A chatbot answers a question. An agent takes an action: it drafts the contract, matches the invoice, files the evidence, updates the record. That difference is why legal, risk and internal audit now sit in the critical path of every AI programme, and why many programmes stall waiting for them.
The fix is not a longer review. It is agreeing the rules once, before the first agent is built, so each new agent is a quick check against a known policy instead of a fresh negotiation.
The problem, and who has it
Business teams want agents live this quarter. Legal wants to know who is accountable when an agent gets it wrong. Risk wants to know how bad "wrong" can be. Audit wants evidence that controls worked, not a promise that they exist. Without shared rules, every agent restarts that conversation.
The steps
1. Keep an inventory of every agent. One register with the owner, the workflow, the systems it touches, the data it reads and the actions it can take. Regulators in several markets, including the Reserve Bank of India's FREE-AI framework, expect firms to know what AI they run.
2. Give each agent a risk tier. Three tiers are enough for most enterprises. Tier 1 reads and drafts only. Tier 2 acts inside limits, such as refunds under a set value. Tier 3 makes decisions that affect customers, employees, money or legal position, and always has a person approve.
3. Write permissions as limits, not intentions. "The agent may approve supplier invoices that match the purchase order within 2% and below the approval threshold" can be tested. "The agent will be used responsibly" cannot.
4. Put approvals where the risk is. Route Tier 3 actions to a named approver with the evidence attached, and set a time limit so approvals do not become the new bottleneck.
5. Log every action, in a form audit can use. Who or what acted, on which record, with what input, under which policy version, and who approved. Keep it as long as the underlying record.
6. Agree the off switch. Name who can pause an agent, how fast, and what happens to work in flight. Test it before go-live, not during an incident.
Who signs off
General Counsel signs the policy and the accountability model. The Chief Risk or Compliance Officer signs the tiering and limits. Internal Audit agrees the evidence it will test. The CIO or CISO signs access and data residency. Get all four in one room for the first agent, and the next ten move faster.
What good looks like
- Every agent in production appears in the register with an owner and a tier.
- New Tier 1 and Tier 2 agents are approved in days, not months.
- Audit can sample any agent action and trace it to an approval and a policy version.
- The off switch has been tested and timed.
Put this playbook to work
Book a call and we will apply it to your organisation with your teams.
