Is AI Coming for Cycling Coaching?
AI is here, we have ChatGPT, Claude, and a host of others. And I'll preface things by saying that yes, I use AI in my coaching, as do many coaches at the World Tour level. But here's the thing: how are these coaches using AI, is it to create training plans per se, or is it for something else?
On the face of it, you have a GPT or Claude a/c (paid or free) and you tell it that your FTP is this, your VO2max is that, you can ride for so many hours per day, and — if you like — you use an MCP or direct connection to your Strava a/c so that it can mine your existing data. Bam! Done, it pulls your data, knows your goals, and creates a training plan. Sort of thing you'd expect AI to be able to do well.
But here's the rub. AI is trained on whatever's publicly available, sound or not, so it tends to infer what you should be doing from the average of what's already been written. It has no idea whether that's actually correct, or just what the majority of people happen to do, it's essentially taking it as gospel.
On the other hand, as I understand it, a platform such as TrainerRoad uses machine learning (ML) to collate all the data points from the athletes on TR and uses that to see what drives gains, but again, while it may drive gains for the majority, it really depends where you are on the "edge case" and how close you are to a plateau. If you're miles away from it, then yes, it likely works; but as you start to approach it, then it doesn't work, and this is when a lot of athletes come to me or other coaches: "App X isn't making me improve any more, my FTP is stuck." As a coach, I then review their data, see what they've been doing, and work my way around ideas and concepts that help them improve and meet their goals. Often, I'll look at data from an athlete — especially outdoor data — and see that apps have completely misinterpreted what it means or how it should be interpreted. I've seen rides where someone has done an outdoor ride where they've climbed a hill multiple times, at the top turned round and descended, then had a few minutes' easy spin and then repeated several times, and the software has suggested that their FTP is significantly less than what they can do. This happens to me all the time: I'll climb a local hill at my FTP or just over, repeat several times, and the app will tell me my FTP is way lower than what I could actually do.
Coaching (at least good coaching) relies on knowledge gained over years. I've a 1st class Hons degree in Sport Science, peer-reviewed research, and I've been a coach for approaching three decades, coaching hundreds of athletes including elite an World Champion, Paralympians, and ultra riders at the highest level. I also developed the method of estimating roughly one-hour power from a MAP ramp test at specific ramp rates, the approach that now sits inside apps such as Zwift, TrainerRoad, Rouvy and MyWhoosh, so I've a fairly good idea of exactly where these algorithms earn their keep, and where a human still wins.
So, how do I use AI? I use AI to help with admin, correct my grammar (sorry, I'm pretty bad with this), help me think of content to write about, and — in terms of coaching — get it to help me analyse some data. Mostly, I've found it poor at analysing numerical data such as power data, mainly because what it's been trained on is in the public domain and simply not really correct. On the other hand, it's useful at assessing some numerical data when looking at trends (such as, say, HRV), and how long-term trends may suggest why people feel good or bad when power data says otherwise (for example, Jim slept badly for six days which finished 10 days ago, and his power has been marginally depressed since then), or looking for athlete comments I've missed, or how a comment may not mean anything much in the moment — but occasional 'off' comments may have a cumulative effect that a coach doesn't notice over a medium term (because you answer comments in the moment). Here, AI is genuinely useful. This is how it's being used at the World Tour level.
Research on how athletes should train, and how coaches should work, points the same way. A recent survey of 78 world-class endurance coaches (Sandbakk et al., International Journal of Sports Physiology and Performance) found the game has moved firmly towards individualisation, away from generic, one-size-fits-all plans. Those coaches were also openly wary of "black-box" algorithms and unvalidated metrics, concluding that the data still needs careful human interpretation to be worth anything. And the "something else"? The factors they rated as mattering most for whether an athlete actually adapts to a plan were the coach–athlete relationship, the balance of life stress, and the athlete's belief in the plan — not one of which an app or an AI can see!
So why are people using apps, etc., for coaching? Well, simply, it's likely down to cost, AI coaching models are dirt cheap in comparison to 1-to-1 coaching, by roughly a factor of 10. Realistically, though, you're trading cost for accuracy.
What would I do as an athlete? It depends on the situation. If I'd just started training and was at the beginning of my journey — yup, rock on with AI software; it'll be fine to dial you in at this stage, if that’s what you want to do. If I'd been training for a while, or the gains were slowing down, then I'd be looking at some sort of halfway house, such as our MAP: The Build where a coach is interpreting your data and building a plan for you, with ideas and algorithms that have been validated over almost three decades of 1-to-1 coaching, but this, too, has its limitations. And as soon as you become an edge case, or the gains aren't happening at the rate you want, or you really want to improve, or the goals themselves become edge-case uber-difficulties, that's where 1-to-1 coaching really shines through.
What would I do as a coach (which I am — although I'm also an athlete)? At present, neither AI nor ML systems can genuinely create training plans that are good. But they can, and do, help with some of the back-end administrative work that isn't particularly well handled in training software at present. Occasionally I'll ask GPT or Claude a question around edge-case scenarios; for example, one athlete I coach has a rare genetic condition, it’s so rare that only nine families worldwide have been identified with the faulty gene and I'll throw ideas at GPT or Claude to see if it can help with their training. Interestingly, Claude came up this morning with a concept I hadn't considered, though in general this is rare. So don't think I'm some sort of archaic coach who isn't moving with the times (albeit I'm sure I'm archaic!) but right now, and for a fair while yet, AI isn't winning at coaching plans, unless your current training is just riding the bike with no structure whatsoever.