AI agent inventory: how to build one in 5 steps
You can't set limits for an agent you don't know exists. The inventory is the first step of any agent governance, and the one most often skipped.
Why start with the inventory
An agent is like an employee with access to your systems. Nobody would hand out an access card without knowing who the person is or what they do. With agents it happens all the time: one team switches one on in its email tool, another connects an automation to the customer database, and nobody writes it down.
Supervisors are heading the same way. Spain's AEPD considers dedicated governance of agents the most important measure, and France's CNIL wants organisations to be able to reconstruct what each agent did. Neither is possible without knowing which agents you have.
What each record should hold
| Field | What to write down |
|---|---|
| Name and purpose | What it does, in one sentence. |
| Owner | A named person, not a department. |
| Tool or provider | The product or model it uses. |
| Data it accesses | Which data, and whether it includes personal or sensitive data. |
| Systems and permissions | Which systems it can read and which it can write to. |
| What it does on its own | For example, replying to customers or creating tickets. |
| What needs approval | Payments, deletions, anything sent outside the company. |
| Other agents it works with | Whether it is part of a chain. |
| Risk level | Whether the chain ends in an AI Act Annex III use. |
| Logging and stop | Where its actions are recorded and how it is stopped. |
| Next review | A date. |
The 5 steps
1. Discover
Ask each area which AI agents or automations it uses. Also check the AI features switched on in tools you already pay for (email, CRM, customer service) and any access keys created for AI services. It is common to find more than expected.
2. Record
Open one record per agent with the fields above. If something is missing, mark it as pending: an incomplete record is better than none.
3. Classify
Look at the whole chain, not each agent on its own. Under the Commission's draft guidelines, when several agents together serve a high-risk use, they are assessed as a single system.
4. Assign
Each agent needs an owner who can answer three questions: what it may do, what it may not do and who stops it if it fails.
5. Review
Set a regular review (for example, every quarter) and another one whenever an agent gains access to new data or systems.
Common mistakes
- Recording only the official agents and forgetting the ones each team switched on by itself.
- Making a department the owner. If everyone is accountable, nobody is.
- Building the inventory once and forgetting it. Agents change permissions faster than people do.
Frequently asked questions
Is a spreadsheet enough?
Yes, to start with. What matters is that it stays up to date and that every agent has an owner.
Does a chatbot count as an agent?
If it only answers, it is an assistant. If it also does things (creates an order, sends an email, changes a record), treat it as an agent.
For information only, not legal advice.
What's your Data Governance maturity?
Free assessment with your priority gaps and the templates on the Data Governance path.