Prima Polis
Every advice call at a Dutch insurance intermediary now writes itself into the client file. Four vendor systems, made to cooperate under financial regulation.
The situation
Prima Polis is an insurance intermediary in Teteringen. Their advisers spend the day on the phone, and every one of those calls is supposed to end as a written note on the client's file: what was discussed, what was advised, what happens next.
That obligation is not administrative housekeeping. Dutch intermediaries work under Zorgplicht, the statutory duty of care, and under the Wft, the Financial Supervision Act. If the advice was given and the file does not show it, the firm carries the risk.
How I framed it
The firm came to me describing a productivity problem. Advisers were losing time to admin and wanted it back.
Sitting with it, the time was the smaller half. An adviser who writes a note four hours late writes a shorter, vaguer note, and vagueness is precisely what the duty of care cannot survive. The exposure was not that notes took too long. It was that the notes which did exist were reconstructions.
Asked for
Advisers are losing hours to writing up calls. Give them that time back.
Actually needed
The firm's evidence of advice is being reconstructed from memory. Make the record a by-product of the conversation instead.
That reframe changed what success looked like. A faster note-taking tool would have been judged on minutes saved. What the firm actually needed was judged on whether the file could be relied on at audit, which pointed at a different system entirely.
What I built
No single vendor does this. The work was making four independent systems behave as one chain: the CRM the firm already ran, a desktop bridge for click-to-call, the telephone platform that places and records the call, and the AI service that transcribes and analyses it.
None of those four products were designed with the others in mind, and the integration points a vendor documents are not always the ones that survive production.
- 01
Establish what the chain actually depends on
Before anything was configured, I worked out where the sequence could break silently. Every report depends on recording being enabled at the telephony layer, which means one misconfigured handset produces no report at all and no error to tell anyone.
- 02
Choose the route into the client file
More than one path back into the CRM looked viable. Only one of them actually carries this kind of call. Establishing which, before go-live rather than after, was the decision the engagement turned on.
- 03
Specify the analysis against the regulation
The system pulls out commitments made and advice given as discrete fields, because those are what the duty of care turns on. A generic call summary would have been easier to build and worth considerably less at audit.
- 04
Design around the duplicate record
The CRM writes its own note the moment a call ends, while the analysis arrives minutes later. Left alone, one conversation leaves two entries. That was resolved deliberately rather than discovered by a confused adviser.
The second stage is the one worth dwelling on, because the failure mode it avoids is the expensive kind.
Why that matters
A route that fails in testing costs an afternoon. A route that passes testing and then stops delivering under load costs a fortnight of silence, and the discovery usually arrives from the regulator's direction rather than your own.
What shipped
- Automated call capture, transcription and analysis
- Structured advice and commitments per conversation
- Reports written back to the correct client file
- Action points logged as tasks
- A resolved single-record path into the CRM
- Working instructions for the firm's own staff
Where it stands
Live and running in daily use. Prima Polis report that it saves hours a day of admin, and the compliance position is stronger than it was, because the record of advice is now created as a by-product of giving it rather than as a task to remember afterwards.
A second phase is in build: my own transcription and reporting pipeline, running on Python and FastAPI with a Next.js review dashboard, which adds an approval step before anything reaches the CRM and grounds product identification against the firm's real catalogue. That one is at pilot stage and I would not yet call it finished.
Next project
Sarfati Associates