Chief Data Officer / Europe
Three decisions that cost us something
- Gross written premium
- €612m
- 2025; Belgian cooperative non-life insurer
- Member-policyholders
- 410,000
- Each holds a cooperative share; about 168,000 claims a year
- Data colleagues
- 64
- Governance, platform, analytics and reporting data; the data protection office sits outside the team
Decisions & consequences
01Recording motor recoveries as their own transactions, and restating the reserve
In 2016 our claims operation changed how motor recoveries were booked. Under the Belgian market's direct-settlement arrangement we often pay our own policyholder first and then recover from the liable driver's insurer. To shorten handling, the claims system began netting the expected recovery against the payment at the moment of settlement. Paid amounts fell in the most recent accident years, and the reserving actuaries, reading those triangles, saw development that looked faster and cheaper than it was. By the third quarter of 2017 they were applying 41 manual adjustments a quarter from a private extract, and the actuarial function's report warned that the motor best estimate might be understated by €6 million to €9 million.
I proposed one claims ledger in which gross payments, expected recoveries and recoveries received were separate entries. Claims operations would own the definitions of those entries; the actuarial team would own a documented reserving view built from them and reconciled every quarter. Élodie Fraipontier, who holds the actuarial function, opposed giving up her team's extract, which she saw as part of its independence. The claims director objected to an extra field on every motor settlement. We agreed that the actuaries would lose the private copy but gain a veto over any definition change that moved a triangle.
The work cost €1.4 million and took fourteen months. Handlers now spend about 4,000 more hours a year recording recoveries, and the first quarter-end on the new ledger, in June 2018, ran six working days later than usual, though still inside the reporting deadline. Rebuilding history from 2011 added €7.2 million to the motor best estimate and showed €19 million of recoveries still outstanding. Manual adjustments fell to six a quarter, and definition changes now need two signatures.
Next time I would rebuild further back. Accident years before 2011 still rest on an approximation that the actuarial function has to qualify in its report every year, and that qualification is the visible price of my decision to save four months.
Launching income-protection triage with a manual step for two languages
About 9,400 income-protection claims reach us each year from self-employed members who cannot work. In 2023 my analytics team built a model to sort new claims into two routes: a fast route, where a handler confirms the medical certificate and starts payment, and a full route with a case manager and our medical adviser. The model reads fields extracted from the certificate, the occupation class and prior claims. It cannot decline a claim or reduce a benefit, and a member can always ask for a case manager.
Testing by language changed the timetable. French-language claims reached the fast route 58% of the time; Dutch- and German-language claims only 31%. The extraction service had been trained mostly on French certificates, so it marked the others incomplete. Our head of analytics wanted to delay launch until retraining was finished. Pieter Dewaelens, the claims director, wanted to launch in September as planned, arguing that nobody would wait longer than before. Launching for French speakers alone would have made language the basis for faster payment, which I would not support.
We launched for everyone in September 2023. For Dutch and German claims, a handler replaced the extraction with a ten-minute completeness check that fed the same routing rules. Two additional handlers cost about €150,000 a year, and the efficiency saving in the budget slipped by fourteen months. After a data protection impact assessment, we retrained the extraction on 5,800 anonymised certificates and removed the manual step in November 2024.
Median time to first payment fell from 19 days to 10, and the difference in fast-route rates between language groups narrowed to four points. Handlers' overrides of the routing are reviewed monthly with the claims director. If I bought an extraction service again, I would write language coverage and test data into the contract rather than discover the gap two months before launch.
Capping the price gap between renewing and new home customers
In 2024 we compared every home renewal with the price our current tariff would quote a new customer for the same property and cover. The comparison was harder than it sounds. For 31% of older policies we lacked construction year or roof details, so the analytics team imputed them and showed the effect of that assumption separately. The result was uncomfortable for a cooperative. Members with ten or more years of tenure paid a median 13% more than an equivalent newcomer, and those over seventy paid 17% more, because years of renewal increases had compounded while new customers received introductory discounts.
Pricing belongs to the chief actuary, not to me, and the commercial director wanted to keep the introductory discounts that supported our growth through Flemish brokers. European supervisors had already warned the market about differential pricing at renewal. I presented the measure to the executive committee and then to the board's audit and risk committee, with the imputation range beside each figure. I recommended a cap rather than identical prices for everyone, because genuine risk differences between properties are legitimate.
In January 2025 the board approved a cap: no renewing member would pay more than 2% above the equivalent new-customer price, phased in over two renewal cycles. Introductory discounts were reduced. Home premium for 2025 came in €5.3 million below plan, new home policies through brokers fell 11%, and two broker groups moved part of their new business elsewhere. My office publishes the gap monthly to the pricing team and quarterly to the committee.
By September 2026 the median gap was 3.1%, and lapses among long-standing members were unchanged. What I would change is the timing. I had the data for this comparison when I became chief data officer in 2019 and treated renewal outcomes as a pricing matter rather than a data question. Our longest-standing members paid for that with five more years of increases.