What is an AI cycling coach?

The two layers, and why the split matters

Almost every product in this category is two things wearing one name.

The engine. A deterministic calculation layer that takes your ride files and your recovery signals and produces the standard measures: training load, an acute and chronic balance, power zones, an FTP estimate, an aerobic decoupling figure. Given the same inputs it produces the same outputs, every time. This part is arithmetic and sport science, and it has been around far longer than the current wave of AI.

The language layer. A model that reads those numbers, your history and your question, and answers in sentences. This is the part that is new, and it is the part the marketing is usually about.

The reason to care is narrow and practical. A language model is extremely good at explaining a number and structurally unsuited to producing one: it generates the most plausible continuation, not the correct result. Ask it to compute your weekly load and it will return something in the right range, formatted convincingly, and occasionally wrong in a way you cannot see. Ask it to explain a load your engine computed, and it is doing the thing it is actually good at.

So the first question to ask any AI cycling coach is not what model it uses. It is: which of your numbers are calculated, and which are generated?

What data one actually needs

An AI cycling coach with no data is a chatbot with a bicycle theme. What separates the two is what it can read about you.

• Power, from every ride. The only load signal that does not move with heat, stress or caffeine. It is what makes training load comparable week to week.
• Heart rate. Useful for what power cannot see — how hard a given effort actually cost you today. It lags behind effort and drifts on long rides, which is a feature if you are reading it as a cost signal and a trap if you are reading it as an intensity target.
• Resting heart rate, heart-rate variability and sleep. These describe how you are coping with the load, not how fit you are. They are only informative against your own recent baseline, measured the same way at the same time of day.
• What you tell it. How the session felt, what your week looks like, that you are travelling on Thursday. This is data too, and it is the part no sensor supplies.

A coach that only reads completed rides can tell you what happened. A coach that also reads recovery and your own account of the week can tell you what to do next. Those are different products.

How adaptation should work — and how much it actually buys you

Every product in this category promises a plan that adapts. It is worth knowing what the evidence says about that promise, because it cuts both ways.

The encouraging half. Thirty recreational runners were split between a predefined plan and one where the load was raised, held or lowered twice a week based on nocturnal heart-rate variability, perceived recovery and a heart-rate-to-speed index. Over fifteen weeks, 10-kilometre time improved by 6.2% ± 2.8 in the adjusted group against 2.9% ± 2.4 in the fixed group (p = 0.002). In that 10-kilometre test, 81% of the adjusted group were high responders against 23% of the fixed group — and while 8% of the fixed group finished moderately worse than they started, none of the adjusted group did (Med Sci Sports Exerc, 2022).

The sobering half. A systematic review with meta-analysis pooled the trials comparing heart-rate-variability-guided training against predefined training. Guided training was clearly better at improving the vagal HRV measures themselves (SMD 0.50, 95% CI 0.09–0.91) — but for the outcomes you care about, the advantages were small and not statistically significant: maximal aerobic capacity SMD 0.20, capacity at the second ventilatory threshold 0.26, endurance performance 0.20. The authors' own summary is the sentence to keep: if guided training beats predefined training at group level, current data suggest it does so only by a small margin — though with less likelihood of negative responses (Int J Environ Res Public Health, 2021).

Read together, that is a modest and useful claim: adapting to your signals is unlikely to transform your season, and it is a reasonable way to avoid the weeks that wreck one. Anyone selling you more than that is selling you more than the evidence.

🔴 The limits on both: the first is thirty runners, not cyclists, in a single study; the second is a review whose authors note that the basics — which HRV index, what measurement position, fixed or rolling baseline — are still unsettled.

The numbers are softer than the interface suggests

An AI cycling coach will show you an FTP figure to the watt. It is worth knowing how firm that anchor really is, because everything else is calculated from it.

A laboratory study compared FTP — taken as 95% of a 20-minute time trial — against critical power in 17 moderately trained cyclists and triathletes. On average the two agreed: a mean difference of 7 ± 13 W and a very strong correlation (r = 0.969). But the limits of agreement ran from −19 to +33 W, and the typical error was 13 W, or 5.6% — which the authors note sits above the 5% usually accepted in sport science. Their own wording is careful but pointed: arguably, they write, these findings question the underlying physiology of the FTP concept (Front Physiol, 2020).

