AI health coaches: what they actually know about you

What they actually are

An AI coach is a language model with access to your recent metrics and a set of instructions about how to talk about them. It is a layer on top of the data, not a new sensor. That matters more than it sounds: it does not improve accuracy, it does not add measurement, and it cannot see anything the ring did not record. What it changes is the interface — you can ask a question in your own words instead of hunting for the chart that answers it.

Where they genuinely help

Two things. First, translation: turning “HRV 38, RHR 61, temperature +0.4” into a sentence about what those tend to mean together is real work, and doing it on demand is a good use of the technology. Second, context recall — pulling the relevant week of history into an answer rather than making you scroll. Reviewers consistently find coaches most useful for exactly this class of question, where your own numbers change the answer, and least useful for general fitness questions, where the reply is indistinguishable from any chatbot’s.

Where they quietly guess

A coach cannot know why a metric moved. It sees a low HRV; it does not see the argument, the infection incubating, the room at 26°C or the fact that you started a new medication. So it infers, and inference presented in fluent prose reads like knowledge. Language models also produce confident wrong answers on occasion, which is tolerable when you are drafting an email and less so when the topic is your heart. And the advice tends to regress toward the generic — sleep more, hydrate, take a rest day — because that is what most of the training data says.

The question worth asking: where does the conversation go?

Chatting with a coach means sending your health context to a server. Vendors differ on what happens next. Oura states that Advisor runs on infrastructure it controls and that conversations are not routed through third-party AI providers; other products are built on general-purpose model APIs. Neither is scandalous — but it is a question with a real answer, and it is worth reading the specific policy rather than assuming. Note also that for cloud-synced wearables the underlying metrics were already on someone’s servers; the chat adds a second flow, not the first one.

The unglamorous alternative: rules and your own baseline

Most of what a coach usefully says can be produced without a language model at all: compare today against your own rolling baseline, flag what is outside your normal range, and cite the research where research exists. That approach is deterministic — the same inputs give the same answer, every time — auditable, and it can run entirely on your own machine because it needs no server. It will never chat with you. In exchange it will never invent a mechanism, and it does not need your health data to leave the building.

How to use one well, if you use one

Ask it what in your data drove the answer, not just what to do. Treat any specific claim about causation as a hypothesis to test rather than a finding — you can test most of them yourself with a tag and three weeks. Give it the context it cannot see, since a coach that does not know you flew yesterday is guessing about your resting heart rate. And keep the failure mode in view: an assistant that produces an explanation for everything can make a noisy week feel meaningful, which is a slower version of the same trap as checking your score every morning.

Vitra takes the other route: its guidance is research-backed rules plus your own rolling baselines, computed locally on your Mac or PC, with no model API and no health data leaving your machine. You get the reasoning rather than a chat window — which is less magical, and considerably easier to check. General information, not medical advice.

Frequently asked questions