How a high-volume e-commerce team handled 15,000 peak-season calls without expanding the team
Refunds and replacements at peak ran 15,000 calls a month. The human team was burning two to five minutes per call on intake before they ever got to the actual problem. We built an agent that ran the full intake itself, resolved the simple cases end to end, and handed humans pre-populated context on the rest.
- Peak monthly calls handled
- ~15,000
- Refunds/replacements resolved end-to-end
- ~3,000/mo
- Intake time reclaimed
- 2 to 5 min per call
- Callback context
- 100% structured
The shape of the problem
Every single refund or replacement call started the same way. Refund or replacement? What's the order number? Delivery address? Where did you buy it (Amazon, in most cases)?
Two to five minutes per call. Every call. Before the human ever got to the actual problem.
At 500 calls a day in peak season, the math stops being marginal.
What we built
We fronted the customer-care line with a voice agent that runs the full intake conversationally. Before the agent asks the buyer a single question, it has the order state in front of it from the order management system. So no "let me look that up", no repeating the order number three times.
Then the path forks.
- Simple cases get resolved on the call. Clear policy, clear order, clear request. The agent actions the refund or replacement and the buyer hangs up done.
- Complex cases get a callback. The agent collects the full intake, schedules the callback, and writes the context to the case record. When the human picks up the conversation, they pick it up halfway through. Not from zero.
The simple-case automation only got greenlit after we ran a sample of actual peak-season recordings against the rules. Roughly 80% of the calls fit one of four intake shapes. The rest stayed human.
What it did to the operation
- ~15,000 calls a month handled at peak
- ~3,000 refunds and replacements resolved without a human touching them
- Two to five minutes of intake reclaimed on every single call
- 100% of human callbacks landed with structured context
The first two numbers are the headline. The third is the one that actually changed the operation. At 15,000 calls, that's tens of thousands of minutes a month the team stopped spending on intake.
What we'd change
Train two people inside the company to make small agent changes themselves, on day one. We did that here too late. Now it's standard scope on every customer-care build we ship. If your team can't tweak the agent without us, we've built the wrong thing.
When this build fits
Any team that's asking the same intake questions on every inbound call. The savings aren't in any single call. They're in the sum across the month.