Someone in your company is already using Microsoft Copilot or ChatGPT. Probably several people. The question landing on quality teams right now isn't whether to allow it — it's where the line sits, and how to write that down in a way people actually follow.
Short version: general-purpose assistants are fine for plenty of work in a regulated company, and they cannot be the system of record for anything GxP. The useful part is understanding exactly why, because that tells you where the line goes.
Why a general assistant can't hold a GxP record
It isn't about the quality of the writing. It's about four things a regulated record needs that these tools don't provide.
- No record-level audit trail. Part 11 and Annex 11 expect a tamper-evident trail showing who created or changed a record, when, and what changed. Your tenant may log that a chat happened; that is not an audit trail for the content.
- No electronic signature. There is no compliant signing step, no re-authentication, no signature meaning.
- The version changes underneath you. The provider updates the model on its schedule. A validated state you cannot hold still is not a validated state — see change control and revalidation.
- No defined intended use. A general assistant does everything and is therefore validated for nothing. Validation starts with a narrow, written intended use.
None of that makes the tools unusable. It makes them unusable as the place a GxP record lives.
Where they're genuinely fine
- Summarising published guidance or a standard you already have access to.
- Drafting internal, non-GxP material — a meeting agenda, a project update, a job advert.
- Rewriting your own text for clarity, where you verify every sentence and the result goes through the normal controlled process.
- Helping someone understand a concept before they do the regulated work in the validated system.
- Coding and data work outside GxP systems.
The common thread: nothing leaves the assistant and becomes a record without a person putting it through the controlled system, where the audit trail and signature live.
Where people get into trouble
- Pasting a deviation or an investigation in to "help write it up". You have now sent regulated content to a system outside your quality system, and the draft has no traceability back to source.
- Asking it to interpret a regulation and acting on the answer. It will sound confident and is sometimes wrong — and there is no citation you can put in front of an inspector.
- Generating test scripts or validation documents and filing them. If the output becomes a GxP deliverable, the tool that produced it is in scope. See GxP AI validation.
- Uploading client or pre-approval material. Often barred by your own contracts, whatever your AI policy says. See data privacy in GxP AI.
"Can we validate Copilot?" — the honest answer
You can validate a specific, narrow use of a tool. You cannot validate a general assistant in the abstract, because validation is always against an intended use, and "answers any question" is not one.
So the question to ask is: what exactly would we rely on it for, what would go wrong, and would a person catch it? If the answer survives that, you write the intended use, set acceptance criteria, test on real examples, and keep a human approving. At that point you generally discover you want a tool built for the job — one with an audit trail, signatures, version control and a defined scope — which is the difference between an assistant and a validated system.
The draft EU GMP Annex 22 points the same way: it does not expect generative models to be making critical GMP decisions, and where generative AI supports lower-risk work, the control is a qualified person reviewing the output.
A policy people will actually follow
Blanket bans fail. People use the tools on their phones instead, and you lose visibility. A workable policy is short and specific.
- Name the approved tool and tenant. An enterprise tenant with contractual terms is a completely different risk profile from a personal account.
- List what must never go in: patient data, pre-approval material, client-confidential documents, anything from a GxP record.
- State the output rule in one line: nothing from an assistant becomes a GxP record without going through the controlled system and a named approver.
- Say where to go instead. If people need AI help with validation or clinical documents, point at the validated route — that is what GxP Copilot and TraceDraft exist for.
- Train on real examples, including a confidently wrong answer. See SOPs and training for AI.
For the wider regulatory picture, see GxP AI and AI governance services.
Where to go next
Explore Cybroscape GxP Copilot for AI-native validation, Cybroscape TraceDraft for source-traceable clinical documentation, or book a demo to see either on your own data.