---
title: What evidence do AI agents need in GxP-regulated life sciences?
url: https://opencontroller.lyzr.com/blog/gxp-evidence-for-ai-agents-in-life-sciences
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# What evidence do AI agents need in GxP-regulated life sciences?

**An AI agent that touches GxP records needs the same evidence as any computerised system: a validated, version-controlled configuration, secure time-stamped audit trails, controlled access, and change control. On top of that, the EU's draft GMP Annex 22 expects defined intended use, independent test data, confidence thresholds and human review, and says generative AI and LLMs should not be used in critical GMP applications. Plan agent use cases around that line.**

Pharmaceutical, biotech and medical device companies are piloting agents for deviation summaries, batch record review, document control and regulatory writing. The question from quality assurance is always the same: can we evidence this the way we evidence everything else? This guide sits under our pillar on [mapping AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks). For background on the category, see [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## Which rules apply today?

| Rule | Status | What it means for agents |
|---|---|---|
| FDA 21 CFR Part 11 | In force | Electronic records and signatures need validation, audit trails, access controls and record retention |
| EU GMP Annex 11 (Computerised Systems) | In force; a revised draft was consulted on in 2025 | Risk-based validation, data integrity, audit trails and periodic review |
| EU GMP Annex 22 (Artificial Intelligence) | Draft, published for consultation in July 2025; consultation closed October 2025 | AI-specific expectations for models in GMP use |
| FDA draft guidance on AI to support regulatory decision-making for drugs and biologics | Issued as draft in January 2025 | A risk-based credibility assessment framework for AI models |

Part 11 section 11.10(e) requires secure, computer-generated, time-stamped audit trails that independently record operator entries and actions that create, modify or delete electronic records, without obscuring earlier entries. If an agent creates or edits a GxP record, the agent is an operator in that sense, and its actions need the same trail.

## What does draft Annex 22 say about agents?

Annex 22 is still a draft. As of August 2026 it had not appeared as a final text in EudraLex Volume 4, and EMA's work plan targeted a final text for the European Commission in Q4 2026. Check its status before relying on it. The draft position is clear on several points:

- It covers static, deterministic machine learning models used in GMP.
- Dynamic models that keep learning in use are out of scope and should not be used in critical GMP applications.
- Generative AI and large language models should not be used in critical GMP applications. In non-critical uses, qualified personnel must confirm outputs are fit for purpose.
- Intended use must be defined, including the human operator's responsibility where a model informs a decision.
- Test data must be independent of training data, with acceptance criteria approved before testing.
- Confidence thresholds should be set, with outputs below them flagged as undecided.
- Models fall under change and configuration control.

Most agents run on LLMs. Under the draft, that steers them towards non-critical tasks with qualified human review: drafting, summarising, searching and preparing, not releasing batches or making quality decisions.

## What evidence should an agent have?

Use this checklist for any agent in a GxP environment.

1. **Intended use statement.** What the agent does, which records it reads and writes, and whether the use is critical. Approved by quality.
2. **Risk assessment.** Impact on product quality, patient safety and data integrity, following your existing computerised system validation approach.
3. **Versioned configuration.** Agent version, model and model version, prompts, tools and data sources, under change control. A model provider update is a change.
4. **Validation records.** Test cases, acceptance criteria approved before testing, independent test data, and results.
5. **Audit trail.** Time-stamped records of every action that creates, modifies or deletes a GxP record, attributed to the agent and to the person it acted for.
6. **Access control.** A unique identity per agent and least-privilege permissions to GxP systems.
7. **Human review record.** Who reviewed each output, what they decided and when.
8. **Periodic review.** Scheduled checks of performance, incidents and drift, with results recorded.
9. **Retention.** Audit trails kept at least as long as the records they relate to, as Part 11 requires.

Our guides to [AI agent audit log fields](/blog/ai-agent-audit-log-fields) and [how long to retain AI agent logs](/blog/how-long-to-retain-ai-agent-logs) go deeper on items 5 and 9.

## How do you keep agents inside a validated state?

The hardest part is change. A hosted model can change behind a stable name. Pin dated model versions where the provider offers them. Treat prompt edits and tool changes as configuration changes. Re-run the validation test set before any change goes live, and keep the results. An agent registry that records every version, and which version produced which output, makes this auditable.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, runs multi-turn and red-team evals, and scores production traffic with LLM judges. Policy as code has built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, and can be extended with your own GxP controls. It deploys as Lyzr SaaS or on customer-controlled infrastructure. These produce evidence for your validation; they do not replace your quality system. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can LLM-based agents be used in GMP at all?**
Under draft Annex 22, generative AI and LLMs should not be used in critical GMP applications. Non-critical uses with qualified human review remain possible. Check the final text when it is published.

**Is an AI agent an operator under 21 CFR Part 11?**
If it creates, modifies or deletes electronic records, its actions need the secure, time-stamped audit trail that section 11.10(e) requires, attributed to the agent and the person it acted for.

**Does a model update require revalidation?**
Treat it as a change under change control. Assess impact, re-run the validation test set and record the result before the new version goes live.

## Key facts

- 21 CFR 11.10(e) requires secure, computer-generated, time-stamped audit trails of operator entries and actions that create, modify or delete electronic records.
- EU GMP Annex 22 (Artificial Intelligence) was published as a draft for consultation in July 2025; consultation closed October 2025; as of August 2026 no final text was in EudraLex Volume 4, with EMA targeting a final text in Q4 2026.
- Draft Annex 22 covers static, deterministic models and says generative AI and LLMs should not be used in critical GMP applications.
- Draft Annex 22 expects defined intended use, independent test data, pre-approved acceptance criteria, confidence thresholds and change control.
- The FDA issued draft guidance on AI to support regulatory decision-making for drugs and biologics in January 2025.
- Opencontroller provides a versioned agent registry, evals, LLM judges and policy as code with built-in framework mappings; it produces evidence, not compliance.
- Opencontroller starts from $250K per deployment per year.

## Related

- [What should an AI agent audit log contain?](https://opencontroller.lyzr.com/blog/ai-agent-audit-log-fields.md)
- [How long should you keep AI agent logs and traces?](https://opencontroller.lyzr.com/blog/how-long-to-retain-ai-agent-logs.md)
- [What evidence does ISO/IEC 42001 expect for AI agents?](https://opencontroller.lyzr.com/blog/iso-42001-evidence-for-ai-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
