It's Tuesday. Your plan says 8x800m at 5k pace.

Here is what your watch already knows. You slept five hours. Your HRV is 15% below your seven-day average. Last Thursday's tempo ran 8% higher on cardiac drift than it should have. And it will be 31 degrees by the time you get out of the door.

The plan still says 8x800m at 5k pace. It said that in January, when it was written.

Adaptive coaching says something else: move the session to Thursday, when your calendar is clear and the forecast drops nine degrees. Run forty minutes easy today instead. If your HRV recovers overnight, you'll do the full set. If it doesn't, you'll cut to six reps and redistribute the load across the weekend.

Same runner, same data, two different weeks. One ends with a completed session and a recovered athlete. The other ends with a session you survived, followed by three days of heaviness you can't quite explain.

That gap is what this article is about.

What actually changes about your week

Forget the technology for a moment. Here is what adaptive coaching does differently, in the situations that actually come up.

The session that shouldn't happen today. The example above. Poor sleep, suppressed HRV, accumulated fatigue from a session that cost more than it should have. A plan cannot see any of it. A good coach would move the session. An adaptive system moves it and tells you why.

The session that should be harder. This is the half nobody mentions, and it matters more than the backing off. Sometimes you arrive at a Tuesday fresher than the plan assumed. You tapered into a quiet work week, you slept nine hours, your HRV is sitting above baseline. A static plan gives you the same session it was always going to give you. Adaptive coaching should recognise the window and use it, because those weeks are where fitness actually gets built and most runners waste them running to a number set months ago.

The week you missed. You get ill on Monday and lose four days. A plan carries on regardless, so week seven arrives whether or not you did week six. The classic failure is a plan that keeps progressing volume through a week you spent in bed, then hands you a long run your legs have no business attempting. Adaptive coaching reads the gap and rebuilds the ramp, which is the practical application of training load management.

The conditions. Your plan was written for a temperate Tuesday. It is now 31 degrees, or icy, or you are three timezones from home. Effort targets should move. Pace targets should move with them.

The niggle. You mention your Achilles has been grumbling. A PDF has no field for that. A system you can actually talk to can factor it in, reduce the eccentric load for a fortnight, and flag it if the pattern repeats.

None of this is exotic. It is what a good coach does every week. The difference is that 95% of endurance athletes don't have one.

So what counts as AI coaching?

The term has been stretched to cover three very different things. Worth knowing which you're looking at.

A chatbot bolted onto a plan. You get a pre-built schedule and an assistant that answers questions about it. The AI does not change your training. This is a fine product, but calling it coaching is like calling cruise control self-driving.

A rule-based adaptive plan. Your schedule shifts according to predetermined rules. Miss a session, it redistributes. Improve on a test, it bumps your targets. Genuinely adaptive, but mechanically so. It knows you missed Tuesday. It does not know whether you were ill, travelling, or simply could not face it, and those three should produce different responses.

A system that reasons. It looks at HRV trending down across four days, cross-references a 15% volume increase and a transatlantic flight on Tuesday, and concludes that today's intervals should become an easy spin. Not because a rule fires at a threshold, but because it weighs the relationship between sleep disruption, accumulated fatigue and injury risk the way a coach would.

The distinction that matters: rule-based systems are brittle. They handle the scenarios their designers anticipated and fail at everything else. Your training week is mostly made of everything else.

What it takes to build one properly

Three ingredients, and most products have one or two.

Physiological guardrails. Large parts of training science are settled and should not be left to a language model's judgement. Progressive overload. Acute-to-chronic workload ratios. Taper protocols. These belong hard-coded. A system that breaches sensible progression because its model had a creative idea, or because you argued for a harder session, is a system that will injure people.

A model of you specifically. How fast you recover, how heat affects you, which sessions historically precede your niggles. This is the part that earns the word personalised: not picking from three difficulty levels, but building a longitudinal record of one athlete. We call ours the Unified Athlete Record.

Reasoning that can be inspected. The newest ingredient. A language model can take the guardrails, your history and the mess of your current week, and synthesise a decision. Crucially it can explain that decision in plain language, which is the only way you can tell whether to trust it. We call this the glass box, and it's the feature we'd defend hardest. A recommendation you can't interrogate is just a number with better marketing.

Where it falls down

It cannot fix bad inputs. If your optical heart rate sensor is reading badly, or you skip the check-ins that tell the system how you actually feel, the coaching degrades accordingly. The advantage is that a decent system will at least flag the inconsistency, which is something most human coaches have no time to do across a full roster.

And the technology is young. Language models are powerful and can be confidently wrong. The systems worth trusting are the ones that are conservative about injury risk and honest about their limits. We have spent a long time on this and are not finished.

On human coaches, the usual line here is that AI will never replace a good one. We won't write that, because we are not sure it's true. Perhaps a fifth of coaches are excellent and worth every penny. The rest are issuing near-identical plans, checking in sporadically, and charging between $150 and $500 a month for a PDF and a WhatsApp thread. They are not reading your HRV before prescribing tomorrow. A well-built system can beat that, not because the technology is magic, but because the median is lower than the industry admits.

How to evaluate one

Four questions, whoever you're assessing, including us.

1. Does it change my actual training based on my data, or only on my answers to a sign-up quiz?

2. Can it explain why it made a given recommendation, in language I can check?

3. Does it read the devices I already own, without me exporting anything?

4. Does it have guardrails that stop it doing something reckless with my health?

If you want the longer version of question one, we wrote a whole piece on adaptive versus static plans. If you want to understand the single most useful signal these systems read, start with HRV-guided training.

Key Takeaways

• A training plan is written once, months before you run it. Adaptive coaching decides what today's session should be, using what your body is actually doing this morning.

• The value is not only in backing off. A system that never tells you to go harder on a good week is wasting your best training days.

• Three things get sold as AI coaching: a chatbot beside a fixed plan, a rule-based scheduler, and a system that genuinely reasons. Only the third changes what you do and why.

• Rule-based adaptation is brittle. It handles the scenarios its designers imagined, and most of your training year is made of the ones they didn't.

• Building it properly needs all three of hard-coded physiological guardrails, a longitudinal model of you, and reasoning you can inspect in plain language.

• It cannot fix bad data or an athlete who skips the check-ins, and any system claiming otherwise is overselling.

• Ask four questions: does it change my training, can it explain itself, does it read my devices, and does it have guardrails.

A coach is not a plan, and it is not a dashboard. A coach notices, thinks, and responds. That is now buildable. The question is who builds it carefully.

Zepho is adaptive coaching for serious runners. It's live and available now on the App Store.