What Part 11 says a signature is
21 CFR 11.3(b)(7) defines an electronic signature as a computer data compilation of any symbol or series of symbols executed, adopted, or authorized by an individual to be the legally binding equivalent of the individual's handwritten signature. Section 11.3(b)(8) ties a handwritten signature to a present intention to authenticate a writing. A signature is therefore an attestation by a person: this individual reviewed, approved, or takes responsibility for this record.
Subpart C then makes the link to a person explicit. Section 11.100(a) requires each electronic signature to be unique to one individual and never reused by, or reassigned to, anyone else. Section 11.100(b) requires the organization to verify the identity of the individual before it establishes or sanctions their electronic signature. Neither requirement can be met by software acting on its own behalf.
This page is not legal or regulatory advice; how Part 11 applies to a specific record is a decision for your quality and regulatory functions.
The Part 11 provisions that matter here
Read together, the signature sections of Part 11 assume a human signer at every step. The table maps each provision to what it means when an AI agent is part of the workflow.
| Provision | What it requires | Implication for an AI agent |
|---|---|---|
| 11.3(b)(7) | An electronic signature is executed, adopted or authorized by an individual | The agent is not the individual; it cannot be the signer |
| 11.50(a) | Signed records show the signer's printed name, the date and time, and the meaning of the signature (such as review, approval, responsibility or authorship) | The name and meaning must belong to the person who signs, not to the agent that drafted |
| 11.70 | Signatures are linked to their records so they cannot be excised, copied or transferred to falsify a record | An agent must not copy or apply a stored signature to a record a person has not signed |
| 11.100(a) | Each electronic signature is unique to one individual and not reused or reassigned | No shared or service-account signature for the agent |
| 11.200(a) | Non-biometric signatures use at least two distinct identification components, are used only by their genuine owners, and misuse requires collaboration of two or more individuals | Signature components must never be stored where an agent can use them |
| 11.300 | Controls for identification codes and passwords: uniqueness, periodic checks, loss management, transaction safeguards, device testing | Agent credentials are separate from, and cannot unlock, a person's signing credentials |
Where delegation goes wrong
The risk is rarely an agent that announces it is signing. It is a workflow that quietly lets the agent's output become the approval. Section 11.200(a)(2) says signatures are to be used only by their genuine owners, and 11.200(a)(3) requires that any attempt by someone else to use an individual's signature needs the collaboration of two or more individuals. A design in which one person's credentials are available to an agent breaks that assumption.
- Shared or embedded credentials: a reviewer's user ID and password stored in a script so the agent can click Approve.
- Batch approval of agent output: a person signs a list of records the agent prepared without reviewing each one, so the stated meaning of the signature (review or approval) is not true.
- Service accounts with signing rights: an account used by the agent that is authorized to apply a signature, which conflicts with 11.100(a).
- Copied signature manifestations: the agent inserts a person's name, date and meaning into a document, which is what 11.70 is meant to prevent.
- Unclear accountability: no written policy holding the signer responsible for actions under their signature, which 11.10(j) expects.
What EU GMP Annex 11 adds
EU GMP Annex 11 section 14 says electronic records may be signed electronically and that electronic signatures are expected to have the same impact as hand-written signatures within the boundaries of the company, be permanently linked to their respective record, and include the time and date they were applied. Section 15 goes further for batch release: the system should allow only Qualified Persons to certify the release of batches and should clearly identify and record the person releasing or certifying them.
The European Commission consulted from 7 July to 7 October 2025 on a revised Annex 11 and a new Annex 22 on artificial intelligence. The consultation draft of Annex 22 states that it does not apply to generative AI and large language models, that such models should not be used in critical GMP applications, and that in non-critical applications qualified personnel should always be responsible for ensuring outputs are suitable for the intended use, a human-in-the-loop. That text is a draft and may change; it is not adopted guidance.
What the agent can do, and what it should record
An agent can do most of the work that comes before a signature: assemble the record, check it against the specification, recompute values, flag discrepancies, and route it to the right person. The signature itself stays with a named, qualified individual using their own credentials, applied after they have reviewed the record. OWASP's guidance on Excessive Agency (LLM06:2025) recommends human-in-the-loop control so that a person approves high-impact actions before they are taken; approving a GxP record is such an action.
