application, 2003.
- FDA, warning letter to Purolea Cosmetics Lab, 320-26-58, April 2, 2026.
- eCFR, 21 CFR Part 11 and 21 CFR Part 211.
- European Commission, EU GMP Annex 22, Artificial Intelligence, consultation draft, 2025.
- European Commission, EU GMP Annex 11, Computerised Systems, 2011, and EudraLex Volume 4.
- EMA, Report of the multistakeholder workshop on expert contributions to AI guidance development (Annex 22), EMA/156789/2026, dated June 30, 2026, published October 2, 2026.
- MHRA Inspectorate, Use of AI for GXP inspection responses, June 29, 2026.
- Vendor pages: as linked in the vendor table.
I work with Quality and AI leaders on questions about using GenAI and agents in GxP workflows defensibly. If you are working out which of your AI uses sit on which row, get in touch. You can also decide which GenAI tools fit your GxP work.