What counts as an AI agent, and what belongs in the inventory?

01A working definition

Three properties together make something an agent. It is given a goal rather than a single instruction. It decides its own sequence of steps toward that goal, using a model to choose what to do next. And it can act: call tools, use other systems, change things outside the conversation.

Official definitions say much the same. The OECD, in a paper on agentic AI in organisations published in September 2026, describes AI agents as systems that perceive and act on their environment with a degree of autonomy, using tools as needed to achieve specific goals. Singapore's IMDA notes there is no consensus definition, but that agents usually plan, decide and act over multiple steps toward a goal a user sets.

02Agent, assistant, automation

• An assistant answers a person and stops. The person decides what to do with the answer. A chat tool used for drafting is an assistant.
• Automation follows a fixed path someone designed. A workflow that sends each new invoice to a model for a summary and files the result is automation, even though it uses AI at one step.
• An agent chooses its path. Given 'resolve this support ticket', it decides to look up the customer, read the order history, issue a refund and write the reply, and does each of those through tools.
• The line moves. An assistant given tools and permission to use them becomes an agent. A workflow with a model choosing which branch to take is most of the way there.

03Why the definition matters

The word is used loosely. Products are relabelled as agents, and some things that act like agents are never called one. An inventory built on the label misses the second group, which is the one that carries the risk.

Ownership is thinner still. In Gravitee's survey of 750 technology leaders, published in June 2026, 7.2 percent of organisations had a named individual formally accountable for AI agent behaviour.

The numbers are moving fast. In KPMG's Q2 2026 AI Pulse Survey, 53 percent of large US enterprises had AI agents deployed and 18 percent were orchestrating several agents across workflows, double the previous quarter. Deloitte's State of AI in the Enterprise 2026 found 21 percent have a mature governance model for agents.

04What belongs in the inventory

Use a functional test rather than the vendor's label. Include anything that meets any of these:

1. It can take an action in a system (write, send, change, pay, delete) without a person approving that specific action.
2. It runs on a schedule or a trigger, without a person starting each run.
3. It holds credentials or an API key of its own, or acts under a service account.
4. It can call other agents, or be called by them.
5. It spends money on the organisation's behalf, through model calls or anything else.

Assistants that only answer a person still belong on the register of AI in use, because of the data they see. They need a lighter entry than an agent does.

05Edge cases

• AI features inside licensed software. Include them if the feature can act, such as sending email or updating records, and record the vendor as well as your internal owner.
• Coding agents. Include them: they change code and run commands, often with the developer's credentials.
• Agents in testing. Include them once they hold real credentials or touch real data. A test agent with production access is a production agent.
• Personal agents. An agent a member of staff runs on their own account against company systems is in scope, and is the hardest to find.