Refunds, cancellations and failed payments are the support tickets where customers most want an answer now and teams most want a person involved, because money moves. Both are reasonable. The way through is to be precise about which of these requests an agent can resolve end to end, which it should prepare for a person, and which architecture keeps the money side safe even when the agent gets something wrong.
§ 01What an agent can resolve end to end
- In-policy refunds and returns. The order exists, it is inside the return window, the item qualifies, the refund is under the value you set. The agent checks each condition against the order record, triggers the refund for that order, and tells the customer when to expect it.
- Exchanges and replacements for standard cases, such as wrong size or a damaged item with a photo, where your policy says what to do.
- Subscription cancellations. Identity confirmed, cancellation applied in the billing system, proration rules followed, confirmation sent. If you run a standard retention offer, the agent can make it once.
- Failed payments. The agent explains what happened in plain words, sends a secure update link from your payment provider, and retries on your schedule. It never takes card details itself.
- Refund status questions. "Where is my refund" is a lookup, and it is one of the highest volume tickets after order tracking.
§ 02What the agent prepares, and a person decides
- Refunds outside policy, such as late returns, used items or goodwill gestures.
- Refunds above your value threshold, however clean the case looks.
- Disputes and chargebacks. Once a bank is involved, a person owns the case.
- Repeat refund requests from the same customer or address in a short period.
- Cancellations from customers on annual contracts or with custom terms.
For these, the agent still does most of the work. It pulls the order, the history and the relevant policy line into one view, writes a suggested reply, and the person approves or changes it. That is usually where the time savings are, because the slow part was finding everything, not deciding.
§ 03The architecture that keeps money safe
The biggest risk is not a rude reply. It is the agent supplying a number. If the model can choose the refund amount, a confused or manipulated conversation can produce a refund nobody intended. The fix is structural, not a better prompt:
- The amount comes from the system of record. The refund tool takes an order ID, and your system calculates the refundable amount. The agent cannot pass in a figure.
- Hard limits live outside the model. Value thresholds, refund counts per customer and per day, and allowed reasons are checked in code before anything executes.
- Every action can be repeated safely. If the agent retries after a timeout, the customer is not refunded twice. We covered this in the double charge write-up.
- Approval for anything outside the rules goes through a queue with a timeout, so nothing sits forever. See approval timeouts.
§ 04Start with one request type
Pick the highest volume, lowest risk category, usually refund status or in-policy returns, and run it with a person approving every action for the first two weeks. Track how often the person changes what the agent prepared. When that number is low, remove approval for routine cases below the threshold and keep it for the rest. Then add the next category.
§ 05Helpdesk AI or a custom agent?
Many helpdesks now include an AI agent, and for FAQ-style answers it is often enough. The difference shows up when the agent has to take actions across systems, such as checking the order in one place, the subscription in another and the payment in a third. We compare the two routes in build or buy an AI support agent, and cover the wider support picture in AI agents for customer support.
§ 06Cost and timing
A single agent covering one request type on one store or billing system and one helpdesk costs 3,000 to 5,000 USD and takes 2 to 3 weeks. Covering refunds, cancellations and failed payments together, across several systems, sits in the 15,000 to 50,000 USD range over 6 to 12 weeks. Running costs are typically 30 to 250 USD per month. The detail is in our cost guide.
──
If refund and cancellation tickets are eating your support team's week, a 30 minute discovery call will show which ones an agent can close safely and which should stay with a person. Engagement models and prices are on our Hire page.