There’s a particular kind of sadness reserved for the runner who gets injured three weeks before their goal race. You’ve done the long runs. You’ve eaten the overnight oats. You’ve told everyone at work about your target time. And then your Achilles tendon, which has been quietly grumbling for the last six weeks, stages a full walkout.

Most of the time, this isn’t bad luck. It’s a training error. The plan said 18 miles on Saturday, so you ran 18 miles on Saturday, even though you’d slept four hours, your resting heart rate was 15 beats above baseline, and you’d been on your feet all day at a conference. The plan didn’t know any of that. The plan doesn’t know anything about you. The plan is a PDF.

This is the gap that AI coaching is trying to close. Not by replacing human coaches with chatbots, and not by sprinkling the word “personalised” on a static training schedule. But by building something that actually responds to what’s happening in your body, your life, and your training. In real time.

The trouble is, the term “AI coaching” has already been claimed by so many products doing so many different things that it’s become nearly meaningless. So let’s be specific about what it actually is, what it isn’t, and why it matters.

The problem with plans

The conventional training plan is one of sport’s great inventions. Lydiard’s base-building philosophy, Daniels’ VDOT system, Marius Bakken’s Norwegian Method (a personal favourite with at least 2 marathons trained for this way); these are brilliant pieces of coaching distilled into a format that millions of runners can follow. They work. And for a lot of people, they work really well.

But they share a fundamental assumption: that your life will cooperate. That you’ll sleep well, stay healthy, avoid stress, and hit every session at the prescribed effort. That the conditions on the day of your tempo run will be roughly what the plan-writer imagined when they designed it.

Reality, of course, has other ideas. Your toddler wakes up at 3am. You catch a cold. Work sends you to a different timezone. It’s 34°C and your easy pace feels like running through treacle. You’re fatigued in a way you can’t quite articulate but your legs feel like they’re made of wet cement.

A plan doesn’t adapt to any of this. A good human coach does; they’ll check in, read between the lines of your Training Peaks upload, and adjust. But 95% of endurance athletes don’t have a coach. And even those who do are limited by the coach’s availability, the information they can realistically process, and the number of athletes they can manage at once.

This is the actual problem AI coaching is trying to solve. Not “runners need technology” — runners have plenty of technology. It’s that the gap between having data and using it intelligently remains enormous.

What “AI coaching” actually means (and what it doesn’t)

If you search for AI coaching apps today, you’ll find roughly three categories of product masquerading as one.

The first is the static plan with a chatbot bolted on. You get a pre-built training schedule, often a good one, and alongside it, an AI assistant that can answer questions about your plan or offer encouragement. The AI doesn’t change your training. It’s a knowledgeable companion, not a coach. This is a perfectly fine product, but calling it AI coaching is a bit like calling cruise control self-driving.

The second is the algorithm-adjusted plan. Your training schedule shifts based on rules: if you miss a session, it redistributes the load. If your fitness test improves, it bumps your targets. If you’re in a recovery week, it dials things back. This is genuinely adaptive, but in a mechanical way. The rules are predetermined, the adjustments are formulaic, and the system doesn’t really understand why you missed that session; whether you were injured, hungover, or just couldn’t be bothered.

The third is what’s now emerging: genuinely intelligent coaching. This is where things get interesting. Modern AI systems (specifically large language models) can reason about your training in context. They can look at your HRV trending downward over four days, cross-reference it with the fact that you’ve increased volume by 15% and flew transatlantic on Tuesday, and conclude that today’s interval session should become an easy spin. Not because a rule says “if HRV < X then reduce intensity”, but because the system understands the relationship between sleep disruption, accumulated fatigue, and injury risk.

The difference matters. Rule-based systems are brittle. They work well for the scenarios their designers anticipated and poorly for everything else. AI coaching, in its more ambitious form, aims to reason like a good coach would; considering the whole picture, weighing competing signals, and making a judgment call.

Where the data comes from

For any of this to work, the system needs information. The good news is that us endurance athletes are, somewhat obsessively, already providing it.

Your GPS watch tracks pace, distance, elevation, cadence, and heart rate. Most watches also measure HRV, skin temperature, blood oxygen, and sleep stages. Chest straps give cleaner heart rate data. Power meters on bikes measure output directly. Even your phone logs steps, movement patterns, and screen time (a surprisingly useful proxy for recovery quality). And a cornucopia of devices provide other health signals. 

The interesting shift is what’s happened to all this data over the past two years. The major platforms have opened their APIs. What was once locked inside each manufacturer’s proprietary app can now flow into a unified system that sees everything. Your sleep from Whoop, your run from Garmin, your bike session from Zwift, your subjective energy from a morning check-in — all in one place.

This is the raw material. The question is what you do with it.

The three ingredients

Useful AI coaching requires three things working together, and most current products have one or two of them but not all three.

Physiological rules. There’s a body of training science that’s well-established and shouldn’t be left to AI judgment. Progressive overload principles. Acute-to-chronic workload ratios. The relationship between aerobic base and race performance. Taper protocols. These are deterministic; they should be hard-coded, not guessed at. A system that violates the 10% weekly mileage increase rule because its language model had a creative idea, or gave in to your pleading for a harder run, is a system that will hurt people.

