A Deliberate Rep: Answering an Angry Customer Without Inventing Policy

Theory is cheap. Here's a deliberate rep, start to finish, on a task that runs all day in every support queue: replying to a furious customer about an edge case the policy doesn't quite cover. It takes about ten minutes — and the place the capability gets built is not where you'd expect.

The task, and the bar

A customer bought 22 days ago. Your return window is 14 days. They're angry, they're partly right — the product page was ambiguous — and they've cc'd two people. You need a reply that holds the line without making an enemy. Before you touch AI, set the bar — the three criteria a tough reviewer would use:

1. Every policy claim is traceable to policy that actually exists.
2. Acknowledges the specific grievance, not "your frustration."
3. Ends with one concrete next step and a date.

Writing the rubric first is the move most people skip. On this task it's also the move that keeps you out of trouble.

The vending-machine version

Prompt: "Write a reply to this angry customer asking for a refund." You get something warm, fluent, and dangerous:

I'm so sorry for your frustration! We completely understand. Good news — our 30-day satisfaction guarantee means you're still covered, so I've gone ahead and started your full refund. Thanks for being a valued customer!

Read it against the rubric. There is no 30-day satisfaction guarantee — the model invented one, because that's the sentence that makes an angry email go away (fail). "Your frustration" is boilerplate that never names the ambiguous product page (fail). And it commits to a refund nobody authorized (fail). 0 / 3.

This is the failure mode that makes support different from a status update. The status update was fluent and empty. This one is fluent and expensive — it just promised, in writing, a policy your company doesn't have.

The rep

Now do it deliberately. Paste in the actual policy text, the customer's order history, and the one thing the model can't infer: the exception you're actually allowed to authorize — a one-time credit, no manager approval needed under $80. Iterate twice.

You land much closer. The invented guarantee is gone, the offer is real. Read it against the rubric and two gaps remain: it says "as per our policy" without ever saying which policy, and the next step dissolves into "we'll be in touch shortly."

Score, then close the gap yourself

This is where the capability lives. The model produced a good-enough draft; your judgment finishes it:

• "As per our policy" becomes "our returns window is 14 days from delivery — you're at 22."
• "We'll be in touch shortly" becomes "I've issued the credit today; it lands in 3–5 business days."
• And the sentence the model would never write, because it requires knowing your own company is imperfect: naming the ambiguous product page and saying you've flagged it.

The final artifact:

Re: refund request — order #4417. You're right that the product page was unclear about sizing, and I've flagged it to the team this morning. Our returns window is 14 days from delivery and you're at 22, so I can't process a standard refund — but I've issued a one-time $60 credit to your account today. You'll see it in 3–5 business days. If the sizing is still wrong when you reorder, reply here and I'll handle it personally.

Holds the line. Names the real problem. One next step, with a date.

What this rep proves

Notice what actually built the capability. It wasn't the AI draft — the model wrote the reckless version and the good one. It was the rubric (it caught an invented policy before it reached a customer) and your judgment (knowing what the policy says, what exception you can authorize, and that the product page was genuinely ambiguous).

Anyone can generate warmth. Warmth that doesn't quietly commit your company to a guarantee it never offered is a different skill. The fluency was free. The sentence you didn't send is what you got paid for.

This is one rep. The UofAi method is built to give you dozens of them, each one evidence. Start free.