An MHRA inspection does not end when the inspectors leave the building. For most companies, the document that carries the most weight is the one written afterwards: the response to the deficiency report. It is the company’s chance to show that it understands what went wrong, has dealt with the immediate risk, and has changed something so that the problem does not return. It also shapes how the compliance team classifies the outcome, whether further action is needed, and how soon the site is likely to be seen again.
In June 2026 the MHRA Inspectorate added a new dimension to this. In a post on its inspectorate blog dated 29 June 2026, it described AI-generated material appearing in responses sent to its compliance teams, and set out what it expects. The message was not “do not use AI”. It was closer to “whatever you use, you remain accountable for every word”. This guide explains what the MHRA said, why generic responses fail even without AI, how a strong response is built, and how to verify an AI-assisted draft before it goes anywhere near an inspector. If you are searching for how to respond to an MHRA inspection report, MHRA deficiency response help, GDP or GMP inspection response support, or CAPA writing for MHRA findings, this is written for you.
What the MHRA Actually Said
The inspectorate’s post relates specifically to submissions made to MHRA compliance teams following GxP inspections, which covers GMP, GDP, GCP and pharmacovigilance. Its main points are:
- Organisations are using AI tools to draft inspection responses, and the MHRA has seen examples of this
- Problems it has encountered include references to MHRA guidance that does not exist, citations of regulatory frameworks that are not relevant, and responses to serious deficiencies that appear designed to mislead rather than address the underlying problem
- Submissions must be factually accurate and verifiable, technically reviewed by experienced people, supported by evidence and signed off by an accountable individual
- Disclosure of AI use in a response is offered as an option, not a requirement
- Poor quality, inaccurate or overly generic responses may be rejected, returned for revision, treated as a sign of higher inspection risk, or escalated for regulatory action
- Companies with robust processes meet the standard whether or not AI was involved
The closing line of the post is the one most worth remembering. Those who use AI responsibly will thrive, while those who use technology to obscure inadequate responses, fill resource or expertise gaps, or hide a lack of knowledge behind serious deficiencies will be scrutinised. One reported example was a response running to more than ninety pages that never addressed the deficiencies raised.
Notice what the MHRA did not do. It did not name approved or banned tools, and it did not set a rule about drafting methods. It set an outcome standard, which means the burden sits entirely with the company: can you prove that every statement is true?
Why Generic Responses Fail, With or Without AI
AI has not created the problem of poor responses. It has made them faster to produce and more fluent, which makes them harder to spot at a glance. The failures inspectors see are familiar:
- Responding to the wording of the observation rather than to the underlying failure
- Naming a cause such as “human error” or “lack of awareness” and stopping there
- Promising to “review” or “reinforce” a procedure, with no owner, date or measure of success
- Correcting the single example cited and ignoring where else the same weakness may exist
- Attaching evidence that does not match the claim, such as a training record that predates the event
- Padding the response with regulatory citations that do not connect to the finding
A fluent, confident paragraph that says none of the things an inspector needs to hear is exactly what a language model produces by default. That is why the MHRA’s concern is less about the tool and more about what happens when nobody with real knowledge reads the output critically.
The Anatomy of a Strong Response
For each deficiency, a response that survives scrutiny covers the same ground in the same order. It helps to build a standard template and use it every time:
- Acknowledge the finding in plain terms, without softening or arguing, unless there is a genuine factual error, which should be stated with evidence
- Assess impact: which products, batches, patients, records, time periods and customers could be affected, and what was concluded about risk
- Describe immediate containment: what was done at once to stop the problem continuing, with dates
- Identify the root cause using a recognised method, and show the reasoning, including why other causes were ruled out
- Set out corrective and preventive actions, each with an owner, a due date and the evidence that will show completion
- Define the effectiveness check: how and when you will confirm the action worked, and what result would count as failure
- Extend the review: where else the same weakness could exist, and what you checked
- Attach evidence that genuinely supports each claim, indexed to the paragraph it supports
An inspector reading this should be able to trace every claim to a document without having to ask. If they would need to ask, the response is not finished.
Treating Critical, Major and Other Deficiencies Differently
Not all deficiencies are equal, and the response should reflect that. For critical and major findings, which suggest a significant departure from the required standard or a risk to product quality or patient safety, the inspector expects to see senior ownership, a clear risk assessment, and visible evidence of change already under way, not just a plan. For lower-level findings, a proportionate response is better than a long one. A short, precise answer to a minor point looks confident. A ten page answer to the same point looks anxious.
