I built an AI fitness coach, then realized AI was only solving half my funnel

I spent 8 years working in tech, as a Product Manager and Designer. Today I'm a certified personal trainer running two things at once: FitDots, an AI-powered training app, and a 1:1 coaching business on top of it. This post is about the part nobody warns you about when you build an AI product: figuring out exactly where the AI's job ends and where a human has to take over — and building your funnel around that line instead of pretending it doesn't exist.

The problem I actually had:

When I started building FitDots, my first instinct (like most of us building AI products) was "automate everything." Generate the workout, adapt it to time and fitness level, ship it, done.

That part worked. Give the system 15 to 45 minutes, a fitness level, and feedback from the last session, and it produces a structurally solid workout, every time, at zero marginal cost. That's a genuinely good use of AI: bounded inputs, bounded outputs, no ambiguity.

But I kept seeing the same failure mode with users: people completing every workout in the app and still quietly losing motivation, or drifting toward burnout, weeks before it showed up in any metric I was logging. The AI had no signal for that, because the thing it needed to detect wasn't in the data. It was in tone, in the gap between what someone reports and what's actually going on, in change over time that only a human tracking someone across weeks would notice.

That's when I stopped thinking of "add more AI" as the roadmap, and started thinking about the product as two separate businesses stitched together.

Splitting the business, not just the product:

Here's the model I landed on, and the reasoning behind each piece:

Free tier (FitDots + a nutrition guide as lead magnet): this is the part that scales for $0 marginal cost. Adaptive workouts, and a free resource I built called "The Lean Nutrition System - Foundations" that covers the four levers that actually drive body composition (protein intake, treating carbs as fuel, adjusting for training/rest/travel days, not spiraling after one bad meal). None of this requires judgment calls — it's a framework, so it's exactly what you should be handing to an algorithm.

Paid 1:1 coaching: this is where the actual margin is, and it's deliberately NOT automatable. It's positioned for people who've already got structure from the free tier and are still stuck — the ones where the bottleneck is diagnostic (is this a protein number, a stress response, or a pattern they can't see because they're the one living it?) or behavioral (getting someone to be honest about their week, or to not abandon the whole plan after one missed session).

The free tier isn't a watered-down trial of the paid tier. It's solving a genuinely different problem (logistics) than the paid tier (diagnosis + accountability). That distinction turned out to matter a lot for how I write copy, price things, and decide what to build next in the app vs. what stays human-only forever.

The four things I stopped trying to automate:

Useful to spell out, because "AI can't replace human connection" is such a vague claim it's basically useless as a product decision. Concretely, here's what doesn't scale, and why:

Assessment — telling training fatigue apart from early burnout. Requires context on the person, not just their inputs.
Edge-case nutrition — disordered eating patterns, medical history affecting metabolism, the gap between reported and actual intake. A framework can't catch what a person isn't reporting.
Behavioral coaching — getting someone to reframe a setback instead of spiraling into all-or-nothing thinking. This is closer to therapy than to logistics.
Accountability — compliance changes when there's a real person on the other side who'd notice if something's off. This one might be the most economically important of the four; it's most of why people pay for coaching instead of just following a plan.

Why this matters if you're building an AI product?

If you're building something AI-powered right now, the question I'd push on is: what's the actual shape of the problem you're solving, and does it match what AI is structurally good at (bounded inputs → bounded outputs, applied at scale) or what it's bad at (judgment calls under incomplete information)?

Most AI products I see get pitched as "AI does everything now," which is a fine hook but a bad roadmap. The version that's actually defensible, at least for me, was admitting AI covers maybe half the value chain — and building a real (paid) business on the half it doesn't.

Happy to go deeper on the funnel numbers (free → paid conversion, what actually gets people to book a call) if that's useful to anyone building something similar.