MHRA on Using AI to Draft GxP Inspection Responses: The New Standard

What the MHRA Published, and What It Actually Says

The post appeared on the MHRA Inspectorate blog under the categories AI, Artificial intelligence, Compliance matters, and GxP inspections.¹ It opens by narrowing its own scope: it relates specifically to submissions made to MHRA Compliance Teams following GxP inspections. That is a precise target. It is not guidance about AI in manufacturing, AI in clinical trials, or AI as a medical device. It is about the letters, response packages, and CAPA plans that companies send to the agency after an inspection report lands.

That narrowness is what makes it worth reading closely. Most regulatory commentary on AI addresses systems inside your operation. This one addresses the document you write to the regulator about those systems. It is direct feedback on the quality of your correspondence, from the people who read it.

The observed problem

The MHRA states that AI tools are being used to draft inspection responses supplied to the agency, and it credits the technology with real benefits: helping articulate complex technical issues, improving consistency, speeding up routine drafting, and enabling innovation. The post says that used properly, AI can support better regulatory outcomes and improve patient safety. That framing matters, because the agency is not building a case against the tools.

Then it describes what it has actually received. Responses containing references to MHRA guidance that does not exist. Citations of inappropriate regulatory frameworks. Responses to serious deficiencies that, in the agency’s words, “appear designed to mislead rather than address underlying problems.” One inspection response exceeded 90 pages and did not address the deficiencies identified, which the MHRA describes as a disproportionate use of inspector time and expertise. In at least one case, an AI-generated response to a serious deficiency with patient-safety impact contained material inaccuracies, including references to documents that do not exist, and the inaccurate information delayed resolution of the compliance failure. Reviewing that single response required multidisciplinary teams and a full review of previously issued MHRA guidance, taking the effort from roughly four hours to over twenty.

The agency’s summary of the situation is worth registering: the inappropriate application of AI has moved from a theoretical risk to one that has actually occurred. That is a change in how the MHRA is talking about this. Up to this point, most regulator commentary on generative AI in GxP has been anticipatory. This is a report of harm already done.

What has not changed

The section titled “What hasn’t changed” is the intellectual core of the post, and it is deliberately unglamorous. Organizations have always been responsible for the accuracy of their responses to the MHRA. The people responsible for those responses have always been expected to verify factual claims. Technical review has always been a core component of quality systems. Materially false statements to inspectors have always been regulatory offences. The post closes the section with a sentence that should be pinned to a wall in every quality department: “These responsibilities do not change because drafting tools have.”

This is the same logic the FDA applied in the Purolea matter, and the same logic behind the human-oversight expectations in the EMA reflection paper on AI across the medicinal product lifecycle.⁹ Regulators are not inventing new obligations for AI. They are pointing out that existing obligations were never conditional on how a document got written.

The five requirements

The MHRA sets out its expectations as principles rather than tool controls. Every submission or response must be:

• factually accurate and verifiable
• technically reviewed by appropriately experienced people
• signed off by someone with authority and accountability
• supported by evidence for factual claims
• appropriate to the specific regulatory context

The post adds that these expectations apply regardless of how you draft your submissions, and that “Good quality management systems work whether you use AI, templates, consultants, or write from scratch.” The agency states its own position plainly: “our concern isn’t whether you use AI; it’s whether your submissions are accurate, verifiable, and prepared under appropriate oversight.”

The consequences the MHRA is willing to apply

The post does not stop at expectations. It names what can happen when a response fails them. Where the MHRA identifies inaccurate information affecting the conduct of an inspection, it may act. If a response is inaccurate, incomplete, or overly verbose, the agency may reject it or return it for another attempt. It may also treat the organization as higher risk for the purposes of future inspection prioritization, on the grounds of poor CAPA, and it may refer the organization to the Inspection Action Group. The IAG is the non-statutory multidisciplinary body that advises on regulatory action, and referrals there can lead to increased inspection frequency, license variation or suspension, and in serious cases referral onward for enforcement.³¹¹ The post also warns that the agency may assess your verification processes directly, and that repeated patterns may indicate systemic failures warranting regulatory action.

