AI Fitness Coach: What the Category Actually Does
Published September 29, 2026
An AI fitness coach does three separable jobs, and most products are strong at one of them and weak at the other two. Programme generation is largely solved, progression tracking is bookkeeping, and technique coaching is not something any of them can do. Here is how to tell which you are buying.
A disclosure that should shape how you read this: we build a nutrition database, not a fitness product. We have no app in this category to sell you, and the training guidance below is general and widely published. The nutrition half is where we have actual expertise.
What does an AI fitness coach actually do?
Three jobs, often bundled and rarely equally good.
| Job | State of the art | What to expect |
|---|---|---|
| Generating a programme | Largely solved | Sensible splits, sane exercise selection, reasonable set and rep schemes |
| Tracking and adjusting | Bookkeeping, done well | Logged reps drive the next session's target, which humans are bad at doing consistently |
| Coaching technique | Not available | No product can see how you move well enough to correct you |
The word "coach" does most of the marketing work in this category, because it implies the third row. Almost nothing delivers it.
The three kinds of product
Chat assistants. ChatGPT, Claude, Gemini. Free or near free, excellent at constraints and explanation, no memory of your training unless you supply it, and no progression engine. Covered in ChatGPT workout plan.
Dedicated apps with a progression engine. These hold your history and adjust loads from logged performance using defined rules. The genuine advance over a chat assistant, and the rules are often simpler than the marketing implies.
Camera-based form checkers. A separate technology using pose estimation. Useful for gross movement feedback like depth or bar path in good conditions. Not equivalent to a coach watching you, and heavily dependent on camera angle and lighting.
What should you check before paying?
Six questions, and the answers are usually findable in a few minutes.
1. Does it hold your history? If the programme does not change based on what you actually lifted, it is a template generator with a subscription.
2. Can you read the progression rule? A product that explains "add weight when you complete all sets at the target RPE" is being straight with you. One that says "our AI personalises your training" is not telling you anything.
3. Can you export your data? Training logs are worth more the longer they run. A product you cannot leave is a product that does not have to keep earning you.
4. What happens with an injury? The right answer is that it tells you to see a professional. A product that programs around pain is overreaching.
5. Does the nutrition side use a real database? Most of these bundle macro targets and meal suggestions. Ask whether the food figures come from a database or are generated. This is the question most vendors answer vaguely.
6. Does it ever tell you to do less? Deloads, rest and reduced volume are part of training. A product that only ever escalates is optimising for engagement.
Running, strength and bodybuilding are different problems
The category gets discussed as one thing and splits cleanly into three.
Running. The most formulaic and therefore the best served. Easy mileage, long runs, a taper, and a gradual build are standardised enough that automated plans do well. The failure mode is raising volume faster than tissue adapts, which is exactly what a plan cannot observe.
Strength. Well served for programme structure, because linear progression and its successors are heavily documented. Load selection is where it breaks, since the correct weight depends on how you feel today.
Bodybuilding. The hardest, because progress is slow, noisy and visual. Automated volume prescriptions are stated with more confidence than the evidence supports, and this is where you should treat ranges as ranges.
What would a genuinely good one look like?
Useful to describe, because it gives you a yardstick rather than a feature list.
It would ask what you can currently do before prescribing anything, rather than inferring from a signup form.
It would hold a real training log and refer to it. Not "personalised" as a marketing word, but visibly: this lift has not moved in three sessions, here is what changes.
It would state its progression rule in plain language and tell you when the rule expires. Linear progression works for a beginner and then stops, and the honest product says so in advance.
It would prescribe rest. Deloads, lighter weeks, and occasionally telling you to do nothing. A product that only escalates is optimising for engagement.
It would refuse to program around pain and send you to a professional.
It would separate what it knows from what it is guessing. Programme structure is well evidenced. Precise weekly set counts are not. A good product signals the difference instead of stating everything with equal confidence.
Very few products do all six. Using them as a checklist while you evaluate is more useful than reading reviews, because these are properties you can observe in the first week of a trial.
How much of this is new?
Worth some perspective, because the category is marketed as a recent breakthrough.
Automated programme generation is decades old. Spreadsheets that adjusted loads from logged performance were common long before anything was called AI, and the progression rules inside many current products are the same arithmetic with a better interface. That is not a criticism: the arithmetic works, and a good interface is genuinely valuable because it is what makes people record sessions at all.
What language models added is the conversational layer. You can now describe an awkward constraint in a sentence and get a fitted programme, which previously meant either a coach or manual editing. That is a real advance and it is a narrower one than "AI coach" implies.
What has not changed is the thing that limits all of it: none of these systems can see you move, and technique under load remains the main risk.
