AI Coaching for Runners: Pace, HR, Power, and Form Feedback

A runner's watch records three ways of measuring the same effort: pace, heart rate, and power. Each one lies in its own way. Pace lies on hills. Heart rate lies for the first ten minutes and again after an hour. Power tells the truth faster than either but speaks a dialect only its own device understands. A coach, human or AI, earns their fee by knowing which signal to trust in which moment.

This guide covers what an AI coach does with those three signals, what an adaptive running plan looks like when your week falls apart, and an honest answer on form feedback: what watch metrics can flag, and where they stop.

Pace: the signal everyone trusts too much

Pace is the number runners organize their identity around, and it is the least stable of the three. GPS pace wanders under tree cover, in cities, and on switchbacks. On any grade it stops describing effort at all: 5:30/km up a 6% hill is a harder effort than 4:45/km on the flat, and raw pace has no way to say so. Trail runners know this. Their easy days read like disasters on paper.

Grade-adjusted pace corrects some of it by converting hill splits into flat-equivalent effort, and you can convert your own hill splits to flat-equivalent pace to see how big the gap gets. But GAP is a correction, a model on top of noisy data.

Where pace earns its keep: flat, measured efforts. A track session, a road race, a tempo on a known route. There it is the sharpest of the three signals, because it is the outcome you are training for. The pace calculator turns a goal time into the splits that pace-driven sessions are built from.

Heart rate: honest but slow

Heart rate reports internal cost, which pace cannot. The problem is timing. HR lags effort by 30-90 seconds, so short intervals finish before your heart rate arrives at the work. And on long efforts it drifts: the same pace at minute 70 costs 5-10 more beats than at minute 20 as heat and dehydration stack up. Caffeine, sleep debt, and stress all move the number before you take a step.

That lag and drift are why the consensus among running coaches is to use HR for what it measures well: sustained aerobic work and day-to-day recovery state. Zone 2 training lives on heart rate for a reason. So does drift itself, which is a signal, not noise: rising HR at a held pace late in a long run says something about aerobic durability that no single split can. Get your zones set from your max or resting HR with the HR zones calculator before you read anything into them.

Running power: fast, useful, device-specific

Wrist-based running power responds within a couple of seconds. Effort on a hill, into a headwind, on soft ground: power sees it while pace is lying and HR is still catching up. For hill repeats and fartlek-style work it is the best pacing signal on your arm.

The catch is that running power has no agreed standard. Garmin's model, Stryd's model, and Apple's model compute different numbers from different inputs. A watt on one platform is a different watt on another. So power works as a within-device signal: your threshold power on your watch, trended over months on that watch. Compare it across platforms or against a friend's number and it stops meaning anything.

How an AI coach reads all three together

A single metric is an opinion. Three metrics with timestamps are a story. This is the actual work of coaching from data: cross-reading.

Held pace with climbing HR late in a long run reads as fatigue or heat, and the coach should shorten Thursday. Faster pace at the same HR across three weeks of tempos reads as fitness, and thresholds move up. High power with flat pace reads as terrain or wind, and the session gets judged on effort, not splits. A human coach with your file open does this cross-read in their head. An AI coach does it on every activity, because reading the whole file costs it nothing.

This is also where a coach beats a formula. The pace tables and HR zones in a static plan assume the signals agree. Real training data disagrees constantly, and the disagreements are where the information is.

The same cross-read applies to racing. A goal time from the race time predictor gives you a target, but whether that target is live on race morning depends on what the last six weeks of paired pace-and-HR data say about your fitness. A coach that has read every session can tell you the predicted time is stale, in either direction, before you commit to a suicidal first 5K.

What an adaptive plan looks like in a real week

Static plans assume you execute. Adaptive plans assume you will not, and program around it. See the adaptive training plan guide for the general mechanism. For a runner it looks like this:

Missed sessions get absorbed, not stacked. Skip Tuesday's intervals and a static plan leaves you a guilt-driven doubling on Wednesday. An adaptive plan asks what the week still needs. Sometimes the intervals move to Thursday and the long run shortens. Sometimes they die and the week moves on. Missed work is gone. Chasing it is how runners get hurt.