What an AI-Drafted Response Looks Like From the Inspector’s Side

The most practically useful part of the MHRA post is the list of warning signs. The agency says that when it reviews responses it assesses the quality of the oversight process behind them, and it names the patterns that suggest verification was inadequate. Rendered in plain terms, those signs are:

• factually incorrect statements or references to documents that do not exist
• generic language inappropriate to specific circumstances
• a lack of company-specific detail where detail is expected
• citations of regulatory frameworks with no explanation of why they are relevant
• inconsistent technical terminology across submissions
• overly verbose responses that fail to deal with the subject matter appropriately

Anyone who has read a stack of CAPA responses will recognize every item on that list. What follows is what those patterns look like in practice, and why an experienced inspector spots them quickly.

Generic CAPA language that could apply to any company

The clearest tell is a corrective action that would be equally true if you deleted the company name and dropped in a competitor’s. “The procedure will be revised to strengthen controls and reinforce expectations. Personnel will be retrained. Effectiveness will be monitored through routine quality metrics.” Nothing in that sentence is false. Nothing in it is specific either. It names no procedure number, no control, no metric, no threshold, and no failure mode.

Large language models produce this kind of text well, because it is the statistical center of everything they have read about CAPA. The output is fluent, structurally correct, and empty. An inspector who has just spent three days in your facility knows exactly which procedure is deficient and exactly which step failed. Reading a paragraph that could describe any site in the country tells them that nobody who was in the room during the inspection wrote or seriously reviewed the response.

Commitments that are plausible but unverifiable

The second pattern is the commitment with no attachment point. “A comprehensive review of the affected batch records will be undertaken.” Over what date range? Which batches? Against what acceptance criteria? Reviewed by whom, and reported into which forum? A commitment that cannot be checked at the next inspection is not a commitment. It is a sentence.

This is where AI drafting is quietly corrosive. A human under time pressure who does not know the answer usually leaves a placeholder or asks someone. A model does not know that it does not know, so it fills the space with something that reads like an answer. The result passes a fluency check and fails a verification check, and the verification check is the one the inspector runs.

Corrective actions with no named owner or site-specific detail

Inspectors look for ownership because ownership is the difference between an intention and a plan. A corrective action needs a named accountable role, a target date, and a defined deliverable. AI-drafted responses routinely produce actions owned by “Quality Assurance” or “the site leadership team,” due “within 90 days,” delivering “an updated approach.” None of that can be tracked, and none of it survives contact with a CAPA effectiveness review. We have written before about what a serious CAPA effectiveness review looks like, and the 90-day look-back is where weak actions get exposed.

Confident prose describing systems the company does not have

This is the most damaging pattern of the four, and the hardest to walk back. A model asked to describe how a company manages, for example, periodic review of computerized systems will describe how a well-run company manages it. If your organization does not actually operate that process, you have now told a regulator in writing that you do. The MHRA’s non-existent-reference examples are the same failure applied to guidance documents rather than internal systems.

The MHRA states that materially false statements to inspectors have always been regulatory offences. The company that submits a confident description of a control it does not operate has not made a drafting error. It has made a false statement, and the fact that a tool generated the sentence does not move the accountability anywhere.

Why the Response Damages Credibility More Than the Finding Did

Quality leaders sometimes treat the inspection response as administrative closure. It is not. It is the single best opportunity a company has to demonstrate that its quality system works, and it is delivered on the company’s own terms, in writing, with time to prepare. That makes it the highest-signal document the regulator will receive from you that year.

The finding is about a process. The response is about your judgment.

A deficiency raised during an inspection says something went wrong in a process. That is bad, but it is bounded. Most findings are closed with a competent investigation and a proportionate action plan. The response says something different: it shows how the organization thinks when it is being watched. It reveals whether you can identify a genuine root cause, whether you assess impact honestly, and whether the people signing the document understand the operation they are describing.

