I wear a WHOOP every day. I train on TrainerRoad. I log strength work
in Hevy. And at the end of a gym session, I take screenshots of my Hevy
workout and paste them into WHOOP’s AI so it knows what I did.

That workaround is embarrassing to describe, and it is the most
honest product review I can write.

WHOOP’s AI will happily accept a spoken description of a workout. It
is sophisticated. It will also make me stand in a gym parking lot
narrating sets and reps into my phone, which I have done exactly once.
Hevy is not sophisticated. I build the workout ahead of time, hit start
at the gym, and my Watch tells me the next movement. I can skip it, log
reps and perceived effort on my wrist between sets, and never take my
phone out. So Hevy gets the data, and WHOOP gets a screenshot.

Easy beat sophisticated. It usually does.

Where the
sophistication actually earned its keep

I want to be fair to WHOOP, because there is a version of “AI
coaching” it does better than anything else I’ve used, and the contrast
with the logging problem is the whole point.

Earlier this year, during Ramadan, I woke up to a red recovery. HRV
had dropped from the mid-70s to the mid-20s overnight. Blood pressure
was up. I was congested and probably getting sick. I had overslept and
missed the pre-dawn meal, which meant a dry fast, no food or water until
sunset, with two hard appointments on the calendar and a one-hour
endurance ride already scheduled.

Everything about that morning said skip the ride. Except that cycling
is an antidepressant for me, and skipping wasn’t really on the
table.

So I had a conversation with WHOOP’s AI. I told it things it couldn’t
know: the dry fast, the missed meal, the emotional weight of the day. It
adjusted. We landed on swapping the planned ride for a shorter,
controlled ERG session in TrainerRoad, with explicit guardrails on
effort and permission to bail early. Roughly a third less load, same
reason for getting on the bike.

That is what I always hoped this category could do. Not “your
recovery is low, rest today.” Something closer to “given everything
you’re carrying right now, here’s how to make this beneficial instead of
a hole you dig yourself into.”

Notice what made that interaction work, though. It was a decision
point. I had one question, I had context the sensor couldn’t see, and
the AI’s job was judgment. Talking was the right interface for that.

Logging is the opposite. It’s high-frequency, low-judgment, and it
happens with chalk on my hands. Talking is the wrong interface for that,
no matter how good the model is on the other end.

The pattern I keep seeing

Once you notice it, it’s everywhere in this category.

My wife stopped wearing her Apple Watch. She wears her Oura ring
every day and takes it off only to charge it. It asks nothing of her.
The one exception: she pinched a finger between a barbell and the ring
during deadlifts, so it comes off for the gym now. Her trainer does the
same thing, Oura for daily wear and an Apple Watch for workouts. Two
devices, each doing the job that costs the least attention.

WHOOP wins my recovery data because it’s a strap I forget about. Hevy
wins my strength data because a tap on my wrist is cheaper than a
sentence. TrainerRoad wins my cycling because the trainer logs the ride
whether I remember to or not.

The apps that own my data are the ones that made capture disappear.
The AI shows up later, at the moment I have a question. Not one of the
tools I actually use every day asks me to do work in order to be
smart.

A new entrant
that has the right raw material

A former Apple product lead recently announced Meridian, a strength
tracker that uses on-device computer vision to count sets and reps from
your phone’s camera, with a voice-only fallback. By his account,
strength is the top fitness goal in the U.S. and only about one in five
people log strength workouts consistently, and the pitch is that
automatic logging fixes the consistency problem.

I’m the target user, so I asked the obvious question in his comments:
what does voice-only give me over what I already have on my wrist? I’m
not sure it does, yet.

But the screenshot he posted had a skeleton overlay on the lifter.
That means the pose data already exists. The step from “three reps” to
“your back is rounding, reset” is the thing a live trainer gives me and
no app currently does. If that ships, the input-friction question
changes entirely, because I’d point a camera at myself for form feedback
in a way I would never talk to my phone for a rep count.

The rep counter is the easy version. The form coach is the
sophisticated version that would actually be worth some friction.

What I’d take into a build

I’ve shipped health and fitness apps, and I’ve been on the wrong side
of this more than once, so this is as much a note to myself as to anyone
else.

Capture should cost nothing. If your feature needs the user to
describe what they did, you’ve already lost to whichever app logs it on
their wrist or from the trainer or the sensor. The bar is a screenshot
pasted into a chat box, and that bar is low.

Put the intelligence at the decision, not at the data entry. The
moments where a conversational AI is welcome are the moments where the
user has a question and private context the sensor can’t see. Those are
rare, and they’re where the model earns its keep.

Design for the two-device reality. Most serious users already run a
passive daily wearable plus a workout-specific device. Fighting that is
a losing battle; integrating with it is the product.

And the sensor you already have is probably underused. Meridian’s
skeleton overlay is a reminder that a lot of us are sitting on richer
signals than the feature we shipped. The question is whether the next
feature asks the user for more, or gives them something back for what
they’ve already handed over.

Sophistication is not the moat. Making the sophisticated thing feel
easy is.