Expect an auditor to ask first for your register of AI agents and how each was classified under the EU AI Act. Then come the provider's instructions for use, evidence of human oversight, the action logs, monitoring reports, training records, notices to people, and any impact assessments. Each must match the agent as it runs today.
What documentation does the Act require a deployer to hold?
The Act does not hand deployers a single documentation list. The records follow from the duties. A deployer of high-risk systems must use them according to the instructions, assign competent oversight, control input data, monitor and report, keep logs, inform workers and affected people, and, in some cases, complete a fundamental rights impact assessment. Every deployer must ensure AI literacy and disclose AI interaction where Article 50 applies.
Each duty leaves a document or a record. An auditor asks for the record because it is the only way to see the duty being met. If you cannot produce it, the auditor will conclude the duty was not met, whatever actually happened.
Who will ask for it?
National market surveillance authorities enforce the Act. For financial institutions, existing financial supervisors are often given the role for AI used in their sector, as explained in who enforces the EU AI Act for AI agents. Internal audit, external auditors reviewing controls, and customers performing due diligence will ask for overlapping sets.
Firms that run agents in the UK as well as the EU should note that UK regulators ask through existing rules rather than a single AI law; EU AI Act vs the UK approach to AI agents compares the two. Building one evidence set that serves both saves duplicate work.
Even agents used only by staff are in scope. Internal use is still deployment, and whether the EU AI Act applies to agents used only internally depends on what they do, not who sees them.
What makes agent documentation different?
A model's documentation can describe one output. An agent's documentation must describe what it can do, which changes every time a capability, source or connection is added. Three consequences follow:
- Documentation is per capability. Classification, oversight and testing differ between reading a balance and sending a payment.
- Configuration is part of the evidence. The brief, the sources and the limits are as important as the model. An auditor will want the version that applied on a given date.
- The action record carries most of the weight. For agents, the log is where oversight, monitoring and incident handling become visible. Our view, set out in auditors will ask for the verdict record, is that a record of each decision and its reason is the document an auditor reads first and trusts most.
What should the evidence file contain?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Know what you run | Agent register: each agent, capability, owner, provider, role | Register, versioned, signed by owners |
| Classify correctly | Classification per capability against the Act's tiers | Classification memo with reasoning |
| Use as instructed | Provider's instructions on file; configuration mapped to them | Instructions version and mapping |
| Human oversight | Named overseers per capability; hold points before adverse or irreversible actions | Assignments, training, approval records, override rates |
| Input data | Source register with owner, refresh and trust class | Register and source references in each action record |
| Logs | Append-only action record covering every allowed, held and refused action | Retained records; integrity checks |
| Monitoring | Periodic review of agent behaviour and outcomes | Reports, actions taken, suspensions |
| Incidents | Escalation route to provider and authority | Incident log with notification times |
| AI literacy | Role-based training | Records by role; quality checks on approvals |
| Transparency to people | Disclosure in each channel | Text and channel captures |
| Notices to affected people and workers | Templates and consultation | Copies sent; minutes |
| Impact assessments | FRIA where required; DPIA where GDPR requires | Signed assessments and update history |
| Configuration history | Change control on brief, sources, limits and model | Change log with approvals |
The classification memo is worth writing carefully. If you decided an agent was not high-risk, the auditor will test that reasoning first; the method is in how to classify an AI agent under the EU AI Act.
How do you keep it current?
Most documentation fails by going stale, not by being absent. Three habits help:
- Tie documents to change. Any change to an agent's capabilities, sources or limits must trigger review of the register entry, classification, oversight assignments and training. Make it a step in change control, not a reminder.
- Generate what you can from the record. Monitoring reports, override rates and refusal counts should be produced from the action record, not written by hand. Generated evidence is current by construction.
- Date everything. Each document should show the period it applies to. An auditor reviewing a case from six months ago needs the configuration and instructions that applied then.
What will a supervisor or auditor ask?
- Give us the register and the classification for each agent.
- For this case, show what the agent did, what it relied on and who approved.
- Which instructions for use applied on that date?
- Show us the overseers' training and their approval history.
- What did monitoring find, and what changed as a result?
- Show us an incident and how it was handled.
What will the documentation not cover?
It will not cover provider obligations such as technical documentation and conformity assessment, though you should hold the provider's declaration and instructions. It will not substitute for GDPR records, which have their own requirements. And it will not answer questions under sector rules that ask for different evidence, such as a bank's model validation. For how these fit, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
How long should agent documentation be kept?
At least as long as the logs it explains, and in line with your sector's record-keeping rules. Keep superseded versions, because a case can only be judged against the configuration that applied at the time.
Can documentation be kept in the agent platform?
Yes, if it is exportable, access-controlled and cannot be silently changed. An auditor must be able to receive it in a usable form without depending on a live login.
What is the most common gap auditors find?
Records that show what the agent said but not what it did, and oversight that exists on paper but leaves no trace of anyone approving or overruling anything.