The MHRA makes this connection explicit. Its post says that responses to inspection findings should demonstrate effective corrective and preventive action, meaning genuine root causes identified through evidence-based investigation, proportionate impact assessment, actions that address underlying issues rather than symptoms, and verification of effectiveness. It goes on to say that where responses show superficial or generic CAPA without evidence of genuine root cause analysis, that will be identified as a quality system weakness regardless of how the response was drafted. The agency points readers back to its own earlier writing on what effective investigation looks like in a GMDP environment.²

A weak response converts one finding into a systemic concern

Consider the arithmetic from the inspector’s chair. A single deficiency is a data point. A deficiency plus a response that misstates the regulatory framework, cites guidance that does not exist, and commits to unverifiable actions is a pattern. The MHRA post says as much: repeated patterns may indicate systemic failures warranting regulatory action, and the agency may assess the verification processes themselves.

This is how a contained finding becomes a compliance escalation. The MHRA has described its compliance management escalation route publicly for years, including the point at which cases move from routine risk-based inspection into closer monitoring and, if warranted, referral.³ A poor response is one of the cheapest ways to enter that route, and it is entirely self-inflicted.

The effort is asymmetric, and the regulator noticed

The MHRA’s own example makes the resource point unavoidable. One inaccurate response took a four-hour review to more than twenty hours and pulled in multidisciplinary teams. From the agency’s side, a company that generates that kind of burden is not a neutral party. The post describes the disproportionate use of inspector time and expertise as a strain, and inspectors are human beings with finite attention who will remember which companies waste it.

There is also a length trap here. Volume reads as effort to the author and as evasion to the reader. A 90-page response that does not address the findings is worse than a six-page response that does, because it forces the reader to prove a negative across 90 pages. If your draft has grown because a model expanded every section to a comfortable length, you have made the document harder to accept, not easier.

Where AI Genuinely Helps: Form, Coverage, and Clarity

It would be a misreading of the MHRA post to conclude that AI has no place in inspection responses. The agency says the opposite. The technology helps articulate complex technical issues, improves consistency, speeds up routine drafting, and can support better regulatory outcomes when used properly. The agency also says explicitly that it is not trying to police the technology used in the inspection process.

The line that matters is between form and substance. AI is good at the shape of a document, the coverage of a document, and the readability of a document. It is bad at the truth of a document. Every legitimate use sits on the first side of that line.

Structuring the response

Inspection responses have a predictable architecture: restate the observation, describe the immediate correction, present the investigation and root cause, assess impact and product risk, define corrective and preventive actions with owners and dates, and define effectiveness checks. Getting a consistent, complete skeleton in place quickly is genuinely useful, particularly when a response covers a dozen observations across several functions and multiple authors are drafting in parallel. A model that enforces the same structure for every observation reduces the risk that one gets a thinner treatment than the rest.

Checking completeness against every observation

This is the single strongest use, and it is underused. Take the inspection report, extract every discrete observation and every sub-point inside each observation, and check the draft response against that list. Does each sub-point have a corresponding correction, a root cause, an impact statement, an action, and an effectiveness check? Inspectors frequently note that a response addressed the headline of a finding and ignored a clause buried in the middle of it. Machine-assisted coverage checking is well suited to exactly this, because it is a matching problem, not a judgment problem.

Tightening language and removing padding

Given the MHRA’s specific complaint about verbosity, an editing pass that shortens sentences, removes repetition, and cuts hedging is directly aligned with what the agency wants to receive. This is the reverse of how most teams use generative tools. The instruction should be “make this shorter and more precise without changing any factual claim,” not “expand this section.”

Consistency of terminology across a package

The MHRA lists inconsistent technical terminology across submissions as a warning sign of poor verification. In a multi-author response, the same system will be called three different things and the same procedure will carry two different numbers. Checking a package for term consistency against a controlled glossary is mechanical work that a model does well and that humans do badly under deadline.

