Context
Customer calls at a business in regulated finance, where what a customer commits to on a call needs to become a record someone can act on.
The problem
Commitments made on a call live in the rep’s head until a note gets written, so things slip. And an AI summarizing the call will happily invent an amount or date that was never said.
Constraints
- Personal data could not reach the AI model.
- Every extracted value had to match what was actually said.
- Live calls and recordings needed to follow the same path.
- Nothing counted until a person confirmed it.
Approach
The call is processed turn by turn, with personal data scrubbed before anything reaches the AI. It extracts clean records (a commitment with an amount and date, a dispute, a callback, an escalation), each with a confidence score, and every value must match the transcript exactly. They appear as one-click cards a person confirms. The transcription source is pluggable, so a live voice stream and a recording follow the same path.
Key engineering decisions
Redact before the model
Personal data is removed from each turn before it reaches the AI, rather than trusting the model to ignore it.
Typed, verbatim-grounded records
Output is a fixed set of record types, every amount and date must appear exactly in the transcript, and each record carries a confidence score.
One path for live and recorded calls
The transcription source is pluggable, so a live voice stream and a recording produce the same turns and share one extraction path.
A person confirms every record
Each record is a one-click card, and it only counts once a person confirms it.
Result
Commitments become real, actionable records instead of half-remembered notes, with a person confirming each one before it counts.