A Verification Workflow for AI-Assisted Drafts
If your team uses AI to help draft, treat the output as an unverified first draft written by someone who has never seen your site. A simple six step workflow keeps you inside the MHRA’s expectations:
- Step one, ground the tool: give it only the inspection report text and your own approved facts, and instruct it to cite nothing it was not given
- Step two, build a citation log: every regulation, guideline or MHRA document named in the draft is opened and read by a person, and recorded with the exact section relied on. If it cannot be found, it is deleted
- Step three, fact check against records: every date, batch number, headcount, procedure number and result is checked against the source document, not against the draft
- Step four, technical review: a subject matter expert who was not the drafter confirms that the root cause analysis and the actions make technical sense
- Step five, accountable sign off: a named senior person, such as the Responsible Person, QPPV, Head of Quality or Qualified Person, reads the complete response and signs it
- Step six, retain the trail: keep the version history, the citation log and the review records so that you can show how the final text was produced
The test to apply at the end is simple. If the inspector phoned the signatory tomorrow and asked them to explain any sentence, could they do it without looking anything up?
Should You Disclose AI Use?
The MHRA offers disclosure as an option and has said that transparency is welcome. A practical approach is to include a short statement that AI tools supported the drafting of named sections, and that all content was verified by named individuals against source records. This is useful when the response is long or technical, and it removes any suspicion that something was hidden. It is not a substitute for accuracy. A disclosure on top of an unverified response simply tells the inspector exactly where to look first.
If you decide not to disclose, the standard does not change. The response must still be one that its signatory fully understands and can defend.
Seven Warning Signs in a Draft Response
Before submission, read the draft looking for these:
- Phrases such as “robust”, “comprehensive” or “best practice” with nothing measurable behind them
- References to guidance or regulations that nobody on the team can put their hand on
- Actions without owners, dates or evidence
- Statements that contradict the facts recorded in the inspection report
- The same wording repeated across different deficiencies
- A tone that argues with the inspector instead of answering the question
- A length out of proportion to the finding
An Illustration: Before and After
Consider a finding that a wholesaler’s temperature excursion procedure was not followed for a delivery. The example below is an invented illustration.
A weak response reads: “The company takes temperature control extremely seriously and has reviewed its procedures in line with applicable guidance. Staff have been reminded of the importance of following the process. The company is committed to continuous improvement.”
A strong response reads: “On 14 March a delivery of refrigerated stock arrived with a logger showing a reading above range for 3 hours and 40 minutes. The goods were not quarantined, and 6 units were supplied to two customers before the excursion was identified on 16 March. Both customers were contacted on 16 March, and the units were returned and quarantined by 18 March. A stability assessment concluded that the stock should be rejected, and the Responsible Person’s decision is recorded in deviation 26-041. Root cause analysis found that the goods-in procedure did not require logger data to be reviewed before put-away, and that the receiving team had not been trained on the requirement. The procedure was revised on 28 March, the whole team was trained and assessed by 4 April, and a monthly check of 20 consecutive deliveries will confirm effectiveness, with the first review due on 30 April. Evidence is indexed as items 1 to 6.”
The second version is only a few lines longer, and it gives an inspector everything needed to close the point. Notice too that nothing in it could have been produced by an AI tool without access to the company’s own records, which is exactly why the human verification step matters.
Timing and Practical Management
A response is a project, and it needs managing like one:
- Read the report once for the facts and once for what each observation implies about your system
- Agree early who owns each deficiency, and set an internal deadline well before the regulator’s
- Ask for an extension in good time if a root cause investigation genuinely needs longer, rather than submitting something half-finished
- Use a standard template so that every answer covers the same eight points
- Keep a record of every reference and every piece of evidence as you go, not at the end
- Hold a final read-through with someone who was not involved in writing it
- Agree your position on disclosure of AI use before drafting begins
Remember too that a response is read by people who have inspected many sites. They know what a real root cause analysis looks like and what a template looks like. The most credible responses are specific, modest in tone and honest about what is still in progress.
A Second Pair of Expert Eyes Before You Submit
A deficiency response is one of the few documents in regulated business where the reader’s judgement about your competence is formed from a few pages of text. Quality and Vigilance Ltd supports companies at exactly this point, drawing on its work across GvP, GMP and GDP audits and inspection support, as well as its pharmacovigilance and medical device services. An independent reviewer who has no stake in the original finding often spots the gaps that the writing team cannot see, whether the draft was written by hand or with AI assistance. The help that makes the biggest difference usually looks like this:
- Independent review of a draft response against the inspection report, checking that every observation is answered and every claim is supported by evidence
- Verification of every regulatory reference, so that nothing cited is misquoted, irrelevant or non-existent
- Root cause analysis and CAPA design support, including effectiveness checks that an inspector will accept
- Preparation of a clear evidence index and a defensible, accountable sign-off process
- Advice on whether and how to disclose AI use, and on putting a verification workflow in place for future responses
Get in touch with Quality and Vigilance Ltd today to have your MHRA response independently reviewed before it is submitted