When software changes information in a regulated system, an inspector will ask who is accountable, what was recorded, and whether a person’s change can be told apart from one an AI initiated. Part 11 was written before any of this existed. Here is how to answer anyway.

21 CFR Part 11AI governance

Key takeaways

  1. No regulator has yet published specific guidance on AI-initiated actions in Part 11 systems — the answer has to be built from the existing framework.
  2. Audit recording should be part of how a record is written, not a step each feature has to remember.
  3. Every record answers who, what, when and why — and AI actions are tagged so they can be found and counted.
  4. AI never produces a signature or an approval. Those require a person.

A question the regulations have not answered yet

FDA has not published specific guidance on AI-initiated actions within Part 11 systems, and neither has EMA or PMDA. What exists is the Part 11 framework and the ALCOA+ expectation of data integrity. Any company deploying AI in regulated software has to interpret those for itself — and should expect the interpretation to be tested.

Record as part of writing, not after it

Most enterprise software treats audit logging as a feature: a developer adds a line to record each action after performing it. That has two weaknesses. People forget, and the record is created separately from the change, so the two can drift apart. AI makes both worse, because an AI process may take dozens of actions in quick succession.

The stronger design makes the audit record a consequence of writing the record at all. Nobody decides whether to log a change; it is logged because it happened.

The four questions every audit record answers: who, what and when, and whyEVERY AUDIT RECORD ANSWERS FOUR QUESTIONSWhoThe person, from theirauthenticated sessionFor AI-initiated actions:the AI, tagged distinctly,and the person accountablefor itWhat and whenWhat was affected and whatwas doneThe time according to thesystem — never accordingto whatever askedWhyWhich fields changed, fromwhat to whatIn what context, and forwhat stated reason
Figure 1The same four questions apply whether a person or an AI initiated the change.

ALCOA+, applied to AI

Principle What it means when AI acts
Attributable The record names the AI as the actor and the person accountable for that use of AI.
Legible A consistent, structured format — not free text only its author can interpret.
Contemporaneous The time is recorded as the action happens, by the system’s clock.
Original The audit record is the original record of the event, not a summary of something kept elsewhere.
Accurate and complete Values are checked before they are stored, and gaps in the record are detectable.
Consistent, enduring, available Every part of the platform records the same way; records are retained and searchable on request.

Making tampering visible

Each record can carry a cryptographic fingerprint calculated from its own contents, together with the fingerprint of the record before it. Together they form a chain: alter a record after the fact and its fingerprint no longer matches the copy held by the next one. This is not blockchain — it is a simpler mechanism that makes tampering detectable by anyone who can read the trail.

Records chained by fingerprints: each record stores the fingerprint of the one before it, so altering a record breaks the chainA CHAIN OF FINGERPRINTSRecord 41fingerprinta3f9…previous: …Record 42fingerprint7c1e…previous: a3f9…Record 43fingerprintchanged!previous: 7c1e…Record 44fingerprint9d4a…previous: e02b…Alter record 43 after the fact and its fingerprint no longer matches the copy held by record 44 — the chain visibly breaks.
Figure 2A fingerprint chain makes after-the-fact changes detectable without any special infrastructure.

Telling AI from people

An auditor will increasingly ask questions like “show me every document classification an AI changed last quarter”. The system should answer directly, which means AI-initiated actions are tagged differently from a person’s actions and from the platform’s own internal processes — at the moment they are recorded, not reconstructed afterwards.

AI is not an individual. A signature has to be.

The signature boundary

One line should not move: AI does not sign or approve. Part 11 §11.100 requires an electronic signature to be unique to one individual, and an AI is not one. Signing requires a person to authenticate themselves, and the signature records their identity, the time, the meaning and a fingerprint of exactly what they signed.

Open questions for the industry

Some questions remain open across the industry: whether the audit trail should record which model and version produced a suggestion that led to a change; how chains of AI actions should be linked as one line of reasoning; and what “attributable” should mean when software acts on a person’s behalf. The safest position today is conservative — record more, let AI propose, and keep every consequential step with a person.

How DnXT approaches it

In DnXT, AI-written comments and suggestions are labelled as AI and confirmed by a person before they change a record; AI does not sign; electronic signatures record the signer, time and meaning and are hashed with the content they sign; and content versions, review records and the planning audit trail are chained by fingerprints so that changes after the fact are detectable.

Show your auditors the record

We will walk your quality team through how DnXT records people, AI and signatures.

Book a walkthroughValidating cloud software