Refund eligibility check

Nobody should let an agent move money. Everybody can let an agent do the reading: pull the order, apply the policy in order, and hand a person a decision with the clause attached. That removes most of the work and none of the accountability.

The job on one screen

Runs when	A refund request arrives, or a support agent opens the refund panel.
Reads	Order date, amount, product type, prior refunds on the account, and the written policy.
Decides	Eligible, not eligible, or needs a human, plus the clause that produced the answer.
Produces	A recommendation with a citation, ready for one click.
Stops when	A clause matches, or the amount is over the limit where a person must decide.
Tools it needs	Orders API (read only), policy document, account history.
Autonomy to start at	Approve. It does the work and stops before the irreversible step.

Optional: load a real model

Run it

The policy, in the order it is applied

First match wins, and the trace names the clause. If you cannot write your policy as a list like this, an agent is not your first problem.

Clause, in order	Outcome	Reason given
Amount over 500	escalate	Above the automatic limit, a named approver signs it off. The agent still gathers everything first.
Three or more prior refunds	escalate	Repeat refunds are an abuse signal, and abuse decisions are not an agent's call.
Within 14 days and usage under 20%	approve	Standard cooling off. Clean case, full refund, no argument.
Within 14 days but heavy usage	escalate	The policy is silent on heavy use inside the window, and silence means a person decides.
Between 15 and 30 days	escalate	Discretionary window. The agent prepares the case and a human chooses.
Nothing matched	reject	Outside the refund window. The reply explains the window and offers the credit alternative.

Where this one goes wrong

Policy written as prose

If your policy cannot be listed as ordered clauses, the agent will invent an ordering. Write the list first, then automate it.

The silent exception

Goodwill refunds that everyone grants but nobody wrote down. Those must be an escalation, not a rule the agent guesses at.

Stale policy copy

The agent applies version 3 while support applies version 4. Point the agent at the same document humans read, not a copy.

How you would know it is working

Measure	Why that one
Agreement rate with the approver	How often a human accepts the recommendation unchanged. This is the number that earns more autonomy.
Time to decision	The real saving. Gathering the facts is most of the work.
Escalation mix	Which clause escalates most. That clause is the one your policy needs rewritten.

Earning more rope

Assist	Show the recommendation and clause in the support tool, human does everything.
Approve	Pre-stage the refund with the amount filled in, a person clicks approve. Money still moves on a human action.
Auto	Small amounts only, with a hard cap, a daily total limit, and every decision logged. Many teams correctly never go here.

Never automate the payment. Automate the reading, the policy lookup, and the paperwork. The click that moves money stays with a person who can be asked why.

Related: Invoice matching · Guardrails · Ticket triage · all agent jobs · Agent Lab home

Free from AI School - no signup, everything runs in your browser.