AI-agent governance checklist for enterprise architects

AI agents are applications with a model inside. They have owners, data access, the capabilities they serve and a lifecycle, and they run on several platforms at once. Enterprise architecture already has the discipline to govern that: an inventory, links to the rest of the estate, and regular sweeps for what doesn't line up. These 24 checks apply it to agents.

1. Inventory

1. Every platform where agents run is listed, including agents built into SaaS products you already license.
2. Each platform's agent inventory can be exported or read through an MCP server or API, not kept by hand.
3. Each agent has an id, a name, a platform, an owner and a status.
4. Each agent's record says which data it reads and which systems it can change.
5. The inventory has a date, and someone owns refreshing it.

2. Registration in the architecture

6. Each agent links to an application record in the EA repository, by id where possible.
7. An agent whose record names an application the repository doesn't hold counts as unregistered, not as registered under a typo.
8. Each agent maps to the business capability it serves.
9. The shadow-AI sweep has run: every agent that touches customer or personal data and has no application record is listed with its owner.
10. Matching agents to applications by name is checked by a person, or by a method that shows its evidence and confidence for each pair.

3. Duplicates

11. Agents doing the same job for the same users on different platforms are listed as pairs.
12. The same product deployed under qualified names (EU, US, Pilot, V2) is tracked as one product, not as duplicates.
13. Each duplicate pair has a decision: keep one, merge, or keep both with a reason.

4. Risk tier

14. Each agent is classified from what it does, not only from the class its owner recorded.
15. Agents in a high-risk use case under the EU AI Act's Annex III are flagged. The timetable for high-risk obligations changed during 2026, so check the current dates before you plan against them.
16. High-risk agents with no named person reviewing their decisions are listed first.
17. Any use that could be a prohibited practice under Article 5 is escalated at once.
18. The tier and the reason for it are recorded on the agent, with a date and the person who set it.

5. Oversight and change

19. Each agent that can change data in a system of record asks a person first, or has a recorded exception.
20. Agents' actions are logged where an auditor can read them.
21. A change to an agent's model, prompt or tools triggers a review of its tier.
22. Retired agents are switched off on the platform, not only marked retired in the inventory.
23. The register is reviewed on a schedule, and the sweeps (9, 11 and 16) rerun each time.
24. Each sweep's result can be traced to the records it read.

Run the three sweeps

Checks 9, 11 and 16 are the ones that find things. Each is a join across systems: every agent platform's inventory against the EA repository. By hand, across three platforms and a few hundred agents, each one is a spreadsheet exercise that is out of date when it's finished. Run as a standing query across the connected inventories and the repository, each becomes a question you can ask every month. Whatever does the join, two rules keep the result usable: every agent it lists should trace to the record it came from, and every cross-system match should show its evidence.

The three sweeps are three of the questions in EA-Bench, our study of how well AI systems answer the questions EA teams get: AI inventory reconciliation, AI agent rationalisation and EU AI Act classification. The question kit says what a right answer to each contains, and the traps that should fail it.

Start with the sweeps. A first pass usually finds agents nobody registered, agents doing the same job on two platforms, and agents whose risk tier was never set. Each finding is a conversation with an owner rather than a policy debate.