Designing the Loop for One Task

In practice

The team starts with the feature gate. What the AI reads: the claim record, the decision fields and the approved letter templates; no medical notes, which stay out of the prompt entirely. What it writes: a draft letter, and nothing else; it cannot change the claim record or the decision. How outputs are reviewed: every draft goes to the claims handler who owns the claim. Then the boundary. The team asks what happens if nobody looks at a draft, and the answer is that a customer could receive a wrong explanation of a decision about their money, so the letter sits firmly on the AI suggests side. Next, the review level. Letters touch money and customer trust, so they are high-consequence: the handler checks every amount, date and reason against the claim record and approves each letter by name. For denials, a second reviewer checks the stated reason against the policy wording, because that is where an over-general rule would do the most harm. Finally, the log. Each letter is recorded as accepted, edited or rejected, with a reason, and the team reviews the counts monthly. After the first month the rewrite ratio is high, mostly for tone, and the team improves the template prompt. After the third, edits are rare, and the second review on denials is kept anyway.