AI-assisted inspection responses: what the MHRA actually expects
The MHRA Inspectorate recently published its position on organisations using AI tools to help draft responses to GxP inspection findings, and it's worth reading in full rather than assuming it's another blanket restriction. It isn't. The regulator's actual position is narrower and, in some ways, more demanding: use whatever tools help, but the accountability for what goes out the door does not move an inch.
All submissions and responses must be factually accurate and verifiable, technically reviewed by appropriately experienced people, and signed off by someone with authority and accountability. MHRA Inspectorate, Use of AI for GXP inspection responses
What actually went wrong
The blog post is candid about the failures that prompted it. These weren't hypothetical risks raised to justify a new rule. They're documented failures the Inspectorate had to work through, and each one maps to a distinct way an unverified AI draft can go wrong.
| What the MHRA found | Why it mattered |
|---|---|
| Fabricated references to MHRA guidance that doesn't exist | Undermines trust in the entire submission, not just the fabricated section |
| A 90+ page response that never addressed the deficiency raised | Wastes inspector time and reads as evasive rather than responsive |
| Material inaccuracies in a safety-related response | Delayed resolution of a genuine patient safety issue, pushing review time from a few hours to more than twenty |
Read together, the pattern isn't "AI produced bad writing." It's that nobody caught the errors before the response was submitted. A human reviewer who actually checked the cited guidance, or actually read the deficiency being answered, would have stopped every one of these before it reached an inspector.
The standard hasn't changed, only the tooling has
Every requirement the MHRA lists for AI-assisted responses is the same standard that should already apply to any inspection response, AI-assisted or not:
- Factually accurate, with evidence that supports every claim made
- Technically reviewed by someone with genuine expertise in the area the response covers
- Signed off by a named person with real accountability for its content
- Contextually appropriate to the specific deficiency raised, not a generic or templated answer
AI doesn't get a lower bar because it's new, and it doesn't get a higher one either. It gets exactly the same bar, applied by someone who actually checked the output rather than trusting it.
Disclosure is treated as a sign of maturity, not a confession
One detail is easy to miss: the MHRA has set up a voluntary disclosure option where organisations can note which sections of a submission involved AI assistance and confirm that a human verified them. Rather than treating that disclosure as an admission of risk, the Inspectorate says it will view transparent, well-controlled AI use positively, as evidence of a mature quality culture. Hiding AI use, or not having a process to catch its mistakes, is the thing that actually damages trust during an inspection. Where verification processes prove inadequate, the MHRA has been explicit that it may reject the response outright, escalate the organisation to higher-risk status, or refer systemic failures to the Inspection Action Group.
What this means if you're building or using these tools
For any organisation using digital systems that touch inspection-ready documentation, whether that's an eTMF, an audit trail, or a query-response workflow, this is a reminder that AI-assisted drafting is only as good as the verification step behind it. A few practical takeaways follow directly from the cases the MHRA describes:
- Treat an AI-drafted response as a first draft requiring the same scrutiny as one written by a junior team member, not as a finished answer
- Build a named sign-off step into any workflow that touches regulatory correspondence, so accountability is never ambiguous after the fact
- If you disclose AI involvement, disclose it specifically (which sections, what was checked), not as a vague blanket statement
- Treat a fabricated citation or reference as a process failure worth investigating, not a one-off mistake to quietly correct
A tool that helps someone draft a response faster is genuinely useful. A tool, or a process, that lets an unverified draft reach an inspector unchecked is a serious liability, and the MHRA has now shown it is actively looking for exactly that failure mode.
The lesson isn't to avoid AI in regulated environments. It's that the human review step was never optional, and now there's a documented, real-world example of what happens when it's skipped.