Surfacing internal contradictions before the inspector does

A useful review prompt is adversarial rather than generative: read this response and list every claim that contradicts another claim, every action with no owner or date, and every factual assertion with no attached evidence reference. Used this way the tool is a reviewer, not an author, and it is looking for exactly the defects the MHRA says it looks for. This is the same principle behind treating generative tools as verification aids in regulatory document authoring rather than as sources of content.

Where AI Becomes Dangerous: Generating the Substance

The damaging uses share one feature. In each case, the model is producing a claim about your organization or about the regulatory framework that nobody has independently established.

Root cause is the hardest line to hold

Of the prohibited uses, root cause is where teams most often drift. Investigation is difficult, deadlines are short, and a generated paragraph about inadequate procedural detail and insufficient training reinforcement will always be available. The MHRA anticipates this directly: AI tools may support parts of the CAPA process, but they cannot substitute for the technical understanding and company knowledge needed to develop meaningful corrective actions. Where a response shows superficial or generic CAPA without evidence of genuine root cause analysis, the agency will treat it as a quality system weakness.

The practical rule is that root cause must be traceable to evidence collected in your operation: interview records, trend data, equipment logs, batch records, audit trail review. If a stated root cause cannot be traced to a specific piece of evidence in the investigation file, it is not a root cause. It is a hypothesis that happened to be written down confidently.

Citations require a human with the source document open

The fabricated-reference problem is the most publicly embarrassing failure mode and the easiest to prevent. The rule is simple and absolute: no regulation, guidance document, standard, internal SOP number, or study is cited in a response unless a named person has opened the source and confirmed both that it exists and that it says what the response claims it says. That check takes minutes. Not doing it has already, according to the MHRA, delayed the resolution of a patient-safety issue.

A Working Internal Standard for AI-Assisted Inspection Responses

The MHRA sets outcomes and leaves method to industry. That is a reasonable regulatory posture and an uncomfortable one for quality teams, because it means you have to write the method yourself. What follows is a standard that can be adopted as a short procedure or as an addendum to an existing inspection response SOP. It has four parts: permitted assistance, non-delegable accountability, pre-submission verification, and disclosure.

Part 1: Permitted assistance, written down

State the permitted uses explicitly, because an unwritten rule is not a control. The wording can be short. AI tools may be used to organize, format, check, condense, and review inspection response content. AI tools may not be used to originate factual claims about the company, its systems, its products, or the regulatory framework. Anyone who has read the MHRA post will recognize that this is simply their five requirements expressed as a permission boundary.

Two operational points belong here. First, name which tools are approved, because an unapproved consumer tool may also create a confidentiality problem with inspection material. Second, state whether inspection report text may be entered into a tool at all. For many companies the answer will depend on the deployment, and that decision should be made once by information security and quality together, not improvised by an individual at 9pm before a deadline.

Part 2: The accountability that cannot move

The MHRA requires sign-off by someone with authority and accountability. Make that concrete. Each observation in the response should carry a named technical author who is competent in the area, a named reviewer independent of the author, and a single accountable signatory for the package, normally the Qualified Person, Head of Quality, or the equivalent role for the GxP area concerned.

The signatory statement should say what it actually means. Signing is a declaration that the factual claims have been verified, that the described systems exist as described, that the actions are achievable and owned, and that the citations have been checked. It is not a declaration that the document reads well. A signature block that says “reviewed and approved” without defining what was reviewed is the exact gap the MHRA is probing when it says it may assess your verification processes.

This is the same accountability principle that runs through current regulator thinking on human-in-the-loop requirements for pharma AI, and it is the point where the MHRA post and the FDA’s enforcement position converge most cleanly.

Part 3: The pre-submission verification pack

The most valuable artifact this standard produces is a short verification record that is completed before the response leaves the building and retained with it. It is not bureaucracy for its own sake. If the MHRA later asks how you assured the accuracy of your submission, this record is the answer, and it takes under an hour to complete for a typical response.

