Every safety team believes its audit trail is fine, until an inspector picks a case and asks a simple question: who changed this field, when, why, and what did it say before? If the answer takes a day of searching mailboxes and spreadsheets, the audit trail is not fine.
This post explains what the rules ask of an audit trail, what a complete one contains, where the gaps usually are, and what changes when AI starts creating and editing safety data.
What the rules say
United States: 21 CFR Part 11
Section 11.10(e) requires “secure, computer-generated, time-stamped audit trails to independently record the date and time of operator entries and actions that create, modify, or delete electronic records.” Record changes “shall not obscure previously recorded information”, and the audit trail must be kept at least as long as the records it covers and be available for agency review and copying.
FDA’s December 2018 questions and answers on data integrity define an audit trail as “a secure, computer-generated, time-stamped electronic record that allows for reconstruction of the course of events”. The same guidance sets out the ALCOA expectation that data be attributable, legible, contemporaneously recorded, original or a true copy, and accurate. It also notes that when login credentials are shared, “a unique individual cannot be identified through the login”.
European Union: Annex 11
Annex 11 sits in the EU GMP guide rather than the pharmacovigilance guidance, but it states the audit-trail principle more plainly than anywhere else. Section 9 says that, based on a risk assessment, consideration should be given to building into the system “the creation of a record of all GMP-relevant changes and deletions”; that “for change or deletion of GMP-relevant data the reason should be documented”; and that audit trails “need to be available and convertible to a generally intelligible form and regularly reviewed”.
The EU and PIC/S published a draft revision of Annex 11 for consultation on 7 July 2025, with comments due by 7 October 2025. It includes a dedicated section on audit trails, and was published alongside a new draft Annex 22 on artificial intelligence. Until the revision is final, the current text applies.
Data integrity: ALCOA+
The UK MHRA’s GXP data integrity guidance of March 2018 builds on ALCOA with four further attributes, often called ALCOA+: complete, consistent, enduring and available. An audit trail is how you prove the first five and keep the last four.
Anatomy of a complete entry
Put the rules together and every entry in a pharmacovigilance audit trail should answer six questions.
The sixth question, how, is the one most systems were not designed for. It matters as soon as automation or AI is involved, because “the system changed it” is not an answer an inspector will accept.
What an inspector wants to see
Inspectors rarely read an audit trail end to end. They pick a case and ask you to walk them through it. A good audit trail lets you do that from one screen, in minutes.
The gaps we see most often
| Gap | Why it matters | What good looks like |
|---|---|---|
| Work done outside the system (email, spreadsheets, shared drives) | Decisions leave no trace in the audit trail | Every decision is made, or at least recorded, in the system of record |
| Shared or generic logins | No one can be identified as the person who acted | Named accounts; shared accounts only for read-only viewing |
| Changes without a reason | The ‘why’ cannot be reconstructed | A reason required for changes to critical fields |
| Audit trail that can be switched off or edited | The record cannot be trusted | Always on, protected from change, including by administrators |
| Audit trails nobody reviews | Problems are found by inspectors, not by you | Risk-based review of critical data, documented |
| Unreadable exports | Evidence cannot be shown in a usable form | Readable, filterable history for any case or user |
| History lost in migration or with vendors | Gaps in the record for older cases | Audit trails carried across systems and partners |
How AI changes the picture
When AI reads a source document and fills in case fields, it is creating and modifying safety records. That puts it squarely inside the scope of the rules above. Three things follow.
- Record the AI’s action separately from the human decision. The audit trail should show what the AI proposed, what a person accepted, changed or rejected, and who that person was. Merging the two hides exactly what an inspector wants to see.
- Record the model version and configuration. When a model or its prompts change, you need to know which records were produced by which version. Treat model and prompt changes like any other configuration change, under change control.
- Keep the source link. Each value the AI extracts should point to the place in the source it came from, so the reviewer and the inspector can check it.
The CIOMS Working Group XIV report on AI in pharmacovigilance (December 2025) lists governance and accountability among its seven guiding principles, and offers a governance framework to help organisations document actions, manage change and ensure traceability. A complete audit trail is where that principle becomes evidence.
Reviewing the audit trail
Annex 11 asks for audit trails to be “regularly reviewed”; FDA’s guidance says that where regulations do not set a review frequency, you should set one “using knowledge of your processes and risk assessment tools”. In practice, a risk-based approach for pharmacovigilance might look like this:
- Critical case data, such as seriousness, dates that start the clock, reportability and causality, reviewed as part of case QC before submission.
- System-level activity, such as user access changes, configuration and model changes, and deletions, reviewed on a fixed schedule.
- Targeted reviews after a deviation, a late case or a change in a vendor’s process.
Whatever you choose, write it in an SOP, train people on it and keep evidence that the reviews happened.
A quick self-check
- Can you show the full history of any case, including AI actions, in under ten minutes?
- Does every change to a critical field carry a reason?
- Is every action attributable to a named person or a named, versioned automated process?
- Can anyone, including an administrator, switch the audit trail off or edit it?
- Is audit trail review defined in an SOP, and is there evidence it happens?
- Do audit trails survive system migrations and hand-offs to partners?
Sources
- 21 CFR 11.10, Controls for closed systems
- FDA, Data Integrity and Compliance With Drug CGMP: Questions and Answers (December 2018)
- ECA Academy, quoting EU GMP Annex 11, section 9 (Audit trails)
- NSF, Revised EU/PIC/S GMP on documentation, computer systems and AI (draft Annex 11 and Annex 22, July 2025)
- Lachman Consultants, summary of MHRA ‘GXP’ Data Integrity Guidance and Definitions (March 2018)
- CIOMS Working Group XIV, Artificial Intelligence in Pharmacovigilance (report published 4 December 2025)