How to govern AI agents without stopping them

There are two failure modes in AI governance and they arrive in order. First, no governance at all: departments buy copilots, a project team ships an agent, and nobody holds a list. Then, an overcorrection: a committee, a lengthy form, and a six-week approval for a system that categorizes email. The second failure is worse, because teams route around it and you end up back at the first — with less visibility than before.

Start with an inventory, not a policy

You cannot govern what you cannot enumerate. Before writing policy, build the register: every AI system in use, its owner, its purpose, the data it reads, the tools it can call, and the risk tier it sits in. This is unglamorous and it is the single highest-value governance artefact, because every subsequent control depends on it.

Tier autonomy by consequence

The approval burden should track the cost of being wrong, and nothing else. A single global policy applied to every AI action is what makes governance feel like an obstacle.

• Low risk — reversible, internal, no external effect. The agent acts unattended and the action is logged. Example: classifying an inbound message.
• Medium risk — externally visible or hard to reverse. The agent prepares and a named human approves. Example: drafting a reply to a customer.
• High risk — financial, contractual, legal or safety consequence. The agent may recommend and never execute. Example: releasing a payment.

Make the audit trail the deliverable

When something goes wrong, the organization needs to answer four questions quickly: what happened, what data was involved, which system was affected, and whether a human approved it. If your logging cannot answer those in minutes, you do not yet have an auditable AI estate — regardless of how much policy you have written.

Governance done well is what makes the next deployment faster, because the questions have already been answered in general and only the specifics need review. That is the test: if your governance process makes each successive project slower rather than quicker, it is not working.