Manufacturing deviations and CAPA investigation

• When a deviation is raised, the investigator sees the most similar past deviations and their outcomes, matched by meaning across product, equipment, process step and failure mode, not just keywords.
• The assistant suggests likely root-cause areas, each linked to the past cases and data behind it, and prompts for the evidence that would confirm or rule each one out.
• A draft outline of the investigation is assembled: the event description, a timeline from MES and LIMS data, and the impact questions the site's SOP requires. The investigator writes the analysis and the conclusions.
• A trend model looks across all deviations each week and flags rising or recurring patterns by equipment, material, area or process step, for the quality review meeting.
• When a CAPA is closed, the same failure is watched for. If it recurs, the CAPA's effectiveness check is flagged for review.
• The AI is advisory throughout. QA classifies, approves the root cause and the CAPA, and decides batch disposition. The Qualified Person certifies. Nothing the AI produces sits in the batch release path.

• MHRAUK GMP: the Human Medicines Regulations 2012 and the 'Orange Guide'
  →

  The manufacturing licence holder runs a pharmaceutical quality system that investigates deviations, finds the root cause and takes corrective and preventive action. Those are QA's decisions. The assistant helps the investigator find and organise evidence; it never classifies, approves or closes a deviation or a CAPA.

• MHRAThe Qualified Person
  →

  Each batch is certified by a Qualified Person, who must be satisfied that deviations affecting it have been investigated. The assistant has no connection to certification, and nothing in the release path waits for, or depends on, its output.

• ICHQ9(R1) quality risk management and Q10 pharmaceutical quality system
  →

  Q9(R1) asks for risk assessments that manage subjectivity. Suggestions from past cases can anchor investigators on old answers, so the assistant shows evidence for and against, and the investigator records their own reasoning. Q10 expects CAPA effectiveness to be checked and reviewed by management; the recurrence watch feeds that review, it doesn't replace it.

• MHRAGxP data integrity guidance (2018) and computerised system validation
  →

  Records must be attributable, legible, contemporaneous, original and accurate (ALCOA+). The investigation report is the investigator's record, signed in the eQMS. The assistant's suggestions are logged with their sources and model version, and the tool is validated for its intended use, using a risk-based approach such as ISPE's GAMP 5 (second edition) and its GAMP guide on AI (July 2025) as industry guidance.

• European Commission and PIC/SEU GMP Annex 11 revision and draft Annex 22 on AI
  →

  Consulted on from July to October 2025 and still in draft in November 2026. The draft Annex 22 covers static, deterministic models in critical GMP uses and says generative AI and large language models shouldn't be used in critical GMP applications; in non-critical uses, trained people are responsible for the output. It applies directly to sites supplying the EU and is a strong signal for everyone else, so this design keeps the language model in a non-critical, advisory role.

• ICOUK GDPR and the Data Protection Act 2018
  →

  Deviation records name operators and investigators. The trend model looks at processes, equipment and materials, not individual people's performance, and it isn't used for disciplinary decisions. That needs a DPIA, a clear purpose and a retention period, and staff should be told how the tool is used.

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

• The investigator owns the investigation. Suggestions are evidence to consider, and the report records which were used and why.
• Only closed, QA-approved investigations go into the index, so the assistant never learns from drafts.
• QA reviews every investigation that concludes 'human error' alone, whether or not the AI was involved.
• A named owner for the model, the index and the prompts, with change control and re-validation like any other GMP computerised system.
• A monthly review of suggestion usefulness and of trend flags acted on or dismissed.
• Investigators trained before go-live, including what the assistant can't see: anything that was never written down.

1. 3–4 weeks

   Discovery

   Map today's deviation process, measure the baseline, check the quality of past investigation data, agree the eQMS route with its supplier, and agree the intended use and risk level with QA.

2. Gate · must be true to passBaseline agreed, intended use written down, validation approach agreed with QA, DPIA started.

3. 8–10 weeks

   Pilot

   One production area in shadow mode: suggestions are produced, investigators work as normal, and QA compares the two.

4. Gate · must be true to passSuggestion relevance within agreed limits, no influence on classification or disposition, validation evidence complete, head of QA sign-off.

5. 4–6 weeks

   Production

   Suggestions go live for the pilot area, with monitoring, support and a rollback plan. The trend report joins the quality review meeting.

6. Gate · must be true to passMonitoring, support model, rollback plan and investigator training in place.

7. Ongoing

   Scale

   Add areas and QC laboratories one at a time, then reuse the pattern for complaints and change control.

8. Gate · must be true to passEach new area passes the same pilot gate.

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

1. 1

   Goes wrong

   Investigators copy the suggested root cause because it matched last time.

   How the design avoids it

   Evidence for and against each suggestion, the investigator's own reasoning recorded, and QA checking how often suggestions were adopted unchanged.

2. 2

   Goes wrong

   The index is full of old 'human error' conclusions, so the assistant suggests more of them.

   How the design avoids it

   Index only QA-approved investigations, weight recent ones, and review the root-cause categories before go-live.

3. 3

   Goes wrong

   The tool drifts into the release decision because a QP starts relying on it.

   How the design avoids it

   No link to certification, guardrails against disposition language, and the intended use written down and trained.

4. 4

   Goes wrong

   Someone connects the AI straight to MES on the OT network.

   How the design avoids it

   Read-only extracts from a reporting copy, through the DMZ, agreed with OT security in discovery.

5. 5

   Goes wrong

   Validation is left until the end, or skipped because 'it's only advisory'.

   How the design avoids it

   Agree the intended use and risk level in discovery and validate proportionately from the start.

1. Q4 2026 target
   Final texts of EU GMP Annex 22 on AI, the revised Annex 11 and Chapter 4, due to go to the European Commission. Regulators held a workshop in mid-2026 on a risk-based approach to generative AI, so the language model position may change.

   Design to the draft now: language models advisory only, static and versioned models where it matters, and the intended use written down. Review against the final text when it's published.

2. 2026–2030
   The MHRA's Strategy 2030 (October 2026), including a Centre for AI and Digital Health to set out expectations for AI used across the medicines lifecycle.

   Keep the validation pack, the decision logs and the intended use current, so you can show how the AI is controlled when asked.

3. 2027
   The MHRA's AI Medicines Development Action Plan, promised in the strategy.

   Watch it for any expectations on AI in manufacturing and quality, and map them to what you already do under the draft Annex 22.

4. 2028
   The NCSC's post-quantum cryptography milestone: a full migration plan by 2028.

   GMP records are kept for years. Record which encryption the new data flows use, and keep it easy to change.