Refund Eligibility Check Agent · Use case
How can we use an AI agent to check a refund request against our refund policy on our own infrastructure?
A refund eligibility check agent takes one refund request, reads the order record and your refund policy, and sets out which rule of the policy applies, which conditions the record meets, which it does not, and what is missing. An adviser or a supervisor approves or refuses the refund in your own system. The agent issues no refund, no credit and no compensation, recommends neither approval nor refusal, and tells the customer nothing.
The policy has twelve conditions.
The order record has the facts.
- ✗ 01An adviser with a refund request opens the order, the delivery record and the payment, then reads the policy for the rule that fits, with the customer waiting.
- ✗ 02A refund is money leaving the company and a refusal is a decision about a customer, so neither is a step to hand to software.
Where it stops for a person
Always. An adviser or a supervisor approves or refuses.Always. It checks, it does not decide: it does not say "approve" or "refuse", and where the policy leaves room for judgement it says so and quotes the clause. An exception outside the policy is a supervisor's decision. A refusal is told to the customer by a person, who can explain it. It scores no customer and keeps no tally of who asks for refunds.
It says which rule applies.
An adviser or a supervisor decides.
Step · 01
An adviser opens the request and names the order.
Step · 02
The agent reads the order, delivery and payment records through your own APIs, read only.
Step · 03
It finds the rule of the refund policy that fits the request and quotes it.
Step · 04
It sets out each condition of that rule beside what the record shows: met, not met, or not in the record, and lists what the adviser still needs from the customer.
Step · 05
The adviser, or a supervisor above the limit you set, approves or refuses the refund in your own system and tells the customer.
Reads from · connectors
Your own APIs
The order, delivery and payment records, read through your own APIs with the operations that change something left out; and the refund and returns policies, indexed as a knowledge source. It runs with the adviser's permissions and reads the one order named.
Acts through · MCP servers
Nothing. It has no tool that raises a refund, a credit or a voucher, changes an order or writes to a ticket.
What you set up
- Index the refund and returns policies as a knowledge source, and remove the versions no longer in force.
- Import the description of the API that holds orders and payments, and keep only the operations that read.
- Build the agent in Katonic Studio and test it on past requests where you know which rule applied.
- Publish it to one team's advisers first, with the approval limits you already use.
Where it runs
The agent, its knowledge store and the model run on infrastructure you own or control. You choose the model and serve it yourself, so questions, documents and answers are not sent to a public AI service. Order and payment records stay inside the company, and every refund is approved by a person in your own system.
How you get it today
This page is labelled Use case: the agent is built in Katonic Studio. You describe the job, connect what is named above, add test questions and release the agent once it passes them. When the Catalog holds a tested template for this job, this page will carry the Template label and show how to take it.
Does it issue the refund?
No. It has no tool that raises a refund, a credit or a voucher. An adviser or a supervisor approves or refuses in your own system.
Does it recommend approval or refusal?
No. It sets out each condition of the rule beside what the order record shows. The person who decides reads both.
What if the policy does not cover the request?
It says so, quotes the nearest clause and lists what is missing. An exception is a supervisor's decision.