Most training plans are written for a fictional cyclist — the average one. The one who recovers in exactly 48 hours, sleeps eight solid hours every night, and never has a stressful Tuesday at work. But that cyclist doesn't exist. Understanding how an AI cycling coach works is really about understanding what a coaching system finally does when it stops ignoring reality.
The core mechanism: inputs, outputs, and the feedback loop
At its simplest, an AI cycling coach takes in structured data about your training and physiology, processes it against models of how athletes adapt, and spits out a modified plan. That's the skeleton. The flesh is what makes it interesting. A well-built AI system is constantly comparing what it expected to happen with what actually happened. You were supposed to hit 280 watts for 20 minutes; you managed 263 and your heart rate ran 8 beats higher than predicted. That gap matters. A static plan shrugs. A good AI system registers it as signal — your fatigue is higher than the model predicted, and something needs to give in the next block.
The engine underneath this is typically a combination of performance modelling and machine learning. Performance models — often based on Bannister's impulse-response framework or its modern derivatives — simulate how training load accumulates into fitness and fatigue. Machine learning sits on top, learning your individual response coefficients from historical data rather than applying generic population averages. A 2020 study in Applied Sciences (MDPI) showed that a reinforcement learning-based virtual coach could produce training planification on par with human coaches after iterating through a cyclist's actual training history. The implications are significant: the AI's estimates get more accurate the more data it has from you specifically.
Wearables have made this feedback loop far richer. Heart rate variability, resting heart rate, and sleep data from consumer devices give the system daily readiness signals that go well beyond what power files alone can show. Research published in Scientific Reports (2025) found that combining HRV with well-being scores produced better training outcomes than using HRV guidance alone — which is exactly the kind of multi-signal interpretation an AI system can handle at scale, and a human coach would struggle to compute consistently for dozens of athletes.
What this actually looks like in practice
Let's be honest: most riders imagine AI coaching as some exotic black box that spits out a perfect plan from the void. The practical reality is more grounded, and more useful. You connect your training data — Garmin, Wahoo, Strava, whatever you're using — and the system builds a baseline model of your fitness. It estimates your FTP equivalent, your optimal training load range, your acute-to-chronic ratio. Then it starts planning workouts forward. When you complete a session, it re-evaluates. Miss a session? The week gets re-shuffled. Nail an interval session with power numbers higher than expected? The model nudges up its estimate of your current form and may schedule something progressively harder earlier than planned.
The AI coaching concept works best for time-crunched riders who need every session to count. Generic periodization plans are designed with buffer — they assume you'll miss sessions, they assume average recovery. When you have four hours a week instead of twelve, you can't afford average assumptions. An AI system that knows you specifically — your fatigue response, your preferred workout days, your historical performance in different conditions — can squeeze more adaptation out of limited training hours than any off-the-shelf plan.
Common mistakes riders make when starting with an AI coach: treating the first two weeks of data as reliable. They're not. The model is calibrating. If you go full gas in week one to "show the AI what you can do," you'll skew its baseline high and likely get prescribed workouts that tip over into overreaching. The better approach is to train normally, honestly, for the first few cycles. Feed it true data, not peak data.
When it matters — and when it doesn't
An AI coach earns its keep when your training has structure and regularity. If you're riding 3–5 times a week with at least some kind of power or heart rate data, the feedback loop works. The system has something to learn from. If you're riding twice a week casually with no tracking, you'll get a plan, but the adaptation engine has very little to work with — and you might as well be using a generic plan.
It also matters most during base and build phases, where accumulating load intelligently is the whole game. During a peak or taper phase, the margin for error shrinks and the gains are marginal — a good human coach who knows you well still has an edge in those final weeks before an A event, because they can interpret non-quantifiable signals: stress, motivation, how you carry yourself on a call. The question of whether AI can build a genuinely good cycling plan is closely linked to this: it can, for the bulk of your training year, with impressive consistency. But it's a tool, not a replacement for thinking about your own training.
The practical upside for the serious amateur is access to individualised coaching logic that, five years ago, required hiring a coach. The AI doesn't replace the relationship, the experience, or the judgement — but it does make adaptive, data-driven training available at a price and scale that changes what's possible for riders who train alone.
Related reads
Explore LeCoach: your AI cycling coach app
Can AI build a real cycling plan?
Sources
Reche-Soto et al. (2021). Towards an AI-Based Tailored Training Planning for Road Cyclists: A Case Study. Applied Sciences, 11(1), 313. mdpi.com
López-Valenciano et al. (2025). Individual training prescribed by heart rate variability, heart rate and well-being scores in experienced cyclists. Scientific Reports. nature.com