Can an app actually coach you? What your data can and cannot tell you to do

Short answer

An app cannot coach you the way a good human coach does — it has no judgement about your life and no stake in it. What it can do, and a human cannot, is run a controlled before-and-after test on your own body. A systematic review of 137 studies of chatbot health advice found 99.3% did not even report which model version they tested, so treat AI advice as a prompt to look, not an instruction to follow.

What people actually mean when they ask for a coach

Three different requests hide inside the word, and apps are good at one of them.

"Tell me what to do." A specific instruction for today. Most apps answer this with a generic recommendation dressed up in your name — go to bed earlier, take a walk, breathe. It is rarely wrong and rarely useful, because it did not need your data to produce it.

"Keep me honest." Accountability. This is mostly a social function, and an app can substitute for it only weakly: a notification has no relationship with you and nothing to be disappointed about.

"Tell me whether this is working." Did the earlier bedtime do anything? Is the magnesium worth the money? Does that weekly meeting really wreck me, or do I just dislike it? This is the question data can answer better than a human being, and it is the one almost nobody asks first.

A coach can watch you train and see your form. What no coach can see is your heart rate at 3 a.m. That asymmetry is the whole of what an app is good for.

What an AI health coach gets wrong — and why

There is now a large literature on language models giving health advice, and the most useful finding is about the quality of the literature itself. A 2025 systematic review in JAMA Network Open examined 137 peer-reviewed studies evaluating AI chatbots giving health advice. Of those, 136 (99.3%) evaluated closed-source models and did not give enough information to identify which version was tested. All of them lacked an adequate description of the model's configuration. And 136 (99.3%) did not describe a prompt engineering phase in their study.

That is not a finding about whether AI coaching works. It is a finding that we mostly do not know, because the studies are not reported in a way that would let anyone tell. When someone claims an AI coach is "clinically validated," that is the context to put it in.

The second problem is structural and has not been solved by better models. A language model generates plausible text. When it does not know something, the shape of the output does not change — it is just as fluent and just as confident. Our own help centre documentation for Welltory AI Coach says this in as many words: it warns that the models can produce hallucinations or inaccurate answers, that it is an experimental beta, and that health advice from it should be double-checked with a medical professional. We would rather publish that sentence ourselves than have you discover it.

The third problem is the one people feel but rarely name. Advice that arrives without evidence attached is indistinguishable from advice invented on the spot. "Try going to bed 30 minutes earlier" is fine guidance and tells you nothing about you. The thing that makes it worth acting on is being shown the three nights where it already happened by accident and what your body did.

The one thing data genuinely does better than a human coach

Run an experiment on you, specifically.

This is a real method with a name. An n-of-1 trial is a controlled experiment where the single participant is also the entire study — alternating periods of doing and not doing the thing, with the outcome measured the same way each time. It has a formal reporting standard, the CENT extension to CONSORT, published in The BMJ in 2015. It is not biohacker folklore; it is how you establish whether something works for one person when population averages cannot tell you.

The example worth knowing is the 2021 statin trial, also in The BMJ. Two hundred people who had stopped or were about to stop statins because of muscle pain each ran a personal blinded trial: six two-month periods of atorvastatin or placebo in randomised order, not knowing which was which. Across the 151 who completed enough periods to analyse, there was no difference in muscle symptoms between the statin and placebo months. Two-thirds of those who finished reported restarting treatment.

Two lessons sit in that result, and they point in opposite directions.

The useful one: a structured personal test can overturn a belief about your own body that felt completely solid. These people were not imagining their pain. On average, across the group, it was no worse during statin months than placebo months — which is not the same as proving that no individual's symptoms had anything to do with the drug, but it was enough to change what two-thirds of them decided to do next.

The cautionary one: the control is what made it work. An unblinded self-experiment where you already believe the answer tends to produce the answer you believe. This is the honest limit of self-tracking, and anyone selling you self-experimentation without mentioning it is selling you something.

So the realistic version for everyday life is modest: measure the same way each time, change one thing, keep a note of what else was happening, and look at a trend rather than a day. That is weaker than a blinded trial and considerably stronger than a hunch.

One hard limit on all of this: never start, stop or vary a prescribed medication in order to run a self-experiment. The statin trial above was designed and supervised by clinicians, with placebo and blinding, precisely because that is not something to improvise.

What Welltory actually measures here

The coaching layer in Welltory is built on measurements rather than on advice, so it is worth saying exactly what is measured and what is inferred.

The physiological base. Every HRV measurement produces the standard metrics, visible in Measurement details rather than hidden behind a score: SDNN (overall variability), RMSSD (the parasympathetic, recovery-linked measure), pNN50 (the proportion of adjacent beat intervals differing by more than 50 ms), AMo50 (the sympathetic, mobilising counterpart), Mean RR and MxDMn, plus frequency-domain values — Total Power, VLF, LF, HF and the LF/HF ratio — when you measure in 300-beat mode. Each measurement carries an accuracy score, so a reading taken while you were moving is flagged rather than quietly folded into your trend.

Episodes — the unit that makes coaching possible. Welltory detects stress and recovery episodes during the day: stretches where heart rate and the related signals show a clear pattern. It detects the episode from physiology; it has no idea what caused it. That gap is deliberate, and filling it is your half of the work.