A goal race reshapes the block, not one session. When a half marathon moves from ten weeks out to six, the question is what to cut, and the answer depends on where you are: a runner behind on long-run volume cuts different work than one behind on threshold time. A template cannot ask. A coach can, and does, in the chat.

Fatigue changes intensity before it changes injury status. Rising resting HR, HRV trending below baseline, climbing HR at your easy pace: each one is a lead indicator. A coach that reads them swaps a threshold session for an easy 40 minutes and keeps the block alive. Most evidence on training adaptation points the same direction: consistency over weeks beats any single hard session.

The plan knows what else you did. A heavy lower-body session on Monday changes what Tuesday's run should be. Strength work belongs in a runner's week, but only if the running plan can see it. A plan that reads your strength log programs quality runs away from heavy leg days instead of on top of them.

Form feedback: what the data can and cannot say

Time for the honest section. Modern watches report running dynamics: cadence, ground contact time, vertical oscillation, sometimes stride length and balance. These are real measurements and their trends carry information. Cadence dropping 6 steps per minute in the last third of every long run is a fatigue pattern worth acting on. Ground contact time creeping up over a training block can track with accumulating fatigue. Check your own number against the cadence calculator, and read running cadence and form for what the ranges mean.

Here is the limit: an AI coach reading these numbers can flag trends. It cannot see you run. Overstriding, hip drop, crossover gait, asymmetries that load one tendon more than the other: those live in video, not in a wrist sensor's summary statistics. Two runners can post identical cadence and vertical oscillation with completely different mechanics.

So the division of labor is plain. Watch metrics plus an AI coach: trend detection, fatigue-related form decay, cadence work. Video gait analysis with a physio or a running-specialist coach: mechanics, injury-driven movement patterns, technique rebuilds. Any AI product claiming to fix your form from watch data alone is overreaching. What the data supports is narrower and still useful: catching the drift before it becomes a diagnosis.

What Movement Rebels does with a runner's data

The concrete version, so you can judge the claims.

Garmin sync is live in both directions. Activities, HRV, sleep, and training load flow in through Garmin Connect. The coach writes structured sessions, warmup, intervals, targets, cooldown, and pushes them to your watch. You hit start and the watch guides the session. On iPhone, the native iOS app reads Apple Health, covered in the Apple Health AI coach guide.

Every run gets a debrief. A completed Garmin or Apple Health activity triggers an automatic coach analysis in your chat thread. Lap splits, HR drift, plain language. The cross-reading described above, delivered while the run is still fresh.

Weekly plans read your whole picture. Plan generation pulls the last 14 days of training data, your strength log, your nutrition entries, and a persistent athlete profile built from every conversation you have had with the coach. The plan that comes out reflects the week you had, the race on your calendar, and the knee you mentioned three weeks ago.

The rest of the runner's stack is in the same app. A strength library with 3,600+ workouts for the gym days, Rebel Fuel for macro targets that follow the day's session, recovery tools, and a morning brief that reads your overnight data before you decide how hard to go. What is not in the app: a social feed. No kudos, no segments. Activities sync to Strava for the social layer, and the training lives in MR. The full comparison with Garmin's own coaching is in the Garmin AI coach guide.

Who should still hire a human

Some runners need a person. If you are chasing an Olympic Trials or national-class standard, race tactics and taper judgment at that level are still human craft. If you are rebuilding from a complex injury, you need eyes on your gait and a physio in the loop. And some athletes train better because a person they respect expects the work; software does not replace that.

For the rest, which is most runners training 3-6 days a week around a job: the AI coach versus personal trainer comparison walks through the trade in detail. The short version is that an AI coach reads more of your data than a human has time to, answers at 6 a.m. before the run, and costs less per month than one session with a coach.

Pricing

Movement Rebels is free for 7 days: full coach access, full Garmin sync, strength library, Rebel Fuel, recovery tools, no credit card. After the trial, the subscription is $20/month.