Case study · Pensions

A pension-letter assistant: one-day prototype, on-premises production in twelve sprints

A customer-facing assistant that explains pension letters in plain language, defines difficult terms, lists the actions a member must take, and offers translation and audio for non-native speakers; prototyped in a day, then integrated and validated over twelve sprints into an on-premises production deployment that has since been replaced by another solution.

Client engagement · via Jaden Data / entAIngine Production

Mandate

Designed the solution architecture and the AI processes for a member-facing assistant, wrote the prompts, and was accountable for design quality and for delivering to the customer's expectations.

Key decisions

  1. Prototype from reusable platform components in one day, then spend the project on integration, user-experience validation and the on-premises deployment.
  2. A six-step backend process (transcription, generation, translation, text-to-speech) fed by a retrieval pipeline over the letters and a vector store of supporting context.
  3. Continuous validation with internal user groups and a UX validation group instead of a big-bang launch.

Outcome — qualitative

4 languages at launch — Dutch, Polish, Turkish and English Measured
Recognised by a Dutch ministry as a best-practice example of impactful AI integration
Removes barriers for members who struggled with Dutch letters — "they now understand the point being made"
Superseded reached production, then replaced by a later solution
Stakeholders
  • Member communication owners
  • APG's GroeiFabriek innovation hub and Pensioen Schoonmaak
  • Internal user and UX validation groups
  • IT, security and compliance
Constraints
  • Dutch financial regulation and on-premises hosting
  • Members with limited Dutch and varied digital skills
  • Accuracy on financial content
Reuse
  • Six-step assistant process pattern
  • Letter ingestion and retrieval pipeline
  • Multilingual and audio output components

Context

APG is one of the world’s largest pension providers, administering over €500 billion for a large share of the Netherlands under heavy regulation. Written pension communication is hard to follow for members who are not native Dutch speakers. As APG’s own write-up puts it, participants “have difficulty with the Dutch language” and “are embarrassed about that”.

Business problem

Members needed to understand what a letter says, what the difficult terms mean and what they have to do, in a language they read comfortably, without a call-centre interaction for every letter. The tool that came out of it lets a participant scan a letter from Pensioen Schoonmaak and “have it translated, simplified and read to them”.

My mandate

I designed the solution architecture and the AI processes, wrote the prompts, and was accountable for design quality and for delivering to the customer’s expectations.

Decisions

  • Assemble the first prototype in one day from standard platform components, then dedicate the project to system integration and user experience.
  • A conversational interface backed by a six-step process (audio transcription, text generation, translation, text-to-speech) and a retrieval pipeline that parses incoming letters and pulls supporting context from a vector store.
  • Twelve two-week sprints with tight coordination with the validation groups, so adoption problems surfaced before go-live.

Delivery

The work ran from 2023 into 2024, twelve iterations from kick-off to the system running in the on-premises environment. It was developed by Pensioen Schoonmaak together with APG’s GroeiFabriek innovation hub, and unveiled on 8 November 2023 during the Pensioen3daagse event with four languages — Dutch, Polish, Turkish and English — and the potential to scale to over a hundred.

Delivery shape
  1. Day 1 — Working prototype (from platform components)
  2. 12 sprints — On-premises production (integration, validation, environment)
  3. Handover — Run by the client (under Dutch financial regulation)

Outcome

It started as a proof of concept and went to production. APG’s write-up quotes the effect on participants directly: “the app really does remove barriers. They now understand the point being made.” A Dutch ministry recognised it as one of the best examples of impactful AI integration. It has since been replaced by another solution — a normal end to a first-generation system, and the reason this case is written in the past tense.

It ran in production under Dutch financial regulation, was recognised by a Dutch ministry as a best-practice example of impactful AI integration, and has since been superseded.

Reuse

The assistant process pattern, the ingestion and retrieval pipeline and the multilingual output components became reusable platform capabilities.

Evidence

Prototype and sprint timelines are project records. The launch date, the four launch languages, the participant quotations and the development partners are from APG’s own publication, linked in the facts panel. That it has since been replaced is the client’s status as reported to me.

Want the same thing done in your environment?

This case is one of several. If the shape looks like your problem, the fastest route is to send me the constraints you cannot move.

Remote-first, on-site when it matters; NDA on request