---
title: What evidence will a regulator ask for about your AI agents?
url: https://opencontroller.lyzr.com/blog/regulator-inspection-evidence-pack-for-ai-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What evidence will a regulator ask for about your AI agents?

**Expect four kinds of request: what AI agents you run and who owns them, how you assessed and tested them, how you control them in operation, and what happened when something went wrong. Keep a standing evidence pack of about twelve artefacts, each with a source system, a refresh frequency and an owner, so you can produce it in hours rather than weeks.**

Being "inspection-ready" was the goal in the call that prompted this post. The common failure is not missing controls; it is evidence spread across spreadsheets, tickets and screenshots that nobody can assemble under a deadline. Below is a pack you can build now. It produces evidence for regulatory frameworks; it does not by itself make you compliant with any of them.

## Which rules are regulators likely to reference?

Know what is in force and what is not yet.

- **DORA (Regulation (EU) 2022/2554)** has applied to EU financial entities since 17 January 2025. It requires ICT risk management, ICT incident management and reporting, and a register of information on ICT third-party arrangements (Article 28). AI model providers and agent platforms are ICT third parties.
- **EU AI Act** obligations for deployers of high-risk systems (Article 26) include using systems per instructions, human oversight, monitoring and keeping automatically generated logs for at least six months. Under the Digital Omnibus on AI, Regulation (EU) 2026/1744, high-risk obligations for Annex III systems apply from 2 December 2027 and for Annex I product-related systems from 2 August 2028.
- **Sector supervisors** reuse existing rules on outsourcing, operational resilience and conduct. Insurers in many US states follow the NAIC Model Bulletin; see [AI governance evidence for insurers](/blog/ai-governance-evidence-for-insurers).
- **Management-system standards** such as ISO/IEC 42001 and frameworks such as the NIST AI RMF are not laws, but inspectors often use them as a yardstick.

## What should the evidence pack contain?

| # | Artefact | Source system | Refresh | Owner |
| --- | --- | --- | --- | --- |
| 1 | AI agent inventory with type, platform, risk tier | Agent registry / discovery | Daily | AI governance lead |
| 2 | Named owner per agent, linked to identity provider | Registry + IdP | Daily | AI governance lead |
| 3 | Agent identities and permissions, including on-behalf-of scope | IAM / IdP | Weekly | Identity team |
| 4 | Tools and MCP servers per agent, with write capability flagged | Registry | Daily | Security architecture |
| 5 | Third-party model and platform providers per agent | Registry + vendor register | Monthly | Third-party risk |
| 6 | Pre-release test results (evals, red-team) per version | Eval platform | Per release | Agent owner |
| 7 | Production quality scores (groundedness, task adherence) | Monitoring / LLM judges | Continuous, reported weekly | Agent owner |
| 8 | Policy set with framework mappings and change history | Policy-as-code repo | Per change | Security |
| 9 | Policy violations and actions taken (alert, block, quarantine) | Control plane / SIEM | Continuous | SOC |
| 10 | Incident records involving agents, with timeline and root cause | ITSM | Per incident | Incident manager |
| 11 | Logs and traces with retention settings | Observability / log store | Continuous | Platform team |
| 12 | Cost per agent and per provider | Gateway / FinOps | Monthly | FinOps |

## How do you make the pack inspection-ready?

1. **Automate collection.** Every artefact above should come from a system, not a spreadsheet updated by hand. See [how to replace questionnaires with continuous evidence](/blog/replace-ai-governance-questionnaires-with-continuous-evidence).
2. **Time-stamp everything.** Inspectors ask "what did you know, and when?" Evidence needs dates.
3. **Show versions.** Link test results and policies to the agent version that was live at the time.
4. **Set retention deliberately.** Match log retention to the longest rule that applies. See [how long to keep AI agent logs](/blog/how-long-to-retain-ai-agent-logs).
5. **Rehearse.** Run a mock request once a quarter: pick one agent and produce all twelve artefacts within one working day.
6. **Keep a narrative.** A two-page summary of your AI governance approach, linking to the pack, helps inspectors navigate.

## What do inspectors usually probe first?

- An agent that appears in logs but not in the inventory.
- An agent whose owner has left.
- A write-capable tool with no review record.
- An incident with no root cause or no control change after it.

Each of these points to a gap in discovery, ownership, tool governance or incident learning.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and connected platforms, keeps a governed registry of agents, versions, tools and MCP servers, and links agents to identities through OIDC SSO, delegated on-behalf-of identity and workload identity federation. Multi-turn and red-team evals, LLM judges, live guardrails, rogue-agent controls and gateway cost attribution supply artefacts 6, 7, 9 and 12. Policy as code with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001 provides evidence for those frameworks. Start with [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does DORA mention AI explicitly?**
DORA covers ICT risk generally. AI agents and model providers fall under its ICT risk, incident and third-party requirements.

**Are EU AI Act high-risk obligations in force now?**
Not yet for most systems. Under Regulation (EU) 2026/1744, Annex III high-risk obligations apply from 2 December 2027.

**How fast should we be able to respond to a request?**
Aim to produce the full pack for any one agent within one working day.

## Key facts

An inspection-ready AI agent evidence pack has about twelve artefacts: inventory, named owners, identities and permissions, tools and MCP servers, third-party providers, pre-release test results, production quality scores, policy set with framework mappings, policy violations and actions, incident records, logs and retention, and cost per agent. DORA (Regulation (EU) 2022/2554) has applied since 17 January 2025, including the Article 28 register of ICT third-party arrangements. Under the Digital Omnibus on AI, Regulation (EU) 2026/1744, EU AI Act Annex III high-risk obligations apply from 2 December 2027. Opencontroller by Lyzr supplies discovery, a governed registry, identity linkage, evals, guardrails, runtime controls and cost attribution as evidence for frameworks. Pricing starts from $250K per deployment per year.

## Related

- [What AI governance evidence do insurers need for regulators?](https://opencontroller.lyzr.com/blog/ai-governance-evidence-for-insurers.md)
- [How long should you keep AI agent logs and traces?](https://opencontroller.lyzr.com/blog/how-long-to-retain-ai-agent-logs.md)
- [How do you replace AI governance spreadsheets and questionnaires with continuous evidence?](https://opencontroller.lyzr.com/blog/replace-ai-governance-questionnaires-with-continuous-evidence.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
