Ageas says it has stopped categorising people by age and will read behaviour and health instead.

The chief executive of a Belgian insurer has written four thousand words on longevity, and it is better sourced than most things of its kind. There is a reference list. It cites the 2024 Lancet Commission on dementia, the American Heart Association on stroke, Eurostat on household composition, the UN on household size. It names which of its sources are internal. Almost nobody does that.

Which is why it is worth reading closely rather than dismissing, and why the two things worth saying about it are both narrow.

The sentence

Here is the passage, in the middle of a section arguing that age has become a poor guide to what a person needs:

As insurers we become increasingly aware of age bias and the risk in defining the needs of individuals by age alone. When people stop recognising themselves in the solution offered by virtue of age, institutions lose the ability to truly understand the populations they serve. So, the lesson we have embraced is to stop categorising people by age.

And immediately after:

This is where Data and AI can play such an important role. Data and AI allow us to look at behaviour, life events, preferences and health indicators to ensure greater relevancy.

Taken as written this is an argument against a crude variable and in favour of a better one, and as an argument about customer experience it is hard to object to. A sixty-year-old starting a company and a sixty-year-old caring for a parent do not want the same letter.

The difficulty is the verb. Categorising is doing a great deal of work, and an insurer categorises people for several quite different purposes: to decide who receives a mailing, to decide which product to show someone, to decide whether to accept a risk, and to decide what to charge for it. Those are the same activity to a data team and four different activities to a regulator.

The essay does not say which it means. It may well mean only the first two, and the context — relevance, intervention timing, recognising yourself in the solution offered — points that way.

Why the distinction is not pedantry

Because of where the fourth one sits. Annex III of the AI Act lists, among the high-risk uses:

AI systems intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance

Marketing segmentation is not in that line. Risk assessment and pricing is. The rules for Annex III systems apply from 2 December 2027, which is fourteen months away.

So an insurer saying it is replacing age with "behaviour, life events, preferences and health indicators" has described something that is either outside the Annex entirely or squarely inside it, depending on a word the sentence does not contain. For a life and health insurer that is not a shade of meaning. It is the difference between a product decision and a conformity assessment.

I do not think this is concealment. I think it is an essay written about customers that has not been read back by someone thinking about Annex III, which is an ordinary thing to happen to a good document. But the sentence is now public, it is the most quotable line in four thousand words, and it will be repeated by people who will not add the distinction.

There is also a smaller irony, which the essay supplies itself. The replacement for age is health indicators. Age is visible, cheap to audit and already regulated. Health data is special category data, and a model that reads it is harder for anybody outside the firm to inspect than a birth date ever was. Moving from one to the other may well be fairer. It is certainly less legible.

The seven numbers

The essay's second feature is its figures, and this is where its honesty and its unverifiability sit in the same place.

• Pension preparedness

  One in three Belgians over 50 has a pension pot of EUR 50,000.

• Prevention

  75% would adapt their habits if it helped prevent future illness.

• Employer-funded care

  52% of working Belgians would finance long-term care through a flexible employer benefit.

• Informal care

  24% of Belgians aged 50+ provide informal care daily or weekly.

• Ageing at home

  More than 95% want to; fewer than half manage it.

• Administrative burden

  42% feel overwhelmed by organising care.

• Protection mandates

  13% have one. 41% do not know what it is.

All seven rest on three items the essay's own reference list marks 'Internal source, no public link available'. Sample size, fieldwork dates, question wording and method are not stated. This desk does not say the numbers are wrong; it says nobody outside Ageas can check them, and each of them supports a product the company sells.

Note what the essay does here, because it is better than the convention. It marks them. It does not dress them as independent research, does not route them through a think tank, and lists them under their own name in a bibliography that also contains the Lancet. An essay that wanted to launder these numbers had easy ways to do it and took none of them.

What the marking cannot do is make them checkable. The figure that most of the argument leans on — one in three over-fifties with fifty thousand euros — is the one that would most repay publication, because it is the number a policymaker would quote. It comes from an item the reference list calls a segmentation workshop.

One other citation is worth a line. The centenarian figure, an eightfold rise to about 3.7 million by 2050, is credited to Pew's analysis of UN projections; the reference list dates that Pew publication to 2016. The same reference list cites the UN's World Population Prospects 2024. The newer source is in the bibliography and the older number is in the text.

The borrowed metaphor

The essay takes its framing from Dario Amodei's essay on technology's adolescence: a period in which capability develops faster than maturity, where something powerful arrives before the structures around it have caught up. Applied to longevity it works well, and it is attributed.

It is then used to introduce a section whose remedy is more capability. Data and AI arrive to solve the gap that the metaphor exists to describe, and the essay does not say which structure is supposed to mature to meet them. That is not a contradiction, exactly. It is an unfinished sentence in a piece that finishes most of its others.

The closing is a vignette from the summer of 2056, looking back on the 2020s as the turning point. It is a prediction with nothing attached that could fail. If I were marking this essay I would ask for one line saying what, by 2030, would show the gap was not closing — and I suspect the author could write it in a sentence, because the rest of the piece is specific enough to suggest he knows.

What this is, and is not

It is a serious essay by someone who has read the Lancet Commission and can name three household-composition sources. The direction it argues — earlier conversations, prevention as a first-order concern, pension provision that does not rest entirely on the state — is defensible and widely held, and this desk has no quarrel with it.

It is also a document in which an insurer announces it will stop using the one segmentation variable its customers can see, and does not say which of its four meanings of categorise it is retiring. Fourteen months before Annex III applies, that is the sentence to put the question to. Nobody has, including us: no comment was sought from Ageas for this piece, which is a limitation of it rather than a feature.