Instead of a signature, the agent should leave an authorship trail. Section 11.10(e) already expects secure, computer-generated, time-stamped audit trails of entries and actions that create, modify or delete electronic records, and those expectations apply when the actor is an agent.
The Annex 22 consultation draft offers an analogy only, because it is written for static, deterministic models in critical GMP applications and states that it does not apply to generative AI and large language models, on which most agents are built. Within that scope, section 3.3 says that where a model gives an input to a decision made by a human operator, and the effort to test the model has been diminished, the description of intended use should include the operator's responsibility; section 10.5 separately says that, under the same condition, records should be kept from that process. Neither section is adopted guidance. The signer's review is itself part of the record.
- Agent identity and version, and the person who owns the agent.
- The inputs used, including documents, data and the model and version.
- Each action taken and each field drafted or changed, with timestamps.
- Checks performed and their results, including anything flagged for a person.
- Who the record was routed to, and the signer's separate signature event with its stated meaning.
How to set this up in practice
Separate the agent's identity from every human signer's identity, and grant the agent no permission to apply a signature. Use the authority checks that 11.10(g) describes so only authorized individuals can electronically sign. Show the signer what the agent did, so the meaning recorded under 11.50(a)(3) is true. Validate the workflow, including the hand-off, as you would any computerised system, and document the decision about which records are in scope for Part 11 or Annex 11.
Keep the agent's credentials out of the signing path entirely. Section 11.300 expects controls that keep each combined identification code and password unique, check and revise issuances periodically, deauthorize lost or compromised tokens, and detect and report attempts at unauthorized use. An agent running under a person's session, or holding a token that can unlock their signing components, would undermine every one of those controls. Give the agent its own identity for its own actions, and keep signing a separate, interactive step for the person.
Train the people who sign. Section 11.10(i) expects that people who develop, maintain or use electronic record and signature systems have the education, training and experience to do their tasks. When an agent drafts the record, that training should cover what the agent can and cannot check, how to read its flags, and when to reject its output rather than correct it silently.
The safe pattern is an agent that drafts and checks, and a named person who signs. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment of current AI use and data exposure. Status, October 2026. In use in Atlas Bio's own agent platform today: a separate verification step that checks agent work, and a hold that stops a task until a person decides. Built and tested on synthetic data, not yet used with a client: the AI risk assessment tooling (AI discovery, data-flow and vendor-terms mapping, and a NIST AI RMF gap report) and a record of what data an AI used, what it produced and who reviewed it. In development: a customer-controlled AI environment in your own Azure tenant, an AI gateway, agent permissions, evidence linkage and notifications. Each engagement states which controls are built. Decisions on GxP scope, validation and compliance stay with your quality function.
Can an AI agent apply a reviewer's signature on their behalf if they authorize it?
That is the pattern Part 11 is written to prevent. Section 11.100(a) forbids reuse or reassignment of a signature, and 11.200(a)(2) says signatures are used only by their genuine owners. The person should sign the record themselves after reviewing it.
Can an agent be listed as the author of a GxP record?
The agent's contribution should be recorded in the audit trail, with its identity and version. The meaning on the signature, such as authorship, responsibility or approval under 11.50(a)(3), should belong to the person who signs.
Does a digital certificate issued to the agent change the answer?
No. Part 11 defines a digital signature as a type of electronic signature, and electronic signatures belong to individuals. A certificate can authenticate the agent's actions in the audit trail, but it is not a person's signature.
Does this apply to records that are not regulated?
Part 11 applies to records kept under FDA records requirements and to records submitted to the agency. For exploratory work outside those records Part 11 does not apply, though predicate rules and internal policy still may; decide and document the boundary before agents are used.
- eCFR: 21 CFR Part 11 - Electronic Records; Electronic Signatures
- European Commission: EudraLex Volume 4, Annex 11 - Computerised Systems
- European Commission: Stakeholders' consultation on EudraLex Volume 4 - Chapter 4, Annex 11 and new Annex 22
- European Commission: Annex 22 Artificial Intelligence - consultation draft
- OWASP GenAI Security Project: LLM06:2025 Excessive Agency
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →