What an AI fitness app really is: four unrelated products sold under one name

It is Sunday night and you have three apps open. All three say AI coach. All three show a workout on a dark background with a graph behind it. One is free to start, one is about a tenner a month, and one costs more than your gym membership. Nothing in the listings tells you what is doing the thinking, so you choose on star ratings, the one signal that says nothing about the mechanism.

The reasonable assumption is that these are competing versions of one product and the difference is polish. They are not competing. One is a set of written if-then rules running over your logged sets. One is a language model that explains a Romanian deadlift beautifully and cannot change your Thursday. One is a qualified human with a phone. One is a fixed catalogue of filmed sessions with a quiz in front of it. All four are entitled to the phrase, and all four are priced as though they were each other.

Once you can name the types, most of the buying decision collapses into a question about you rather than about the market. This piece is the taxonomy on its own. The full six app comparison lives in the best AI personal trainer app in 2026.

Our bias, stated first

We make Pocket Fit and we sell a subscription. A taxonomy written by one of the things being classified deserves suspicion, so read the type we assign ourselves as a claim from an interested party rather than a verdict.

What we can do is make it checkable. Every factual claim about Freeletics, Future, Caliber, Centr and Ladder comes from that company's own site, help documentation or App Store listing, verified for the hub comparison in July 2026, with sources at the end. Prices move and vary by region, so treat every figure as "at the time of writing".

One honesty tax up front. Two of the products below put a real qualified human in your pocket, and for a meaningful number of people that is simply better than any software, ours included.

The four things "AI coach" is allowed to mean

Four distinct technologies wearing the same jacket. Most real products are a blend of two of them, which is fine, and the blend is almost never disclosed, which is not.

Type 1: a deterministic rules engine

A fixed, written set of if-then rules operating on the data you logged. Hit the top of the rep range twice, add the smallest increment your equipment allows. Miss the target twice, hold. No model is involved at the moment of decision, which is why this is the least glamorous type and the one most likely to be doing the actual work when your lifts move.

Its virtue is that it is auditable: you can read the rule, predict tomorrow's prescription, and catch it when it is wrong. Its limit is that it only knows what you told it.

Type 2: a generative or retrieval model

A language model, usually with retrieval over some knowledge base, that answers questions, explains movements, suggests substitutions and rewrites a session in plain English. This is what almost everyone pictures when they hear "AI coach", and it is genuinely useful.

It is also the type most often asked to do the job it is worst at. A model that generates plausible text generates a plausible working weight with exactly the same confidence as a correct one. Used well it sits in front of a Type 1 engine and translates it. Used badly, it is the engine.

Type 3: a human coach, delivered through software

A marketplace or agency that matches you with a certified trainer who writes your programme, messages you when you disappear, and watches a video of your squat. The app is the delivery mechanism, not the intelligence.

It still gets marketed with AI language, because there is an algorithmic layer somewhere in the product, usually in the tracking rather than the coaching. If your constraint is accountability rather than knowledge, this is the type that solves it and no software price point does.

Type 4: a static library with a recommender on top

A fixed catalogue of pre-made programmes or filmed sessions, plus an onboarding quiz that routes you into one. Nothing is generated for you and nothing changes because of what you did yesterday. It is usually called "personalised" rather than AI, and occasionally called AI anyway.

A library can be excellent. Production quality, real instructors, breadth across strength and mobility and yoga: none of that is a criticism. The only claim here is that it is not adaptive, so you should not pay adaptive prices for it.

Which apps sit where, and why the type decides the price

Applying it honestly, ourselves included. Primary type first, secondary layer second.

Now read the prices against the middle column and the spread stops being mysterious. Future lists 50 US dollars for a first month and 199 a month after that. Caliber runs a free tier, Plus in the region of 6 to 12 dollars a month, Pro around 19, and Premium one-to-one coaching starting around 200. Centr lists an annual subscription at 179.99 dollars before discounts, Ladder 29.99 a month or 179.99 a year. Pocket Fit is free to start with no card.

That range spans roughly two orders of magnitude and tracks almost perfectly with one variable: whether a human being is on the other end. Type 3 costs what a person costs, because it is a person. Type 1 and Type 2 cost what software costs. Type 4 costs what a content catalogue costs, and a good catalogue is worth paying for.

The failure mode is a Type 4 product priced like a Type 3 one. Then you are paying for production values rather than coaching, and the thing you thought you were buying was never in the box.

Where the category's claims run thin, including ours

"Adaptive" is not a defined term, in the marketing or in the science. Greig and colleagues reviewed autoregulation in resistance training for Sports Medicine and found that although the practice dates to the 1940s, the field uses adaptation, readiness, fatigue and response interchangeably, which leaves the implementation of any given autoregulation strategy ambiguous. Their paper exists to propose operational definitions, because the literature did not have consistent ones. If sports scientists cannot agree what adapting means, a store listing will not pin it down for you.

