Gave Underwriting one number it could trust, on Databricks
The case in 30 seconds
- Problem
- Underwriting priced a €1B portfolio on numbers it could not fully trust. The loss ratio differed between the Excel report, the PowerBI dashboard and the actuary’s script. Lost tenders were archived, never analysed. Three people could query the data.
- What I did
- Made Underwriting the owner, not the customer. Built a one-page case around two things the business could name — one reconciled loss ratio and a win/loss view of every tender. Cleared governance before moving data, then ran a four-month pilot with a team of fewer than eight.
- Impact
- One governed number from the actuary to the CEO. Month-end commentary on day one instead of day five. 12 Excel reports retired, 2 pricing models in production, every past tender searchable.
How it happened
- 2025 — Where it started
Underwriting is where the company makes or loses money. The data was good — the problem was what happened after the system of record. Every team pulled its own extract into Excel, PowerBI or a Python script on a laptop. Each returned a different loss ratio; month-end was spent arguing over the right one. Lost tenders sat in SharePoint, never analysed.
- The ownership decision
I changed the owner before the architecture. The Head of Underwriting became sponsor; I delivered. Written down on day one — the system of record stays the master; Databricks is the analytics and AI layer on top. Not a Digital project cut at the next budget review, but an Underwriting asset with a P&L line.
- Two use cases, and no more
A reconciled loss-ratio dashboard, because the Chief Underwriting Officer, the actuary and the CEO were reading three different numbers. Tender win/loss intelligence, because it was visible at CEO level and nobody owned it. The business case fit on one page: cost of the status quo, expected gain, three-year run cost. Everything else was refused.
- Blockers cleared before moving a row
Group IT already had a Databricks tenant; reusing it saved months. The data science office co-owned governance — group standards, not a rival stack. The CISO and DPO signed off classification, masking and hosting up front, with Unity Catalog as the answer. Source-system extraction — as much a contract question as a technical one — was settled in month one.
- A small pilot, in the open
Fewer than eight people, one Underwriting data scientist seconded, a demo every four weeks, FinOps guardrails on spend. Go/no-go at month four on adoption metrics, not on architecture diagrams.
- What it looks like now
- First working day of the month: the dashboard refreshed overnight from one governed source. The Chief Underwriting Officer, the actuary and the CEO read the same number. Commentary starts at 9:00.
- An underwriter opens a tender and gets the five closest past proposals — what was priced, who won, why it was lost. The segment loses on service levels, not price. The draft is ready that afternoon.
- A claims handler asks in plain language what was agreed on deductibles with a partner. The answer is the clause, the date and the source email. No mailbox search.
- Two models feed pricing in production, retrained on full history, versioned. When a data scientist leaves, the work stays. The CISO review takes an hour — access, masking and lineage are on a screen.
- TodayDay 1month-end commentary, down from day fiveOne governed loss ratio from the actuary to the CEO12Excel reports retiredThe reporting team analyses instead of copying2pricing models in productionRetrained on schedule, on full history, versioned
What I took from it
Nobody funds a data platform. They fund one number they can finally trust.