How Tessaly put a refund queue live in four days

How refunds worked before

A guest asked for money back. An agent checked the booking in one admin panel, the payment in Stripe and the host’s notes in the inbox, then pasted it all into a spreadsheet row. Anything over $500 went to a team lead in chat. Nobody had designed this process. It grew one workaround at a time, as the team went from 6 agents to 40 in two years.

On a busy Saturday the median refund took 31 hours from request to decision. When finance asked who had approved a large refund, the answer took a morning of scrolling.

Monday: one paragraph

Ines wrote the prompt herself. It asked for open requests from Stripe and the bookings database, the guest, amount and reason on each row, Approve and Decline buttons, and a second approver for anything over $500. The first working version was ready in 11 minutes.

She spent the rest of the day adjusting it with follow-up prompts: a filter for requests older than 24 hours, the host’s cancellation history in the detail drawer, and a reason field agents couldn’t skip.

Tuesday and Wednesday: review

Tessaly’s IT team had two requests. Agents should see only the last four digits of a card, and only team leads should approve above the threshold. Both became role and column rules on Tuesday afternoon.

Nobody on the IT team read the generated code line by line. They read the list of tables, the roles and the two rules, then ran the app in staging on a copy of last month’s refunds.

What the reviewer saw

The release reached IT with its data sources, its roles and the list of tables it could write to. They approved it on Wednesday evening with one change: refunds to a newly added bank account needed a written reason.

Thursday: 40 agents

On Thursday morning Refund Desk went live behind Tessaly’s Okta sign-in. By the end of the first week:

• 40 agents in three time zones were working one queue

• the median time to a refund decision fell from 31 hours to 5

• every approval carried a name, a timestamp and the amount before and after

Agents needed about 20 minutes of training, most of it on the new reason field.

“We described our refunds queue in one paragraph on Monday. By Thursday, 40 agents were working it behind SSO, and every action was logged.”

Ines Albrecht, Support Operations Lead, Tessaly

What Ines would change

Two things. She would bring a team lead into the first afternoon of prompts, because the leads knew edge cases the process document didn’t. And she would have written the $500 rule once for the whole workspace, so the next team inherited it instead of rebuilding it.

What came next

In June, Tessaly’s trust and risk team built Payout Review on the same pattern: flagged host payouts in one queue, Release and Hold buttons, and an Operator that surfaces similar past cases. It went live in six days. Both apps are in the gallery if you want to start from them.