Riders are right to ask what is deciding their training. Plenty of apps now describe themselves as coaching by artificial intelligence without ever saying what that means in practice, and a plan you cannot interrogate is a plan you have to take on faith.
So here is the honest answer for WattPlan, in the order it actually runs.
Two systems, and only one of them writes your training
WattPlan has a deterministic layer and a conversational layer, and the split between them is the whole point.
The deterministic layer builds and repairs your plan. It is ordinary code following published exercise physiology. Given the same inputs it produces the same plan every time, and it is covered by an automated test suite that checks the calculations against known values.
The Coach is the conversational layer. It reads your numbers and explains them, and it can ask the deterministic layer to make specific changes on your behalf. It does not invent training.
That ordering matters. The language model never decides what a workout should be. It picks from a library of structured workouts by exact name, and the request then goes through the same code path as any other change to your week.
What the maths actually does
Every ride you complete is reduced to a few numbers. Weighted Power is your intensity-weighted average power, which reflects the cost of a ragged ride better than a plain average. Intensity is that figure divided by your FTP. Load is the training stress of the ride, derived from intensity and duration. When a ride arrives with no power data, Load is estimated from heart rate using a zone-weighted impulse calculation with gender-specific coefficients.
Those daily Load figures feed an impulse-response model published by Banister in 1975. Two exponentially weighted averages run over your history: one long, at around 42 days, and one short, at around 7. WattPlan shows you the long one as Fitness, the short one as Fatigue, and the difference between them as Form. That is the whole of it. There is no hidden adjustment.
A few other established models sit alongside it. Critical power and W' come from the Monod two-parameter model, fitted by linear regression across your power curve. Estimated VO2max comes from the Hawley and Noakes regression on your best five-minute power. Personal bests are checked across nine durations, from five seconds to an hour. Zones follow Coggan's framework as percentages of FTP.
None of this is novel, and that is deliberate. You can look all of it up, and if you have worked with a coach or kept your own spreadsheet you will recognise every model in the list.
How the plan gets built
When you create a plan, the generator works backwards from your goal date through phases: base, build and peak. If you are training for a real event it adds a taper. If your goal is general fitness or a higher FTP, with no start line to arrive at, it does not taper. It finishes on a consolidation week that holds your load steady and adds a 20-minute FTP test, because tapering towards nothing just throws away fitness.
Within each week the generator works to a target daily Load, then places workouts on the days you said you can ride, respecting the minutes you gave for each day. Your long ride goes on your long-ride day. Sessions are drawn from the workout library rather than improvised.
One number worth understanding is Target Fitness, the peak the plan builds to before any taper. It is bounded by your available hours and by how fast the plan is willing to raise your load. If that number comes out lower than you hoped, the plan is telling you something true about the time you have given it.
The rules your week has to obey
This is the part most plans skip. Changing one day is easy. Keeping the week sensible afterwards is the hard bit, and it is where a plan usually falls apart.
Every change routes through a repair step that enforces the same set of rules on the days it is still allowed to touch:
- No workout lands on a day you marked unavailable.
- No two demanding days sit back to back. If a change creates a pair, the later one is eased rather than deleted.
- No two rest days sit back to back by accident. One becomes easy endurance, preferring a day you can actually ride.
- A deliberate recovery block is left alone. Intentional rest is not an error to be corrected.
- Your long ride returns to your long-ride day if it has drifted.
- Days in the past, days you completed, and days you locked are never touched.
Those passes run repeatedly until the week stops changing, because fixing one day can disturb its neighbour. Every adjustment is recorded with what the day was before, when it changed, and why.
A worked example: two missed workouts
Say you have threshold intervals on Tuesday and a sweet spot workout on Thursday, and work swallows both.
WattPlan notices the Tuesday miss on its own. It looks for a past day that had a workout on it, was not a rest day, was not marked complete, and has no ride recorded against it. It only considers misses from the last three days, because a workout you skipped a fortnight ago is history rather than a problem to solve.
Then it looks for somewhere to put the session. The rules are narrow on purpose. The slot has to be in the future, editable, and on a day you said you could ride. It has to be a rest or easy day, so that absorbing one workout does not quietly cost you another. The workout has to fit the minutes you declared for that day, so a two-hour ride is not dropped onto a sixty-minute slot. And if the missed workout was a hard one, the slot cannot sit next to another hard day.
If no slot passes all of that, WattPlan does not cram it in. It leaves the workout missed. That is the correct coaching answer, and it is worth saying plainly: a single missed workout costs you very little, while a week bent out of shape to accommodate it can cost you the block.
With Thursday also gone, the week is now light. The repair step runs over what remains, dissolves any pair of rest days that appeared, keeps the weekend long ride where it belongs, and leaves you with a week that still makes sense rather than a pile of debris and a guilt trip.
Everyone gets the recommendation and a single tap to apply it. On Premium, a missed key workout can be repaired overnight, and the change comes with one-tap Undo and a record of exactly what moved.
What the Coach may and may not do
Ask the Coach a question and it answers from your own plan, your recent rides, your Fitness and Fatigue, your availability and your event. Ask it to change something and it can, using a fixed set of actions: change a workout, schedule rest, swap two days, reduce the week's volume, place a recovery block, move your event, or note something you have told it.
Four constraints apply to all of it.
It must name an exact workout from the library. It cannot describe a session it has made up. It must use absolute dates, because conversations outlive the day they started. It must ask before making a change you did not request. And whatever it does passes through the same repair rules as any other edit, with the previous state saved.
The instruction it works under is explicit: it is an advisor, not an autopilot. It recommends one clear next step rather than handing you a menu, and it asks before acting on its own initiative.
What it does not do
It does not diagnose anything. WattPlan is training software, not a medical device, and it cannot tell whether you are ill, only that your numbers have moved.
It does not promise watts. No plan can, and any product that puts a number on your improvement is selling you something it cannot deliver.
It does not read your mind about the things it cannot see. Sleep, stress, illness and a heavy week at work all shape what you should do tomorrow, and WattPlan only knows about them if you say so. Telling the Coach is the whole mechanism. It is also why the Coach keeps what you tell it across conversations, so you are not explaining your knee for the fourth time.
And it does not hide its working. Every adjustment carries a reason, and you can always see what a day was before something changed it.
Why this way round
The case for putting a language model in charge of training is that it is flexible. The case against is that flexibility is exactly wrong for load progression, where the same inputs should produce the same answer and where a plausible-sounding invention can cost you a season.
So WattPlan uses each part for what it is good at. The maths decides the training, because the maths is testable and consistent. The Coach reads it back to you in plain language, argues with you about it, and changes it when you ask.
If you recognise the models in this article from the literature or from your coach's spreadsheet, that is the point. WattPlan does the same maths on your own rides, without asking you to keep the spreadsheet.