Machine learning on your data. Your individual response to training is unique. How quickly you recover, your injury patterns, how altitude affects your performance, when in the day you run best; these can be learned from your historical data over time. This is the “personalised” bit that actually deserves the name: not selecting from three difficulty levels, but building a longitudinal model of you. At Zepho we call it the “Unified Athlete Record”. 

Language model reasoning. This is the newest ingredient, and the one that makes everything else sing. An LLM can take the physiological rules, your personal data model, and the messy reality of your current situation, expressed in natural language if you want, and synthesise a coaching decision. It can explain its reasoning (we call it our “glass box” approach). It can have a conversation with you about why today’s plan changed. It can understand that “I’m absolutely knackered and my knee feels weird” is important information even though it doesn’t map neatly to a data field.

Get all three right and you have something genuinely new: a coach that knows the science, knows you, and can think on its feet.

What this looks like in practice

Abstract descriptions of technology are boring. Let’s make it concrete.

Scenario: You’re eight weeks out from a marathon. Your plan calls for a 22km long run at 5:15/km pace. But:

• You slept 5 hours last night (your Garmin knows).

• Your HRV is 15% below your 7-day average.

• This week’s tempo run was harder than it should have been — your cardiac drift was 8% above normal.

• The weather forecast says 31°C and 70% humidity.

A static plan says: 22km at 5:15. Good luck.

A rule-based adaptive plan might say: HRV is low, reduce distance to 18km.

An AI coach looks at the full picture and might say: your acute fatigue is elevated from yesterday’s session, your sleep was poor, and it’s going to be hot. Let’s shift this long run to tomorrow, your schedule shows it’s clear, and do an easy 40-minute Zwift shakeout today instead. If your HRV recovers overnight, we’ll do the full 22km but at 5:30 pace to account for the heat. If it doesn’t, we’ll drop to 18km and redistribute the training stress later in the week.

That’s a fundamentally different kind of response. It’s not just adjusting a number; it’s reasoning about the interaction between multiple factors and making a plan that accounts for what’s likely to happen next.

But what about human coaches?

This is usually the part of the article where someone writes “of course, AI will never replace a good human coach.” We’re not going to do that, because we’re not sure it’s true.

Here’s the uncomfortable reality of human coaching: maybe 20% of coaches are excellent. They bring deep experience, genuine intuition, and the kind of emotional intelligence that makes athletes better. They’re worth every penny. But the other 80%? They’re writing the same generic plans they give to everyone (hello Instagram coach…), checking in sporadically, and charging $150–500 a month for what amounts to a PDF with a WhatsApp group. They don’t process your wearable data in any systematic way. They don’t adjust your plan daily. They certainly don’t know your HRV trend when they prescribe tomorrow’s session. A well-built AI system can genuinely do a better job than the majority of human coaches; not because AI is magic, but because most coaching is mediocre and the bar is lower than the industry likes to admit.

That said, AI coaching isn’t magic either. It can’t fix bad inputs. If your watch gives inaccurate heart rate data (optical sensors on hairy wrists, we see you), or if you consistently ignore the check-ins that help the system understand how you’re feeling, the coaching quality suffers. The difference is that a good AI system can at least detect inconsistencies in your data automatically; something most human coaches simply don’t have time to do across dozens of athletes.

And there’s a maturity question. The technology is new. LLMs are powerful but they can be confidently wrong. Any serious AI coaching system needs guardrails, those physiological rules we mentioned, to prevent the AI from making a creative suggestion that’s also a terrible one. The systems that will earn trust are the ones that are transparent about their limitations and conservative about injury risk. We’ve spent countless hours trying to perfect this, we know we still have work to do. 

Why it matters now

Three things have converged in the past eighteen months to make real AI coaching possible, after years of it being mostly marketing.

First, the AI models got good enough. GPT-4, Claude, and their successors can reason about multi-variable problems in a way that previous generation models simply couldn’t. Telling an AI to “consider the athlete’s HRV trend in the context of their training load and upcoming race” and getting a sensible response is a 2026+ capability, not a 2022 one.

Second, the data infrastructure matured. Open APIs from wearable manufacturers, standardised fitness data formats (FIT, TCX), and cloud platforms that can process real-time data streams mean that building a unified view of an athlete is now an engineering problem, not a moonshot.

Third, the audience is ready. The explosion of endurance events — HYROX went from niche to mainstream in three years, marathon participation hit record levels in 2025, ultra running and gravel cycling are booming — has created a massive population of athletes who are serious about performance, own expensive wearables, and are frustrated that their training technology hasn’t kept pace with their ambition.

These athletes don’t want another training plan. They want something that understands them.

Where this is heading

The next few years will separate the marketing from the real thing. Some apps will continue to put “AI” on what is essentially a decision tree. Others will build something that genuinely earns the word “coach.”

The ones that get it right will share a few characteristics: deep respect for exercise science (not trying to reinvent physiology with machine learning), genuine personalisation built on longitudinal data (not a questionnaire), transparency about how decisions are made (not a black box), and humility about what AI can and can’t do.

If you’re an athlete evaluating these tools, the questions to ask are simple. Does it actually change my training based on my data, or just my answers to a quiz? Can it explain why it’s making a recommendation? Does it integrate with the devices I already own? And does it have guardrails that prevent it from doing something stupid with my health?

The bar for calling something a coach should be high. A coach isn’t a plan. A coach isn’t a dashboard. A coach listens, thinks, and responds. The technology to do that properly is finally here. The question is who builds it with the care it deserves.