The commitment integrity check deserves emphasis. Actions should be entered into the CAPA system before the response is sent, not after. That single sequencing change prevents the most common gap between what a company promised the regulator and what it is actually tracking internally, and it is the gap an effectiveness review will find later anyway.

Part 4: Making the standard usable

A standard that adds three weeks to a response timeline will be abandoned during the first real inspection. Keep it proportionate. For a routine response with a handful of minor observations, the verification pack is a one-page form. For a response involving a critical finding or patient-safety impact, it should be a formal, independently reviewed record, and the accountable signatory should be able to describe from memory how each factual claim was verified.

The Disclosure Question: Should You Tell the Regulator?

The MHRA has offered organizations the option to disclose AI use in responses to compliance teams. It is not mandatory. The agency says it believes transparency benefits everyone, and it describes what a disclosure should contain: a brief statement at the start of the submission, identification of which sections involved AI assistance, and confirmation of human verification and approval. The post states that transparency in the use of AI tools, alongside strong mechanisms to ensure accuracy and reliability of submissions, “indicates a more mature and transparent quality culture,” and that inspectors will consider this positively when assessing compliance.

The case for disclosing

The agency has said, in writing, that it will view disclosure positively. That is not a small thing. Regulators rarely commit in advance to a favorable reading of a voluntary act. The MHRA has also framed disclosure as evidence of quality culture rather than as a confession, which changes the risk calculation considerably.

There is a second argument that matters more over time. If AI-assisted drafting is already happening across your organization, a disclosure statement forces you to know that it happened and to say who verified it. The disclosure is the visible end of an internal control. Companies that cannot produce the statement honestly usually cannot answer the underlying question either, which is exactly the weakness the MHRA says it will probe.

The case against, and why it is weaker than it looks

The objection quality leaders raise is that disclosure invites scrutiny you would otherwise avoid. Two things undercut that. First, the MHRA has already told inspectors what inadequate verification looks like, and those signs are detectable in the document regardless of whether you disclose. Non-disclosure does not make generic CAPA language specific. Second, a company that used AI and did not say so, and whose response is later found to contain a fabricated citation, is in a materially worse position than one that disclosed. The first looks like concealment. The second looks like a verification failure inside a declared process.

The strongest genuine concern is inconsistency. A company that discloses on one submission and not the next has created a pattern that raises questions about the omission. That argues for a firm default position rather than against disclosure.

A practical position

Our view is that pharma and biotech organizations should adopt disclosure as the default for MHRA compliance submissions, apply it consistently, and keep the statement short and factual. Something close to the following is sufficient:

Note what that statement does. It tells the regulator the tool touched form, not substance. It names the boundary. It points at a controlled procedure, which invites the agency to see a system rather than an improvisation. Draft it once, approve it as a template, and stop debating it under deadline.

Where disclosure does not belong

One caution. The MHRA’s offer applies to submissions to its compliance teams following GxP inspections. Do not extend a disclosure statement reflexively into regulatory submissions, marketing authorization documentation, or clinical trial reporting, where different expectations and different legal frameworks apply. Decide the scope deliberately and write it into the standard.

The Purolea Pattern: The Same Failure at a Worse Moment

The MHRA post does not exist in isolation. On 2 April 2026, the FDA issued a warning letter to Purolea Cosmetics Lab that included a distinct section addressing inappropriate use of artificial intelligence.⁴ The company had used AI agents to create drug product specifications, procedures, and master production and control records intended to satisfy FDA requirements, and had used those documents without further review by qualified personnel.⁵ The agency cited the failure to review as a violation of 21 CFR 211.22(c), which requires the quality control unit to approve or reject all procedures and specifications affecting identity, strength, quality, and purity.⁶

The detail that traveled furthest was the company’s response when told that process validation was required before distribution. It said it had not been aware of the legal requirement because the AI agent it used never mentioned it.⁵

Reading Purolea accurately

It is worth being precise about what that case does and does not show, because the coverage overstated it. Purolea was not a sophisticated manufacturer pushing the boundaries of AI. The AI citation sat alongside a long list of fundamental cGMP failures, including absence of process validation. The company subsequently stopped drug production.⁷ The FDA did not signal hostility toward AI, and reading the letter as an anti-AI enforcement action leads to the wrong conclusions.

