What is an AI inventory?

An AI inventory is a register of every AI system an organisation builds, buys or uses, recording each system's owner, purpose, model, data, risk classification and status. It is the starting point of AI governance: you cannot control systems you do not know about. For agents, it must also record what each one is connected to.

Why does an AI inventory matter once agents act?

AI arrives in organisations through many doors. A team builds an agent on a model provider's platform; a software vendor adds an assistant to a tool already in use; a staff member connects an agent to their inbox. Without an inventory, the business learns about them when something goes wrong.

Agents raise the stakes because they act. An unregistered chatbot is a risk to accuracy; an unregistered agent with a connection to the payments system is a risk to money. The inventory needs to capture connections and capabilities, not only names. The wider frame is in agent governance is a bound, not a model.

What should an AI inventory record?

For each system:

• Name and identifier, unique across the organisation.
• Owner: the named person who answers for it.
• Purpose: the job it does, in a sentence.
• Type: model, assistant, or agent that takes actions.
• Model and provider: what it runs on, described generically.
• Data: what it reads, including any personal or special-category data.
• Connections: systems it can write to or act in.
• Risk classification: under internal policy and any applicable law.
• Status: proposed, in testing, live, suspended, retired.
• Last review: when it was last tested or certified.

A ready layout is the AI agent inventory register template. Keeping it current is the harder part, covered in how to maintain an inventory of AI agents and how do you keep an inventory of AI agents?

What does that look like in an example?

A mutual insurer runs its first AI inventory exercise. It expects to find a handful of systems. It finds many more: assistants built into its document tools, two agents built by the claims team on a model provider's platform, and a pricing analyst's script that calls a model through an API. One of the claims agents has write access to the claims system. The inventory does not decide whether that is acceptable, but it puts the question in front of the people who should answer it.

How is an AI inventory different from a model inventory?

Banks have kept model inventories for years under model risk management rules, such as SR 11-7 in the US and PRA SS1/23 in the UK. A model inventory records statistical and quantitative models, their validation status and their use. An AI inventory is broader: it includes systems that are not models in that sense, such as agents and assistants, and it must capture what each system can do, not only what it calculates. How the two connect is discussed in SR 11-7 model inventory for AI agents.

What can an inventory not tell you?

An inventory tells you what exists and who owns it. It does not tell you what each agent may do at the moment it acts. Knowing that an agent has a connection to the payments system is different from knowing it may only draft payments below a limit, that it was tested on your own cases last month, and that its certificate lapses in a few weeks. That second layer is authority, not inventory, as argued in an agent inventory tells you what exists, not what it may do.