How Much Should a Health App Do for You Automatically?
Published · 5 min read
Health software has started acting rather than only reporting. It logs the meal from a photo, moves the workout, adjusts the sleep target, nudges the supplement. The useful question is no longer whether an app can do things on your behalf. It is which things it should do without asking, and how you find out where that line is before something happens you did not want.
The short answer: automate anything that only writes down what already happened, ask before anything that changes what you take or how you train, and stay silent on anything you have not asked about. If a tool cannot tell those three apart, the automation is a liability rather than a feature.
The three tiers, and why the boundary sits where it does
Recording. Capturing a meal, a dose, a session, a sync from a device. Nothing here changes your behaviour, it only stops the history having a hole in it. This should be fully automatic, because hand entry is the part that quietly stops after a few weeks, and the stretch you most want to look back at later is the stretch you stopped logging.
Interpreting. Working out what your own numbers mean, what changed, what is drifting. This should surface on its own, because an interpretation you have to go and ask for is one you will not ask for on a busy week.
Changing. Altering what you take, how hard you train, when you sleep. This should always wait for you. Not because software cannot reason about it, but because the cost of a wrong automatic change is paid by your body and the cost of one extra tap is paid by your thumb.
| What the app is doing | Who should decide |
|---|---|
| Writing down what already happened | The app, every time |
| Telling you what your own data means | The app, unprompted |
| Changing what you take or how you train | You, every time |
| Choosing which of these is which | Also you, without a settings page |
The last row is the one nobody gets right
Most tools do have a boundary. They just make you go and find it. You are handed a settings screen and asked to pick a level of automation before you have any idea how the thing behaves, and you learn your choice was wrong when something happens you would have stopped.
That is the wrong moment to be configuring software. The information you need to set the boundary correctly only arrives after you have watched the tool work for a few weeks, and by then you are no longer in the settings screen.
There is a better signal sitting in plain sight, and it is one you already send: what you reverse. An undo is the clearest statement you can make that you wanted a say in something, and it costs you nothing to send. A tool that never notices them is asking you to say the same thing twice, once by undoing and once in a menu.
This is roughly how NUVARD is built. You set the level per category, so the boundary is yours rather than guessed at. Logging runs on its own, because a record that needs typing is a record that stops. Anything that would change what you take or how you train is drafted, shown alongside the data behind it, and held until you approve it. And if you find yourself reversing something, that is taken as a sign you wanted a say, and you are asked sooner next time, so a choice you made once does not have to be right forever.
Four questions to ask any tool that acts on your behalf
Can you see what it did, after the fact? An action you cannot find in a log is an action you cannot audit. If the app does not keep a plain list of what it changed and when, you have no way to tell a good automation from a quiet one.
Can you undo it in one step? If reversing something takes more effort than it took the app to do it, the asymmetry will wear you down and you will stop reversing things you disagree with.
Does reversing it change anything afterwards? This is the real test, and most tools fail it. If you correct the same thing three weeks running and the fourth week it happens again, you are not being listened to, you are being asked to go and configure.
Is it silent about things you never asked about? A tool with an opinion on everything is a tool you will mute. Restraint is a feature, and the categories you left untouched should stay untouched.
What this looks like in practice on your own stack
You do not need to wait for better software to apply the idea. Three habits do most of the work.
Keep the automatic recording and be sceptical of the automatic advice. The parts of your history that record themselves are the parts you will still have in six months, which is the only timescale on which a personal baseline means anything. Our guide on how long until your baseline is accurate covers what that settling period actually looks like per metric.
When a tool tells you to change something, ask what it compared you against. A recommendation scored against a general population is a statement about strangers. One scored against your own recent history is a statement about you, and only the second is worth acting on. What is a personal health baseline explains the difference.
Notice which corrections you make more than once. That repetition is the boundary your software should already have learned, and writing it down yourself is the manual version of the same thing.
Related reading
- How to Tell If a Health App's Predictions Are Real, the same demand for falsifiability applied to forecasts rather than actions.
- What Happens to My Health Data, what an app may do with the record it keeps for you.
- Too Much Health Data, Where to Start, what to look at first when several apps are all talking at once.
Automation is not the risk. Unbounded automation is, and so is a boundary you have to go and find.
If you want a system that acts where you let it, asks where you want it to, and treats your corrections as worth listening to, that is what NUVARD is built to do. NUVARD is coming soon to iOS and Android, with no date announced yet, and everyone gets the same access on the same day. Join the waitlist at nuvard.ai.
NUVARD provides wellness intelligence. It does not diagnose, treat, or replace medical care.