What AI is actually for in coaching

A while back I built a set of free nutrition tools for the people I was training. Nothing fancy — a cookbook where every recipe’s numbers are checked against USDA values, and a finder that does the same kind of checking across more than eleven thousand restaurant menu items. I built them because my clients kept getting stuck in the same place. Not on the big stuff. On the small, daily moment of: okay, what do I actually eat right now, and can I trust the number on it.

Then a while later I found MealPrep — an Australian site, half a world away from me — and they’d gone after, more or less, the same problem. A macro calculator, a way to compare meal providers, guides for reading a nutrition label. Different surface, same instinct. They describe themselves as a free way to “find the right meal provider.”⁴ Mine was a cookbook. But underneath, we’d both decided the same thing was worth building: not more advice, more trust. The numbers, made real and easy to reach, right at the moment someone’s actually choosing.

I think about that convergence a lot, because it tells you what the job actually is.

Here’s the thing nobody who’s coached for real will be surprised by — clients almost never lack information. They’ve read the articles. They know protein matters and that they should probably eat more vegetables. What they lack is quieter: a way to trust the numbers in front of them, and to get to those numbers without it turning into a research project. “Eat 140 grams of protein” is useless at 7pm when you’re standing in front of the fridge and the only thing you know about the leftovers is that they exist.

So the gap was never knowledge. It was trust, and it was availability. And those two, it turns out, are exactly the things this AI moment is good at.

Coaches are adopting AI fast — faster than almost any profession, if you believe the surveys. One 2026 survey of coaches — run by a coaching-software company, so read it as directional — puts AI use around 91% now, with one usage measure climbing from 15% in 2023 to 75% in 2024-25.¹ Most of that isn’t robots writing workout plans. It’s the boring, load-bearing stuff: about 73% use it for content and research, roughly half for nutrition planning.¹ The availability layer. AI is genuinely good at taking information that used to live in a coach’s head, or behind a paywall, or in nobody’s reach at all, and making it cheap to reach — though cheap to reach isn’t the same as right.

And here’s the part that surprised me, except it didn’t: the same coaches racing to adopt AI are the most certain about its limit. In that same survey, the single strongest consensus — stronger than anything else they were asked — was that 77% believe AI can never replace a human coach.¹ So the field is already living the balance. People feel where the line is. They just mostly haven’t said it out loud.

Let me try to.

The danger with this stuff was never that AI would be wrong. Everything’s wrong sometimes. The danger is that AI is wrong confidently — in a clean little number that looks exactly as authoritative as a right one.

Back in 2017, Stanford put seven popular fitness trackers through a proper test, and the gap it found hasn’t closed. The heart-rate sensors were great, off by less than 5%. The calorie counts were not. Not one device measured energy expenditure well; the most accurate was off by 27% on average, and the worst by 93%.² The senior author, Euan Ashley, put it about as plainly as you can — people are “basing life decisions on the data provided by these devices.”² And the errors weren’t even evenly spread; they got worse at higher body weights, which is to say, worse for a lot of the people most likely to be counting.

The food side is no cleaner, and it’s a different number — that was calories burned, this is calories eaten. A 2025 writeup from a meal-tracking company found one app overestimating a meal by 37% from a database mismatch, plus a systematic skew where the models over-counted some cuisines and under-counted others, because that’s what they were trained on.³ Even getting the food right, the writeup noted, doesn’t “always equate to precise calorie calculations.”³ The app can name your lunch correctly and still miss the number by a quarter.

None of that means the tools are useless. It means the number on the screen is a draft, not a verdict — and somebody has to know the difference. That’s the whole game, honestly. The gap between an answer that looks right and one that is right.

So where’s the line

After building these things and using them with real people, here’s where I’ve landed.

AI’s real job in coaching is the data-and-availability layer. Make the numbers trustworthy — actually traceable to a real source, not just confidently rendered — and make them reachable at the moment of choosing. That’s the part machines genuinely make better, and it’s the part I keep building. When I check every value in a cookbook against a real nutrition database, I’m not doing it to be precious about it. A number you can’t trace is a number you can’t trust, and a client who gets burned once by a bad number quietly stops trusting all of them.

The human keeps the rest. The relationship — the reason someone tells you the truth about the week they actually had. The judgment to know that the “optimal” plan is the one this particular person will actually do. The accountability that a push notification will never be. And the check at the end that almost nobody automates: did this actually work, for this person, or does it just look like it should have.

There are real ways to get this wrong, and they’re worth naming. Lean too hard on AI for nutrition and you can wander into scope-of-practice you’re not licensed for — and that’s a line worth respecting. Outsource the relationship and you’ve automated away the only part that was ever really yours. Trust the clean number without checking it and you’ll hand someone a confident mistake. That same survey had a line about competition I keep coming back to — that the real threat to coaches is “other coaches using AI better.”¹ I’d just add: “better” doesn’t mean more of it. It means using it for the layer it’s actually good at, and keeping yourself firmly in the loop on the layer it isn’t.

If I had to put it in one move, it isn’t adopt-or-resist. It’s narrower. What’s worked for me is using it to make the numbers real, and to put trustworthy numbers within reach of people who couldn’t get to them before — and staying in the room for the part where being wrong actually costs someone something. The tools get the data close. You’re still the one who knows whether it’s right.

That’s the thing MealPrep and I both landed on, independently — different products, same idea. Build the trustworthy version of what people already need. The AI’s just what made that version cheap to build.