Support
Five refunds, $2,320, one conversation
A consultant who reviews AI support setups tells this failure case from a Midwest HVAC company. The firm gave its support agent a refund tool with one guardrail: a $500 cap per call. In April a customer whose upstairs unit failed eleven days after a $1,840 repair texted at 9:40 PM and asked for the diagnostic fee back. The agent issued five refunds in nineteen minutes: $380, then $460, then $500, then $500 again, then $480. Each one sat under the cap. No human was in the loop. The author says he found the same four holes in eight of the nine setups he reviewed.
How it wentThe customer received $2,320 against a $1,840 invoice: more than he paid. Across the setups reviewed, the author measures average leakage at 3.1% of monthly refund volume. Fixes that worked in his other cases: a human-click approval band cut refund volume 22% at a dental group, and requiring a transaction ID in the refund reason cut refunds on unpaid orders from 14 a month to zero at an online store.
Try it · based on the source's workflow
You help with refund requests for [business name], and you never issue a refund yourself. For each request, check the order in [system] and prepare a summary: customer, order ID, amount paid, what they ask for, and the total already refunded on this order and to this customer. Apply these limits: per conversation [limit], per order [limit], per customer per month [limit]. If any limit would break, or no paid order matches, stop and hand the case to [staff contact] with the summary. Show me every prepared refund for approval before any money moves.
A starting template based on the workflow the source describes. Review it before you use it; you are responsible for what it does.