The problem
An immigration firm lives on the phone. Clients call to ask where their case is, new people call to find out whether the firm can help, and everyone wants an appointment. Most of those calls follow a handful of patterns, and every one of them used to take a person away from casework.
The goal was not a chatbot that sounds clever. It was a phone line that resolves the routine calls on its own, hands the rest to a human with the context already captured, and never makes a client feel processed.
What the agent does
I built the voice intake agent on VAPI. It handles three jobs:
- Case status lookups, answered from the firm’s Clio CRM.
- Intake triage for new enquiries, so a lawyer starts the first conversation already knowing what the case is about.
- Appointment scheduling.
It is wired into Clio and into the firm’s Asterisk PBX, so it sits on the real phone line rather than beside it.
The decision that mattered: split the model routing
On a phone call, latency is the product. A pause that would be fine in a chat window feels broken on a call. So I split the routing: gpt-4o-mini on the hot path, at a p95 of around 600ms, and the full model only on the hard 15% of calls.
I did not ship that split on a hunch. I validated it on 200 call transcripts first, checking that the small model held up on the routine calls and that the hard ones were being sent up. The split took about 80% off the token bill.
Evals before model swaps
I designed the evaluation framework for call handling quality before we needed it. It earned its keep during a model swap: it caught the regressions before clients did.
With a voice agent, a regression is not a bad line in a log. It is a client on the phone getting a wrong answer about their visa. The eval set is the only thing standing between a provider update and that call.
Result
60% of the firm’s phone work now runs without a human. Case turnaround dropped by roughly 30%, because the time that used to go to routine calls went back into casework.
On a phone line, the cheapest model that is fast enough is the right default. The expensive one is for the calls that prove they need it.