How AI Training Plans Actually Adapt to Your Runs

We build one of these, so treat this as a look under our own hood rather than a neutral survey. But "adaptive" is the most abused word in training software - every app claims it, almost none say what it means - so a concrete answer is worth more than another brochure.

Here is what actually changes, when, and who decides.

The Two Kinds of Change, and Why They Are Different

Not all adaptation is the same, and conflating the two is how apps oversell.

Small, in-week adjustments. You tell the coach "make this week easier" or "move my long run to Sunday", and a few upcoming sessions shift. Low stakes, instant, reversible.

A full rebuild. Your race date moved, your goal changed, or you fell far enough behind that the block no longer makes sense. That is not a tweak - it regenerates the whole plan from your new reality.

Most apps blur these together. We keep them separate on purpose, because the safe way to handle each is different - and the boring one, the in-week nudge, is what you will use most.

What an In-Week Change Actually Does

When you ask for an easier week, the change is bounded and specific: upcoming, not-yet-done sessions this week get their volume scaled to 85% - both distance and duration. A harder week scales to 112%. Rest days are left alone, and every number is clamped so nothing can collapse to zero: a run never drops below 1 km or 5 minutes however many times you ease it.

Moving a session is a same-week date swap: your Thursday tempo and your Saturday easy run trade dates, so the week stays full and correctly dated rather than developing a hole. Completed sessions are never touched - the app cannot rewrite history you have already run.

Notice what is *not* here. It does not silently rewrite next month. It does not change your goal because you had one bad week. The blast radius of a casual "ease up" is exactly this week's remaining runs, and nothing else.

Who Decides: the Model vs the Rules

This is the part that separates real adaptation from a chatbot with a database.

The language model does one job: it reads your sentence and decides *what you meant*. "This week is brutal, can we back off" becomes an intent - ease this week. "Shift my long run earlier" becomes move-session, Saturday. That is genuinely hard, and it is what an LLM is good at.

But the model does not touch your training numbers. Once the intent is decided, deterministic code applies the change - the 85% scaling, the date swap, the clamps. The model never gets to freelance your mileage, invent a workout, or decide that "easier" means halving your long run. It classifies; the rules execute.

That division is deliberate, and it is the honest test of any "AI coach": ask whether the AI is *deciding your training load* or *interpreting your request*. Ours interprets. If a plan's numbers came straight out of a language model, they would be plausible and unaccountable in equal measure - we wrote about why the model is kept away from the mileage curve.

When It Refuses, and That Is the Feature

If you ask for something the safe rules can't map to a well-defined change, nothing is written. The coach still replies - it does not pretend it did something - but your plan stays as it was. A no-op is the correct answer to an ambiguous instruction, far better than a confident wrong edit to a training plan you are trusting.

And genuinely big changes route away from the in-week machinery entirely. A new race date or a changed goal triggers a full regenerate on the plan page - a fresh periodised block built from your updated inputs, not a patched version of the old one. Small things bend; big things rebuild. Nothing important gets quietly half-changed.

What This Means for You

The practical upshot: you can be honest with the coach without fear of breaking your plan. Tell it the week is too much and it eases *this* week, not your race. Miss a few days and say so, and the remaining sessions absorb it rather than piling up as a debt you have to see every time you open the app.

That is the entire case for an adaptive plan over a PDF: the PDF was written before it met you and cannot respond; this responds within bounds you can predict. If you want the wider picture - what an AI coach costs, where it fails, and how ours is built - the full explainer is here, and the honest comparison against a human coach is here.

Or just generate a plan and, after your first run, tell it how that went.