As AI agents review, classify and check documents, there will be pressure to let them approve or sign too. Part 11 already answers the question: a signature belongs to an individual, and an AI is not one.
Key takeaways
- Part 11 requires an electronic signature to be unique to one individual — an AI agent cannot satisfy that.
- The workable pattern: AI reviews and recommends; a person judges and signs; both are recorded.
- A rule that AI “must not sign” is only as strong as how it is enforced — architecture beats policy.
- Start restrictive. Expanding later is far easier than taking capability away.
If an AI agent can review a document, classify it, check it against validation rules and confirm it meets requirements, why can it not sign it? Because 21 CFR Part 11 says electronic signatures must be attributable to specific individuals. This is not a technical limitation; it is a regulatory requirement with good reasons behind it.
What Part 11 actually requires
An AI agent fails §11.100 immediately: it is not one individual, and it may act for many users through a shared service account. Even if each agent were given its own identity, the purpose of the rule — a named person accountable for a regulated decision — would not be met.
The pattern that works
This is not a workaround. It is how AI should take part in regulated signing workflows: everything up to the decision can be accelerated, and the decision itself stays with a person whose identity, time and stated meaning are bound to exactly what they signed.
Telling an AI not to sign is a policy. Not giving it a way to sign is a control.
Enforce it in the architecture
An instruction in a prompt or a configuration file is a convention, and conventions can be bypassed. The stronger design gives AI no route to sign at all: it can check the status of a signature, list signatures on a document, request a signature from a named person, or verify that a signature is intact — but there is no function that applies one.
Underneath, signatures should be immutable once applied (revocation creates a new record rather than editing the old one), move from pending to signed only when the signer authenticates themselves, carry a meaning from a controlled list, and be bound by a cryptographic hash to the exact document version.
What is still unsettled
FDA has not addressed AI and Part 11 signatures directly; Part 11 dates from 1997. Practice is evolving — some organisations treat a person’s signature as attesting both their own judgment and their review of the AI’s analysis. EU Annex 11, Japan’s ERES guidelines and other frameworks add their own expectations. A global company needs an approach that holds in all of them, which today means keeping signing with people.
Three questions for any vendor: Can an AI agent technically call a sign function, even if told not to? How does the audit trail separate the AI’s analysis from the person’s signature? Is the signature bound to the exact version the person reviewed?
How DnXT approaches it
In DnXT, AI can review, comment and recommend, and its output is labelled as AI. Signing is done by a person who authenticates in DnXT; each signature records the signer, the time and its meaning and is hashed with the content it signs.
AI that never signs
See how DnXT keeps signatures with people while AI speeds up everything before the decision.