Most enterprise AI programmes reach the same wall. The pilot works, the demo impresses, and then someone asks the question that stops it: if we let this agent act, who answers for what it does?
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing cost, unclear value and weak risk controls. The projects that survive tend to answer that question before launch, not after the first incident.
Autonomy is a policy, not a prompt
Telling a model to "be careful" is not a control. Reading a customer record, writing a case note, running code, spending money and deleting data are different kinds of action, with different consequences. The rule for each should be explicit: act alone, ask a person first, or never.
Decide what an agent may do on its own before you decide what it can do.
Proof, not logs
Logs tell a story; proof lets someone check it. When every action and every decision is signed and chained to the one before, you can show a regulator, an auditor or your own board exactly what happened, who allowed it, and when.
Repeatable where it matters
For recurring, regulated work, the right answer is the same answer. Scenarios fix the data, steps and rules once, so the model only fills in what changes. That makes results checkable, and it uses far fewer tokens.
A checklist to start
- List the kinds of action your agents will take, and set a rule for each.
- Decide who approves what, and what context they need to decide.
- Make every action and decision verifiable after the fact.
- Fix recurring work as scenarios before you scale it.
- Keep the data, the rules and the record inside your boundary.