Where you test moves it further. Nine junior road cyclists did a full laboratory series and then a 20-minute time trial on an uphill road: outdoor FTP came out 14–15% above both laboratory thresholds, and the outdoor 20-minute power was 12% higher than the indoor power–duration relationship predicted (J Strength Cond Res, 2023). 🔴 Nine male juniors on one climb — the direction transfers, the percentage does not.

None of that makes FTP useless. The point of an anchor is to be stable enough to train from, and it is. It does mean that a coach which presents a computed number as a measurement is misleading you about its own confidence — and that a good one says so.

What the training model decides, and what it does not

Polarized, pyramidal, sweet-spot-first: the argument about intensity distribution is loud, and the evidence for it is quieter than the argument.

A 2024 systematic review with meta-analysis pooled seventeen studies and 437 participants. Polarized training came out ahead for peak oxygen uptake, but only slightly (SMD 0.24) and only in interventions shorter than twelve weeks and in already highly trained athletes. For the outcomes closer to what a rider cares about, the distributions were indistinguishable: time-trial performance SMD −0.01, and power or speed at the second threshold — the closest laboratory stand-in for FTP — SMD 0.04, 95% CI −0.21 to 0.29 (Sports Med, 2024).

The same pattern shows up in the interval question. Twenty-two well-trained cyclists completed both a one-week block of moderate-intensity intervals and a one-week block of high-intensity intervals. Both improved 15-minute maximal power — 4.9% and 2.8% — with no significant difference between them (p = 0.44); the moderate block was better for power at 4 mmol lactate, 4.5% against 2.1% (p = 0.03) (Eur J Sport Sci, 2025).

🔴 And look at the spread on that last one: the 4.9% carried a standard deviation of 8.7 percentage points, nearly twice the average gain. A group mean promises an individual nothing.

The reading for anyone evaluating an AI coach: be suspicious of a product whose main claim is its training philosophy. The distributions are close enough that consistency, recovery and whether your easy days are actually easy will decide far more than the label on the model.

A plan is not the training

The most under-discussed limit has nothing to do with algorithms. It is the gap between what gets prescribed and what gets done.

Researchers followed two coaches and their junior speed skaters through a four-week block, recording what the coaches intended session by session and what the athletes actually did: 438 intended sessions against 378 executed. Intended training time was 52 hours 37 minutes; what got done was 45 hours 16 minutes — roughly four sessions fewer. Strength sessions slipped most, and individual variation was large, with some athletes exceeding the plan and others well under it (Int J Environ Res Public Health, 2022).

There was a human coach standing right there, and the gap opened anyway — differently for each athlete. 🔴 The limits: 14 junior speed skaters, one four-week block, self-reported logs, observational.

The consequence for this category is direct. The value of an AI cycling coach is not in generating a better plan. It is in noticing, quickly and specifically, when the plan and your week have stopped matching — and adjusting rather than letting you fall behind a schedule you were never going to keep.

The limits, including ours

An honest guide has to say what this software cannot do.

• It cannot see your life. It knows your job got harder only if you tell it. A human coach who has known you three seasons knows before you say anything.
• It cannot make you accountable. A person who notices you skipped Thursday changes behaviour in a way no notification does. On cycling forums, riders paying for coaching say repeatedly that this is what they are paying for.
• It is not a medical instrument. Readiness signals describe how you are coping with training load. They say nothing about health, and a pattern that persists or comes with symptoms beyond training fatigue is a conversation for a qualified professional, not an app.
• It inherits the softness of its inputs. Everything in the section above about FTP applies to every product in this category, ours included.
• A conversational coach is not rare any more. At least six live rivals ship one. Anybody telling you that talking to your training software is the differentiator is describing 2023.

Seven questions worth asking any AI cycling coach

These are not written to make anyone win. They are the questions a rider can actually check, on any product, in an afternoon.

1. Which numbers are computed and which are generated? If a language model is producing your training load, that is the wrong architecture, whatever the marketing says.
2. Does it read your actual rides, or a summary? Full files or nothing.
3. Does recovery change tomorrow, or only get displayed? A readiness score that never moves a session is a dashboard widget.
4. Can you ask it why — and check the answer against your own ride file? An explanation you cannot verify is decoration.
5. Does it say what it does not know? Look for stated limits, confidence, and a willingness to say a number is an estimate.
6. Is the conversation metered? Counters and credit packs change how you use a coach — you stop asking the small questions, which are often the useful ones.
7. Can you get your data out? Ask before you need it, not after.

If you want the same questions applied to specific products with prices and dates, that is a different page: the best AI cycling coach app in 2026.