Why the Industry Doesn't Build the Relationship Layer

If nearly everyone in this field agrees that information alone rarely changes behaviour, why does the entire industry keep building on the information layer?

The agreement is real. Some version of data can guide us, but meaning/identity is what keeps us moving circulates constantly in longevity circles, and it gets nodded at immediately and acted on almost never. The next funding round goes into another diagnostic, another wearable, another coaching model built on a large language model.

This is not stupidity, and it is not bad faith. There are structural reasons, and they are worth naming, because they explain a great deal about what is currently being built and what is not.

Four reasons the money goes to data

The first is arithmetic. Software has near-zero marginal cost: build it once, serve a million people, and the cost of the millionth user rounds to nothing. Relational work has the opposite property. Serving twice as many people requires roughly twice as much of the scarcest input there is, which is a person's attention over time. One of these curves is investable and the other is not. Every incentive in the funding environment pushes toward the first.

The second is measurement. Data produces artefacts: dashboards, adherence rates, engagement graphs, biological age scores that move in the right direction. These fit in a deck. The things that actually predict whether someone will still be living differently in three years... whether they trust the person advising them, whether they feel able to admit a bad month, produce no artefact at all. What cannot be measured cannot be presented, and what cannot be presented does not get funded. This is not a conspiracy. It is a filter, and it runs continuously.

The third is a belief so deeply held in technology culture that it rarely gets stated: that people behave rationally when sufficiently informed. It is an inheritance from the Enlightenment, and it is flattering to everyone who holds it. It also does not survive contact with the evidence. Show a person their glucose curve, their sleep debt, their calculated biological age, and you have changed what they know. You have not touched why they stop at the bakery on the walk home, why the workout gets skipped in the fourth stressful week rather than the first, or why the person who most needs to eat more regularly is often the one most frightened of doing so. Information deficit is rarely the binding constraint. But if you assume it is, building a better information product is the obvious move. And the field has been making that obvious move for a long time.

The fourth is time. Fund cycles run five to ten years and need a demonstrable outcome inside that window. Behaviour change that holds does not really declare itself on that schedule. The interesting question about a health intervention is not what happened in twelve weeks; it is what remains at year four, after a job change, an illness in the family, a winter of not coping. Almost nothing in the current funding structure is set up to look that far, so almost nothing is built to be judged that way.

Put these together and the pattern is not mysterious at all. The layer that gets built is the one that scales, measures, flatters the underlying theory of human behaviour, and reports inside a fund's lifetime. The relationship layer fails all four tests. So it goes unbuilt. And not because anyone decided it was unimportant, but because nothing in the system selects for it.

The current attempt

The obvious response, and the one currently being attempted at speed, is to automate it. If the constraint is that human attention doesn't scale, build attention that does. AI coaches are available at three in the morning, never tired, never impatient, never distracted by their own week. They can hold more context about a person than any human practitioner could retain, notice patterns across months of data, and adapt tone to the individual.

These are real advantages, and I expect the tools to keep improving. Some of this they will eventually do better than a person — synthesis, availability, recall, the patient repetition of things that need repeating. Anyone in this field who responds to that with defensiveness is going to be wrong, and will look it.

But one feature of the relational layer this approach does not reach, and it is worth being precise about what it is.

When someone doesn't show up to a conversation with another person, something happens. Not punishment... more a real event, in a relationship they care about. There is someone on the other side who noticed, and the next conversation begins from the fact of the absence. That absence is usually information: about what got hard, what was being avoided, what hasn't been said yet. The missed session is often where the useful material is.

When someone stops opening an app, nothing happens. Not nothing to them, they may feel a familiar private failure, sure, but nothing in the world. There is no other party. The metric registers a churned user.

This is not a capability gap that better models close, because it is not a capability. It is a property of there being a second party with a stake. You cannot generate accountability by simulating someone who cares; accountability is constituted by someone actually caring, and knowing that they do is the entire mechanism. The same holds for the harder half of the work: the moments where a person needs to hear that the plan they are attached to is the problem. A system optimized to keep users engaged, trained toward responses people rate highly, faces exactly the wrong incentive at exactly that moment. That is not a prompt-engineering problem. It is what the system is for.

What this implies for what to build

The conclusion I draw is not that AI will fail. It is something more awkward, and more interesting for anyone actually building in this space.

The most valuable layer in longevity is the one that structurally resists the business model the industry runs on. It does not scale like software. It cannot be demonstrated on a dashboard. It pays back over a horizon longer than most funds tolerate. These are not temporary limitations awaiting a technical fix. They are the properties of the thing itself.

Which leaves a real question, and I think it is the important one in this field right now: what do you build when you know that the part that matters most cannot be scaled in the ordinary way?

One answer is to stop treating that as a defect. A practice designed around depth rather than throughput is not a failed attempt at a platform; it is a different object with different economics, priced accordingly, capacity-bound by design.

That answer comes with a cost I don't want to write around. Work that cannot scale and must be priced to survive is (by construction) available to a small number of people who can afford it. The scaled products (for all the limits described above) reach millions, and reach people no boutique practice will ever see. There is something uncomfortable in arguing that the most valuable layer of health support is the one that resists distribution, and the discomfort is not resolved by pointing out that it is true. Anyone making this argument, myself included, should be honest that it describes a real problem and not only a defensible market position.

I don't have a full answer to it. The partial ones I find credible are that the relational layer may be deliverable through institutions rather than private practices: clinics, employers, public health programmes, which is a question of policy and funding rather than of technology; and that the tools genuinely do lower the floor even where they cannot raise the ceiling. Neither disposes of the objection. It should sit there.

The other move is to let the layers do what each is actually good at: machines for pattern-finding, availability, synthesis, and the tireless work of surfacing what a person would otherwise miss; humans for continuity, judgment, and the uncomfortable conversation. That combination is more interesting than either alone, and it is not what most products are currently attempting, because most products are still trying to make the relational layer unnecessary rather than to support it.

I don't think it becomes unnecessary. I think it stays scarce, and that scarcity is not a market failure to be engineered away. It is the shape of the work.