An AI agent is useful in operations when it takes routine steps on its own and stops for the decisions that matter. The question isn't whether the model is clever enough. It's what the agent is allowed to do, and what happens at the edge of that.
These five controls make the difference between an agent you can leave running and one you can't.
1. Give it a job, not a goal
"Improve collections" is a goal. "Find overdue invoices and send the standard reminder" is a job. Narrow, repeating tasks with a clear finish are where agents do well, and they're also easy to check. You can widen the job once you've seen how it goes.
2. Set limits the system enforces
Instructions in a prompt are guidance, and a model can misread them. Limits need to sit outside the model: which records it can read, which actions it can take, and how much it can spend or commit. An agent with a £500 limit should be unable to go over it, whatever it decides.
The simplest way to get this right is to hold agents to the same access rules as people, enforced by the server. There's then one set of rules to reason about.
3. Make asking a normal step
At the edge of its limits, an agent should stop and put a clear question to the right person, with the context needed to answer. For example: "Write off this £1,240 invoice? The client is insolvent." The person approves or rejects it, and the agent carries on. Good escalation is what lets you give an agent real work at all.
4. Record every step
Every action an agent takes should say what it did, when, and under which version of your rules. When something looks wrong, you need the actual sequence, not a summary from the agent. A record also lets you review the work in bulk before you trust it with more.
5. Change it the way you'd change anything else
Changing an agent's instructions changes how your business runs. Treat that like any other change: try it on sample data in a draft, review it, publish it deliberately, and keep the previous version so you can step back.
A sensible way to start
- Pick one routine task you could already explain to a new hire.
- Start with the agent drafting, and a person approving every action.
- Review its work for a few weeks, and note what you'd have done differently.
- Let it act alone on the cases you always approved, and keep asking on the rest.
Amazity puts agents on routine steps inside limits you set, under the same rules as your team, with every step recorded. Read about control in Amazity.