Brief an AI agent the way you would brief a new hire

Imagine you hire someone on Monday and give them one sentence: “Handle the customer refunds.” By Wednesday you would expect questions, guesses, and a few refunds you would not have approved.

You would not blame them. You would blame the brief.

An AI agent is in the same position, with one difference. It will not walk over to your desk and ask. It fills the gap with a confident guess. That makes the brief the most important thing you write.

The gap is where the mistakes live

Every unclear instruction is a decision you left to someone else. A person makes it with years of context about how your business works. An agent makes it with none.

Most disappointing agent results trace back to a missing line, not a weak tool. Fix the brief first.

Five lines that make a brief

Write these, in this order, in plain sentences. Keep each short.

1. The job. One sentence on the outcome. “Prepare a reply to each refund request for a person to send.” Note that it says prepare, not send.
2. The inputs. Where it looks, and only there. “The order record, the refund policy page, and the customer’s message. Nothing else.”
3. What done looks like. The finished piece, described so you could check it. “A reply of under 120 words that names the order, states the decision, quotes the policy line it relied on, and offers one next step.”
4. When to stop and ask. The cases it must not decide. “If the order is more than 60 days old, the amount is above the agreed limit, or the message mentions a legal claim, stop and hand it to a named person with a one-line reason.” Replace these thresholds with your own.
5. What it must never do. The short list. “Never promise a refund the policy does not allow. Never edit an order record. Never contact the customer directly.”

Line four is the one most briefs miss, and the one that matters most. An agent with no permission to stop will never stop.

Add two real examples

A new hire learns faster from two worked cases than from two pages of rules. Give the agent the same.

Pick one ordinary case and one awkward one from your own history. Show the input, the reply your best person would have sent, and one sentence on why. The awkward case is the valuable one, since it shows where judgment enters.

Two is enough. More examples tend to make the agent copy their surface, wording included, rather than their reasoning.

Test the brief on old cases before it touches new ones

Take ten past cases where you know the right answer. Include a few that should have been escalated. Run the brief on all ten and mark each result in one of three ways: right, wrong but safe, or wrong and harmful.

Read the wrong ones and ask what line was missing. Add that line. Run the ten again.

When the harmful column is empty on two runs in a row, let it try live work, still with a person approving. Widen its freedom only as its record earns it.

Keep the brief where the agent’s owner can find it

A brief lives. Policies change, edge cases turn up, and a limit that made sense in spring is wrong by autumn.

Give the brief an owner with a name. Keep it in one place. Every time a person overrules the agent, ask whether the brief should have prevented it, and if so, change the line that day. Log the change with the date, so you can tell later why behavior shifted.

A good brief is short enough to read in a minute and specific enough to check. If a new hire could follow it, an agent probably can.

If you want agents that prepare the work while your people decide, our AI Agents & Automation service starts from exactly this kind of brief. And to decide where the review points belong, read where to put the human in an AI workflow.