Where every option is weak
None of them can see you. Worth repeating because it is the whole safety story. Technique under load is the largest injury risk in a gym and the least amenable to remote instruction.
Adherence is still yours. A notification is not accountability. The published work on exercise adherence consistently finds that the programme people sustain beats the optimal programme they abandon.
Recovery is invisible. Sleep, stress and life load determine what you can handle, and a set-and-rep log sees none of it.
The nutrition half is usually the weakest part. A fitness product's core competence is training, so food data is often bolted on, and that is where confident wrong numbers live.
Novelty bias. These systems generate readily, which means asking for something new always produces something new. That is a poor match for training, where the results come from repetition. The tool will never tell you that the best action is to run the same programme for another six weeks, because that is not an output.
They optimise what they can measure. Sets, reps and logged weight are easy to record, so they drive everything. Sleep, stress, work load and how you actually felt are not recorded and therefore do not exist to the system, even though they determine what you can recover from.
How the three kinds handle a real constraint
A concrete example makes the differences obvious. Say you have four days, forty-five minutes, a shoulder that dislikes overhead pressing, and a deadlift you want to keep progressing.
A chat assistant handles all four in one request. It will restructure the split, swap overhead work for incline or landmine variations, and keep the deadlift on its own day. It will not remember any of this next week.
A dedicated app handles the first two if its signup form has fields for them, and usually ignores the shoulder unless it has an injury setting, which many do by simply removing the movement pattern entirely rather than substituting within it. It will remember.
A camera-based tool has nothing to say about any of it, because it operates at the level of a single rep.
That pattern generalises: assistants are best at fitting, apps are best at remembering, and neither is coaching. The reason people bounce between them is that each solves the problem the other creates.
The question nobody asks vendors
Worth adding to the six above, because it is the most revealing single question: what does your product do when the user stops making progress?
A serious answer describes a rule. Volume increases, or an intensity block, or a deload followed by a reset, with criteria for each. A weak answer talks about personalisation and adaptation without describing anything you could observe.
Stalling is the normal state of training beyond the first few months. A product whose only story is "it gets harder over time" has not thought about the situation you will actually be in most of the time.
The half that can be checked
A training plan cannot be verified against a database. Whether four sets of eight suits you is a judgement. Food composition is different: how much protein is in 180 g of cooked chicken thigh has a right answer sitting in a table.
So when a fitness coach product hands you a macro target and a meal suggestion, ask where the numbers came from. If they were generated rather than retrieved, they carry error that compounds across a training block, and the direction is predictable because the under-counted items are oils, dressings and anything eaten standing up.
The structural fix is a tool that queries a food catalog, so an assistant searches, gets a row with macros per 100 g, and scales it to your gram weight. That is this nutrition MCP server. It authenticates with an API key header and is verified with Claude Code and Cursor. Tools and arguments are in the docs, the plan is on MCP pricing, and eating in a surplus while training is covered in tracking a lean bulk. Building a product rather than training yourself? A REST plan.
Before you follow any programme
If you are new to lifting, returning from injury, pregnant, or managing a medical condition, have a qualified human review it. Persistent pain is a question for a physiotherapist, not a subscription.
What we have not measured
We have not tested the products in this category, compared their progression engines, or measured training outcomes. There are no rankings, no scores and no effectiveness claims here, because we have not earned them and we are not a fitness company.
The six questions above are things you can ask any vendor yourself, which is more useful than a ranking that is stale within a quarter.
Related
Frequently Asked Questions
What does an AI fitness coach actually do?
Three separable jobs. Generating a programme, which is largely solved. Tracking performance and adjusting the next session, which is bookkeeping done well. And coaching technique, which no product can do, because none can see how you move well enough to correct you.
Are AI fitness coaches worth paying for?
The part worth paying for is a progression engine that holds your history and adjusts from what you actually lifted, since that is the thing humans do inconsistently. Programme generation alone is available free from any chat assistant.
What should I check before subscribing?
Whether it holds your history, whether you can read the progression rule in plain language, whether you can export your data, what it does when you report an injury, whether the nutrition side uses a real food database, and whether it ever tells you to do less.
Can an AI coach check my form?
Camera-based form checkers using pose estimation can give gross feedback such as squat depth or bar path in good lighting at the right angle. That is not equivalent to a coach watching you, and technique under load remains the largest injury risk in a gym.
Why is the nutrition side usually the weakest part?
Because a fitness product's core competence is training, so food data tends to be bolted on. Ask whether macro figures come from a database or are generated. Generated numbers carry error that compounds across a training block, and it skews in one direction since under-counted items are oils, dressings and unlogged snacks.