Tags and My patterns — the context layer. When an episode is detected, the app asks what was happening. You answer in your own words, by voice or by typing, and tags are extracted from the answer — deadline, coffee, walk, argument — which you can correct before saving. Over time My patterns shows the balance between your stress and recovery episodes, which tags recur most in each, which ones carry more weight than their frequency suggests, and for any single tag: how often it appears, which weekdays it clusters on, whether it is trending up or down, and what your heart rate did while it was active. Tags only count when attached to an episode, which is what keeps this from becoming a diary.

Experiments — the before-and-after test. Take a measurement, do the thing, tag it, take a second measurement within three hours, and you get a report on what changed. The guidance in our help centre is the guidance any experimenter would give: measure immediately before and after so other factors do not creep in, wait for your heart rate to come down after exercise, and test one thing at a time, because multiple tags at once make the effects impossible to separate.

Self-Discovery experiments — a slice of a real day. With an Apple Watch, you can record a 1, 1.5, 2 or 3-hour window of ordinary life — a meeting, a commute, a walk, an evening — and get it back broken down minute by minute: where stress built, where energy went, what brought it back. Two hours is the recommended length, for a reason that is more psychological than technical: long enough for something to happen, short enough that you still remember the details and can match them to the trace.

How we interpret the numbers. Against your own baseline first. The published reference ranges are extremely wide — for RMSSD, pooling the literature gives roughly 13–107 ms, which cannot tell anyone anything. Welltory's working range is 20–89 ms, narrowed with an internal analysis of artefact-free RR intervals from over 600 Welltory users who had measured for at least 90 days — our own analysis, not externally peer-reviewed. Even that is a population range. The comparison that changes what you do tomorrow is against your own history.

The accountability problem, and the one version of it that works

Earlier I dismissed accountability in a sentence, which is unfair to the people for whom it is the entire reason they wanted a coach. It deserves a straighter answer.

Streaks and reminders work on a narrow band of behaviour: things that are easy, quick and already nearly habitual, where the only obstacle is forgetting. For anything that is hard rather than forgettable — training when you are tired, going to bed when you are not ready, eating differently under stress — a notification is pushing against the wrong resistance, and the usual result is that the streak becomes the goal. A broken streak then does real damage, because the thing you have been maintaining is the record rather than the behaviour, and once the record is gone the behaviour has nothing holding it up.

The version that does work is narrower and less satisfying: make the consequence visible rather than the compliance. Not "you missed your goal," but "here is what your body did on the three nights you went to bed after one." The first is a scold from software. The second is information, and it keeps working on the days you ignore it.

This is also why the useful unit is a fortnight rather than a day. A single day tells you almost nothing — the noise in any of these measurements is larger than one day's signal — while two weeks of a changed habit against two weeks of the old one is a comparison you can actually act on. An app that insists on a daily verdict is manufacturing urgency out of noise.

What this still cannot replace

Being direct about this, because the gap matters.

Judgement about your life. Whether to leave the job, whether the relationship is the problem, whether to push through this week or stop. No dataset contains the information needed to answer those, and anything that answers them confidently is performing.

Clinical assessment. HRV is not an ECG. An app cannot detect or exclude a heart condition, cannot diagnose a sleep disorder, and cannot tell you whether a symptom needs investigating. Persistent symptoms belong with a clinician, and the app's role there is to carry a record of what happened, not to form an opinion.

Technique. If your deadlift is wrong, nothing measurable here will tell you so.

And a particular caution for energy-limiting conditions. In ME/CFS, long COVID, POTS and similar conditions, the usual logic of a recovery score can mislead: a reading that looks good can appear during a crash rather than during real recovery. HRV in these conditions is an active research area and we would not claim to know why this happens. If that is your situation, a trend read with a clinician is worth more than any daily number.

Before you decide you need more discipline

The conclusion most people reach from a tracker is a conclusion about their character. The number is down, the streak is broken, and what arrives is I am not trying hard enough.

It is worth noticing how rarely that is the explanation, and how often the data already contains a better one. Resting heart rate drifting up for three days before you feel the cold. The two short nights behind the Thursday where you could not concentrate. The meeting that shows up in your stress episodes every single week while you keep telling yourself you are being dramatic about it.

None of that is a verdict on you. It is a record of what happened, and it is usually kinder than the story you were telling yourself — and more actionable, because a cause you can see is a cause you can change one piece at a time.

Before you blame yourself, check what your body already recorded.

Start understanding your body

Learn what affects your energy, stress, sleep, and daily state. Get the app.

Explore your data

How Welltory helps — and what it cannot do

What it does. Measures HRV and shows the standard metrics underneath, not only a score. Detects stress and recovery episodes from physiology and asks you for the context it cannot know. Builds those answers into patterns over weeks — what reliably drains you, what reliably restores you, and what matters more than its frequency suggests. Supports before-and-after experiments with explicit guidance on how to run them so the result means something. Compares everything against your own baseline rather than a population average.

What it cannot do. It cannot make a decision for you, and it should not pretend to. Our AI Coach runs on a general-purpose language model and can be confidently wrong, which our own documentation says plainly — treat it as something to check, not something to follow. Unblinded self-experiments are vulnerable to your expectations, and knowing the mechanism does not make you immune. It needs weeks of consistent measurement before the baseline means anything, which makes it a poor fit for anyone wanting an answer this week. And it is not a medical device: it does not diagnose, treat or monitor any condition.

Discounts for blog readers: up to 36% off

See what affects your energy, stress, sleep, and daily state with Welltory

Explore your data