What is Trillet AI?
It is a voice AI platform built for phone calls where getting it wrong is not an option: a dental clinic rescheduling a patient, a legal aid service taking on a new client, a government team following up on a housing application, a collections agency setting up a payment plan. The AI on the line has to verify who it is talking to, pull up current information, complete the task and log exactly what was done.
The problem
More than 3,900 businesses across healthcare, legal, finance, government and utilities now run on the platform. The early version, though, had two serious problems.
- It broke its own rules. It was never supposed to deviate from them, yet roughly one call in twenty went wrong, for example offering a time it was not allowed to book. The model wasn’t strict enough about what it couldn’t do.
- It was too slow to feel real. There was a noticeable lag between the caller finishing a sentence and the AI responding, and sometimes it started speaking before the caller had finished, which eroded trust on top of the errors.
The solution
The move to Google Cloud became a global case study, and three things changed as a result.
- The AI holds far more instructions at once. Every client has its own rulebook: opening hours, booking policies, what the AI is allowed to say and do. The new setup lets the AI read hundreds of pages of these rules for each business and stick to them, including the “never do this” guardrails the old system kept ignoring.
- Responses dropped under two seconds. Conversations started to feel like talking to a person rather than an automated agent. If a caller interrupts or changes their mind mid-sentence, the AI adjusts and picks the booking up where it left off.
- The system scales on its own. Infrastructure expands automatically at rush hour and shrinks back afterwards, so there is no idle capacity to pay for.
The results
- <1%
- error rate, down from 5%. The phantom-booking problem is gone.
- ~80%
- lower infrastructure costs
- 85%
- of complex calls resolved with no human involvement
- <2s
- response times, even during 3x traffic spikes
- +10%
- conversions for legal clients, because callers who felt heard were more likely to proceed
The takeaway
The less visible results mattered most. Instead of auditing call logs by hand, the engineering team went back to building: agents that can look at a document a caller shares, such as a medical form, and discuss it in real time, and memory that lets the AI recognise a returning caller and pick up their history.
Voice AI fails when it guesses. Trillet AI’s turnaround came from removing the guesswork: a model that respects hard limits, infrastructure that keeps pace with a real conversation, and a system that scales without anyone touching it. That is the bar for AI in high-stakes work, and it is the bar we hold ourselves to at Vector AI.
