AI cycling coach or a rules engine: what decides your training plan.

On this page

1. What "AI coach" means
2. In their own words
3. Four questions to ask
4. Where Summit sits
5. When another app fits better
6. Questions
7. Sources

An "AI cycling coach" can be a chat you talk to, a model trained on other riders' rides, or a model of your own data that sets the load. Those are different things, and they decide your training in different ways. Summit works another way: it is a rules engine, deterministic, in which no model decides anything, and every change to your plan traces back to a rule applied to your own data.

Which one you want depends on how you want to be coached, and this page is meant to help you choose either way — including when the answer is not Summit. Every competitor fact below is quoted from that company's own pages, with the date we read it and a link to the page.

What "AI coach" means, app by app.

Read the pages of the apps that use the word and three kinds of machine come out. Most apps mix them.

A coach you chat with

You type, a language model answers, and in some apps it edits the plan when you ask. RestOrTrain says it "uses LLMs to solve" training, and its homepage shows the exchange: type that you need a day off and it replies "I've cleared today's schedule." Enduco's "AI Coach Chat" takes 10 messages a week on Standard and 100 on Pro, and you can ask it questions and get plan adjustments back. Athletica's conversational AI Coach reads your rides, power trends and HRV and explains them in plain language. What you get is a conversation, and the plan moves when the conversation says so.

A model trained on many riders' rides

TrainerRoad describes "custom AI models built from the ground up on millions of rides from real cyclists". Its Adaptive Training page says it predicts how hard each workout will feel from your history, performance and recovery, and the plan adjusts to how your workouts went and to a short survey after each one, as Progression Levels.

A model of you that sets the load

Xert turns the power of every ride into a Fitness Signature — "No separate test." — and its programs "adapt and evolve with you as you train"; its Forecast AI maps how your fitness should develop towards a target event date. HumanGO's Hugo, "an AI-powered virtual coach", works from your availability, physiological progress, health metrics and feedback to adapt the plan.

And the mixes: Athletica's coach explains while its plan adapts; JOIN, which calls itself an "AI Cycling Coach", adjusts "based on the workout data, your RPE and your daily readiness score and your availability", with a daily check-in at the centre.

How each app describes itself.

The quotes are each company's own wording, as it appears on our comparison pages, read on the date shown; the middle column is our reading of it, not theirs. Each vendor's pages are the authority on its product, so if a line here decides your choice, check it there.

Four questions to ask any app, AI or not.

The label tells you little. These four tell you what you are buying — and they work just as well on Summit.

Does it decide the session, or only talk about it?

A coach that explains your data is useful; a coach that changes tomorrow's session is a different product. Athletica shows why the question matters: its cycling page says its AI interprets HRV "to adjust training load and recovery", its Coaching Professor post says the AI-Coach "does not change your training load (yet)", and its FAQ says the AI Coach does not automatically modify your training. If that decides your choice, ask Athletica. TrainerRoad's fatigue detection marks yellow and red days from accumulated training stress, and you can decline the change.

Summit: on Elite, the engine decides. Today's session stands, extends or loses intensity, down to soft activation or rest, with nothing to approve. On Free and Pro you get the recovery score; the session that moves with it is Elite.

Can you trace why a session changed?

When a session shrinks you should be able to find out why: which input moved, and what turned it into the change. Ask each app what it shows you when the plan moves, and whether the same inputs would give the same answer tomorrow.

Summit: every decision traces back to a specific rule or calculation applied to your own data, not a black box, and the same inputs produce the same decision every time. The models behind the rules are published ones (below); how they are combined and weighted is ours and is not published — auditable, not replicable.

Does it read your recovery overnight, or ask you?

Some plans move on what you report — an RPE, a survey after the workout, a morning check-in. JOIN is explicit that wearables "should complement your own feeling, not replace it" (its Readiness article). Others read a wearable. Neither is wrong: one trusts your judgement, the other the measurement.

Summit: it reads HRV, resting heart rate and sleep overnight from Apple Health, or from intervals.icu if you train with a Garmin, a Whoop or another device synced there, and scores each signal against your own baseline. There is nothing to fill in each morning. Any wearable that writes to Apple Health works, though only the Apple Watch writes the complete set of metrics Summit uses.

What happens when the data is missing?

A watch left on the charger, a history only a week long, a sync that failed. An app that reads recovery has to do something with a missing night: reuse yesterday's, ask you, or hold the plan. Ask which.

Summit: with a short history the score stays deliberately conservative; with no fresh data it shows no score rather than yesterday's; and with no reading at all, the plan holds exactly as written.

Where Summit sits: a rules engine.

Summit is a rules engine, not a chat. Its engine is deterministic and no model decides anything: fitness, fatigue, form, your readiness and your session are computed by following codified sports-science rules, and each change to the plan traces back to one of them, applied to your own rides and your own nights. Here is the difference with a chat, morning by morning:

The plan works backwards from your event: the phases between now and the date, then one weekly load budget, then the sessions that fit inside it, inside the minutes you have each weekday. Miss a session, flag an illness or change your week and what comes next is recalculated. How the adaptation works, step by step →

The models are the published ones. Load follows the Performance Manager model of Allen, Coggan & McGregor; taper length follows Bosquet et al. and Mujika & Padilla; peak length follows Issurin; intensity distribution follows Seiler & Kjerland. Summit ships that bibliography inside the app, every reference with its DOI, reachable from the screens that use it. What stays ours is how those models are combined and weighted; that lives in the engine and is not published.

What is free and what is not. The recovery score, sleep score and analysis, recovery trend and widgets are free, permanently — not a trial. Pro adds training load (fitness, fatigue, form) and ride analysis. The adaptive session and the plan that moves with your recovery are Elite. Prices are set by Apple per storefront; the plans page has them with the caveat in full.

What a rules engine costs you. You cannot talk it round. Beyond flagging an illness or injury, which stops the day, there is no way to tell Summit you feel worse than your numbers say, and no chat to ask why the plan moved. It is also cycling-only, iPhone-only and needs iOS 26 or later.

When another kind of app is the better fit.

Plenty of riders are better served by one of the apps above, and you should know that now.

For the whole field, with what actually adapts in each app and to what: Best adaptive cycling training apps in 2026.

Questions people actually ask