The accurate signal is narrower and more useful. Existing requirements apply to AI-generated content exactly as they apply to anything else: human oversight, verified outputs, and documented decision records. What is genuinely new is only that companies have started trying to place accountability on a model, and regulators have now stated in two jurisdictions that this does not work.

Why an inspection response is the worse case

Set the two situations side by side. Purolea generated internal documents with AI and did not have qualified quality personnel review them. A company that generates an inspection response with AI and does not have qualified personnel verify it has committed the same control failure. The difference is where the document ends up.

An unverified internal SOP is a control failure waiting to be found. An unverified inspection response is a control failure handed to the person deciding whether your quality system can be trusted. The MHRA lists rejection of the response, higher risk classification, and IAG referral as available consequences. There is no equivalent internal safety net once the document has been sent.

The wider regulatory direction

These two events sit inside a consistent international pattern. The EMA’s reflection paper on AI across the medicinal product lifecycle sets out expectations for human oversight and risk-based validation.⁹ Draft Annex 22 of the EU GMP Guide, published for consultation in July 2025 and analyzed in the PDA Journal in early 2026, addresses AI use within GMP operations and reinforces the same accountability logic.¹⁰ The MHRA has been building an AI regulatory position for several years, from its April 2024 policy paper on the impact of AI on the regulation of medical products¹² and the AI regulatory strategy it published alongside it, which sets out principles of safety, transparency, accountability, and governance running to 2030¹³, to the National Commission now examining the future regulation of AI in healthcare.¹⁴

The direction of travel is not restriction. It is accountability. Every one of these instruments allows AI use and requires a named human to own the output. The June 2026 blog post is that same principle applied to the narrow, high-stakes case of what you send the regulator after an inspection.

What to Do in the Next 30 Days

The MHRA post is short, it is public, and it will be read by your inspector. It is a reasonable assumption that the expectations in it will be applied at your next inspection whether or not they ever appear in formal guidance. The following sequence is achievable inside a month for most organizations.

What good looks like at the next inspection

If an inspector asks how you assured the accuracy of your last response, a strong answer has four parts, and none of them require you to have avoided AI. You can name the procedure that governs how responses are drafted and what tools may assist. You can produce the completed verification record for that specific submission. You can name the qualified people who authored each technical section and the signatory who approved the package. And you can show the CAPA system entries that were created before the response was sent, with owners, dates, and effectiveness checks that match the commitments in the letter word for word.

That answer describes a system. The MHRA has said, in effect, that a system is what it is looking for. Companies with strong verification processes, AI-assisted or not, meet the standard. Those submitting unverified material do not.

Conclusion

The MHRA post is one of the more useful things a regulator has published on generative AI, precisely because it refuses to be about generative AI. It is about verification, accountability, and whether the person who signed a document understood what it said. Those questions predate the technology by decades. What changed is that fluent, confident, entirely unverified prose is now available to anyone under deadline pressure, and inspection responses are written under exactly that pressure. The agency did not respond by prohibiting the tools. It responded by restating the five things a submission has to be and naming what it will do when a submission is not those things.

Our read is that the practical work here is small and the consequence of skipping it is not. A permission boundary, a one-page verification record, a named signatory who knows what the signature means, and a settled position on disclosure. That is a few days of effort, and it converts a genuine risk into a demonstrable control that you can put in front of an inspector. The organizations that will struggle are not the ones using AI. They are the ones that cannot say who verified what, which is the same weakness that produced the Purolea outcome in a different document type.

Sakara Digital works with pharma and biotech organizations building practical governance around AI use in quality and regulatory operations. If you are working out where AI assistance belongs in your inspection responses, CAPA records, and regulatory correspondence, and you want an independent perspective on where to draw the line, we are happy to have that conversation.

For Further Reading