From a home-built AI receptionist to a platform ready for launch

The idea is simple. A pet facility misses calls when the team is busy with the dogs. An AI receptionist answers, takes the booking request, and the staff confirm it.

The founder had built the agent himself. Then the first test calls came back.

It looped on goodbye. It hinted at an availability check it never ran. It said "all set" when the backend call had failed. It did not recognise returning customers, and it kept repeating the same phrases.

An agent like that loses customers on the first call.

We ran scripted test calls and analysed every recording for the transcript and the tone of voice. We audited the agent's configuration.

The result was one prioritised defect list, and before and after recordings for each fix.

A receptionist is only as good as what happens after the call. So the founder handed us the platform too.

QA before launch. About 47 test cases, the riskiest first, across the portal, bookings, payments, SMS and email, the CRM and check-in and check-out. Each one Pass, Fail or Blocked, with evidence. The fix log reached 37 items, about 20 of them on the money path.

Payments that hold up. Stripe Connect Express with live webhooks. A server-side guard: no charge without an active account. Country-aware platform fees for Australia and the US. Monthly cost recovery that never bills twice. An acceptance test with real money.

An event journal that watches the money. Financial and messaging events in one log, expected against actual, checked automatically. A mismatch is flagged in the interface and sent to a chosen channel. Unpaid money gets its own overdue view.