Building an AI agent register: govern your agents like a workforce

Why agents need a register

Agents are different from the AI tools that came before them. They act on instructions, call other tools, reach systems and complete tasks end to end. Some are built on purpose. Many arrive switched on inside platforms you already pay for.

Once an agent can change a record, submit a form or draft a decision about a customer, student or citizen, it is doing work that used to belong to a person. Yet most organisations can't answer the questions for an agent that they answer routinely for staff:

The register is where those answers live.

What to record for each agent

• Identity and purpose: name, what it is for, and who built or supplied it.
• Manager: the named person who reviews its work, approves changes and is the first call when it goes wrong.
• Reach: the systems, data and tools it can access, and with what permissions.
• Authority: what it may do on its own, what needs approval, and what it may not do. Write this as an allow-list, with everything else denied.
• Decisions about people: whether it makes or shapes decisions that could significantly affect someone. That is your link to the Privacy Act's APP 1.7 disclosure.
• Evidence: test results, monitoring, incidents and overrides.
• Cost: licences, compute and the people time spent supervising it.
• Status and review date: pilot, live, paused or retired, and when it is next reviewed.

Authority is earned through evidence

Treat agents the way you treat new staff. They start with narrow authority, often recommend-only, and earn more as evidence builds up: accuracy against a baseline, few overrides, clean incident history. Each step up is a recorded decision by the agent's manager, not a quiet configuration change.

Before any agent goes live, put it through a production gate: written go/no-go criteria, named evidence for each, and sign-off by both the business owner and risk. It is the same discipline as a change advisory board, applied to something that acts on its own.

Report agents alongside people

Boards see headcount, capability and people risk every quarter. Agents belong in the same view. A short digital workforce report should show how many agents are live, what work they carry, the risk rating of each, incidents and overrides, and what they cost against the benefit claimed.

Start with what you already have

If you are preparing for 10 December, you are already building the hardest part. The automated decision inventory finds the programs that shape decisions about people, names their owners and records how they're used. Add the agents that don't touch personal decisions, and the authority, evidence and cost fields, and you have a register.

A practical first 30 days:

1. Run a census: every agent and AI feature in use, built or bought.
2. Give each one a named manager. Anything without one is paused or reviewed first.
3. Write an allow-list for the five agents with the widest reach.
4. Agree a production gate for anything new.
5. Send the board a one-page report with a red, amber or green rating for each agent.

Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of cost, unclear value or weak risk controls. A register does not guarantee success, but it makes the cost, value and controls visible early enough to act on.

Sources

• OAIC: APP Guidelines, Chapter 1 (APP 1)
• Gartner press release, June 2025: over 40% of agentic AI projects will be cancelled by end of 2027

This article is a planning summary, not legal advice. Confirm your obligations against the OAIC's current guidance and with your legal adviser.