Nobody has tested type against type. There is no randomised trial comparing a rules engine, a filmed library and a human coach in the same population over the same months. Every ranking in this category, ours included, reasons from mechanism rather than from outcome data.

The best evidence here is unglamorous and it favours logging. Michie and colleagues pooled 122 evaluations covering 44,747 people for Health Psychology. Behaviour change interventions for physical activity and healthy eating produced an overall effect of 0.31, and self-monitoring explained more of the variation between studies than any other single technique, with interventions that paired it with another self-regulation technique reaching 0.42 against 0.26 for those that did not. That is support for tracking what you do, not for any particular flavour of AI on top of it.

The type is self-declared. We classify products from what their makers publish, and a company can describe a rules engine in language that sounds generative, or the reverse. We call our own progression Type 1 because the rule is written down where you can read it, which at least makes the claim checkable.

How to choose in five minutes

Four questions, in order. A free trial and the app's own support pages usually answer all of them.

• "Show me the rule." Ask what causes the weight to go up. A written answer means Type 1 is in there somewhere. "The algorithm adapts to you" means you have learnt nothing, and possibly so has it.
• "What happens if I log a bad session?" Type 1 and Type 3 change something. Type 2 discusses it sympathetically. Type 4 plays the next video regardless.
• "Who wrote this week?" A generator, a named human, or a catalogue editor three years ago. All three are legitimate. Only two are thinking about you.
• "What happens if I vanish for a fortnight?" Most products pretend it did not happen or punish you for it. Very few plan for it, and this decides whether you are still training in March.

If your honest answer is "I know what to do, I just do not go", buy Type 3 and stop reading comparisons: Future vs Caliber covers that choice, and can an AI fitness app replace a personal trainer covers what you give up either way. If you already have a programme and only need it recorded, you want a tracker, which is Pocket Fit vs Fitbod.

Where the evidence and our claims run thin

This taxonomy is a buying lens, not a regulated classification. Companies blend types and market them under one phrase. Prices cited from public pages move. Greig and colleagues show how loosely "adaptive" is defined in the training literature; that supports scepticism toward screenshots, not a ranking of every app store listing. We sell a Type 1 product with a Type 2 layer, so our examples lean that way.

Put it together

The two words on the icon tell you nothing. The type tells you almost everything: what the product can change, what it cannot, and what it is reasonable to pay. A rules engine moves your loads and cannot notice you are miserable. A language model explains anything and should not be trusted with a number. A human notices you did not train on Wednesday. A library plays whatever comes next.

Pocket Fit is Type 1 with a Type 2 layer, and we would rather say so than call it magic. The numbers come from a written rule: reps first inside the range, weight only off a confirmed plateau of two misses rather than one, because a single bad set is more often a bad night than a real ceiling. The model picks exercises and explains them, and the AI coach rebuilds a session from one sentence when the rack is taken or your shoulder is unhappy. The scheduler moves a missed session into the rest of the week instead of deleting it, Fuel keeps food in the same app so a stall is diagnosable, and the body budget runs one tally across your workout, your streak, your sleep and your nutrition. For the same distinction against the nearest Type 1 rival, read Pocket Fit vs Freeletics.

The founder note, because it explains the design rather than decorates it. Georgi built Pocket Fit after going from 122 kg to winning first place at The Yard Games, having lost 38 kg on the way. The app is not shaped by the good weeks but by the bad ones, which is our story, and why the type we chose is the boring auditable one.

What an AI fitness app really is: common questions

What does AI actually mean in a fitness app?

Usually one of four things: a deterministic rules engine acting on your logs, a language model that explains and rewrites text, a human coach delivered through an app, or a static library with a quiz in front of it. All four are sold with the same two words, and the store listing will not separate them. Ask what causes the weight to go up.

Is an AI personal trainer app worth it?

It depends entirely on which type you buy. A rules engine that generates a programme and progresses load on a published rule is worth it if your problem is not knowing what to do, at a fraction of the 199 dollars a month human coaching runs to. A filmed library is worth it if you value the content. If your problem is that you will not go unless somebody is expecting you, no software price point fixes that and a Type 3 coach will.

Which type is Pocket Fit?

Type 1 with a Type 2 layer. Progression is a written deterministic rule, reps first inside the range and weight only off a confirmed plateau, while a retrieval model handles exercise selection, explanation and the chat coach. We publish the rule so you can check that rather than trust it.

Can a chatbot write my training programme?

It can write something that reads like one, which is the problem. A generative model produces a plausible weight as confidently as a correct one, so it belongs in front of the numbers rather than in charge of them. Explanation, substitutions and shrinking a session to 30 minutes are where it earns its place.

How do I tell a static library from a genuinely adaptive app?

Log a bad session and watch what happens next. An adaptive product changes something, a human coach asks about it, and a library shows you the next video exactly as it would have anyway. The vocabulary is no help, partly because Greig and colleagues found the underlying science uses adaptation, readiness and fatigue interchangeably too.