When a new engineer joins a company, nobody hands them the production database password on day one and says, “Surprise me.” They get a role, a manager, a narrow set of permissions, and a probation period. Yet that is roughly how many teams deploy AI agents today.
After two decades of watching technology programmes succeed and stall, I have noticed that the failures rarely come from the technology being too weak. They come from the boundaries being too vague. An agent that can read email, edit files, call APIs and spend money is not a productivity tool. It is an unsupervised employee with no contract.
The principle is easy to nod along to and surprisingly hard to apply, so let me make it concrete with a use case I would happily propose to any e-commerce platform: the returns and refunds agent.
The job description
Online retail drowns in “where is my order?” and “I want my money back” messages. They are repetitive, emotionally charged, and time-sensitive, which makes them a perfect first job for an agent. Here is how I would write the role.
Responsibilities. Read incoming customer messages, look up the order and shipment status, classify the request (late delivery, damaged item, wrong size, change of mind), and draft a reply in the customer’s language. For clear-cut cases inside policy, such as a parcel stuck in transit beyond the promised window, it may prepare a refund or replacement for approval.
Out of scope. Changing prices, editing product listings, touching payment credentials, contacting customers who have not written in, and anything involving suspected fraud, legal threats or safety complaints. Those go straight to a human.
Access. Read-only on orders, shipments and customer history. Write access only to a “proposed actions” queue, never directly to the payments system. Personal data is masked unless the case genuinely needs it.
The manager and the approval ladder
Every agent needs a named human owner. In this case it is the customer-care operations lead, who reviews a daily sample, owns the metrics, and answers for mistakes. “The AI did it” is not an incident report anyone wants to read at two in the morning.
Then build a ladder of trust. In week one, the agent only drafts, and humans send everything. Once draft acceptance is consistently high, low-value refunds under a small threshold can be auto-approved, while larger amounts, repeat claimants and unusual patterns still wait for a person. Money moving is always the rung you climb last.
Make the work visible
Log every action with the reasoning behind it: which order it looked up, which policy clause it applied, what it proposed. Keep a one-click undo for anything reversible, and a hard human gate in front of anything that is not. Measure the boring things that matter: resolution time, customer satisfaction, refund accuracy, and the rate at which humans overrule the agent. That last number is your honest performance review.
Why the constraints are the feature
The pleasant surprise is that these boundaries do not slow the agent down. They are what make it deployable. A cautious finance director will say yes to an agent that can only propose refunds under a cap, with a full audit trail. She will say no, quite rightly, to one with a blank cheque.
In my experience, the teams that move fastest with AI are not the ones with the most freedom. They are the ones with the clearest org chart, including the robots.