It's 2026 and the old objection is dead. You used to be told a general AI "doesn't have your data." Now you can wire a Strava or intervals.icu MCP straight into Claude or ChatGPT, ask it to analyze your last training block, and get an answer in seconds: fluent, specific, confident. It feels like you just got a free coach.
You didn't. What you got is faster access to your own history — and that's genuinely useful. But surfacing information is descriptive. Coaching is decision-making. The gap between the two is where seasons are won and lost.
Describing your training isn't coaching it
Plugged into your data, the model is excellent at the descriptive layer: pulling up last month's volume, summarizing a workout, explaining what your FTP test result means in plain language instead of making you click through charts. Real value. Keep using it for that.
The trouble starts the moment people extrapolate from describing the past to deciding the future — generating training plans, prescribing pacing, predicting performance. That's a fundamentally different job, and a language model is the wrong tool for it.
It's confident. That's not the same as correct.
A general LLM is, first and foremost, a language model: built to always produce a fluent, well-structured, convincing answer. It will basically never tell you "your data's too thin to say." It pattern-matches text that sounds like analysis — so it'll quote you a TSB it didn't really compute, estimate a CP it has no model for, and read a trend off a power curve it's half-imagining.
You already know this person in real life: the colleague who's effortlessly fluent, deeply persuasive, radiating overall competence — and who gets hired everywhere he applies, often promoted into the next job before reality catches up with the last one. We judge content by how confidently it's delivered, so he excels. An LLM is that person at scale. It sounds like the best coach you've ever spoken to — right up until the plan meets your legs.
It can't actually do the math — and it doesn't know the consequences
Real analysis is computation, not prose. Fitting a power–duration model to your mean-max curve. Simulating a pacing plan against your W'-balance, second by second. Projecting your fitness and freshness forward across a block. These are numerical models that have to be run. An LLM doesn't run them — it generates numbers shaped like the answer.
And because it doesn't simulate your physiology, it has no idea what its own advice will actually do to your body. Ask it to pace a climb and you'll get watts that look reasonable, are coherent, well-argued, convincing — and quietly wrong. The problem was never that the output makes no sense. Quite the opposite. It makes perfect sense, right up until you blow up on the second climb.
It does whatever you tell it
Tell a chatbot "make next week harder" and it makes it harder. Tell it "I feel great, add a third interval day" and it adds one. It has no method to defend and no sense of whether that breaks your progression or buries you in fatigue. It's agreeable to a fault — a yes-man with no training principles. A real methodology tells you no sometimes. That "no" is half of what a coach is for.
The real killer: methodology drift
Here's the one that quietly wrecks a season. Every conversation with a general AI is a fresh improvisation. Feed it your latest mood and it will follow it — too closely. Monday it's all-in on polarized training. Thursday, prompted slightly differently, it loves sweet-spot. Next week it half-remembers a different article and pivots again. Each answer is reasonable on its own. Strung together, they're a random walk.
But adaptation doesn't come from any single clever week — it comes from the same methodology applied consistently over months, with controlled variation, progressively loaded so your body has time to respond. Constantly re-shaping the approach in response to short-term feedback is often the opposite of what drives long-term progression. Switch methods every couple of weeks and you never give any stimulus long enough to work. You're not following a plan; you're sampling ten of them. That's how riders manage to stay busy and stagnant at the same time.
And tomorrow, it forgets
Close the chat and the context is mostly gone. Next session you re-explain who you are, what you're training for, what you did last week. There's no continuity and no accountability — nobody who notices you've quietly skipped your easy days for three weeks straight. A coach's value compounds. A chatbot's resets.
What actually comes next isn't "more LLM"
The fix isn't a bigger, smarter language model. It's a different architecture — one where the language model stops being the thing that reasons about your body and becomes the interface to things that can. An orchestration layer over specialised tools: physiological models, statistical models, simulators, and yes, other language models where they genuinely earn their place.
In that design the agent no longer just responds. It acts, simulates the consequences of those actions, observes what came back, and iterates — before it ever hands you a plan. This is the direction serious AI research keeps pointing to: Yann LeCun's "world models" — systems that can represent an environment, project possible actions, and test their consequences before committing to one. Language returns to what it fundamentally is — an interface. Not a source of truth.
Describing training is easy. Modeling adaptation is not. And that's exactly where the real work starts.
What CoachedBy does instead
That's the architecture CoachedBy is built on. The generative model is the interface you talk to — fluent, fast, easy — but it is not the thing deciding. Underneath sit validated physiological models computing your real numbers, a single coach methodology driving every plan (applied consistently, not reinvented each Monday to match your mood), and persistent memory of your whole history. And crucially, it doesn't just generate a plan and hand it over — it simulates the consequences first:
- Before a single workout reaches your calendar, CoachedBy models its impact on your body and tests whether it's actually feasible given your current fitness and accumulated fatigue. A session that would dig too deep gets reshaped or flagged — not served to you because it "looked reasonable."
- At the week and month level, it projects your fitness and freshness forward across the block, so it catches overreaching before it happens, instead of after you're already buried in a hole.
That's the act–simulate–observe–iterate loop in practice: a system that reasons about consequences, with a generative model as its interface and a real coach standing behind all of it.
So use it — for what it's actually good at
A general AI with your data plugged in is a brilliant, well-read intern: great for "what is polarized training?", "how should I fuel a four-hour ride?", "explain my FTP test result." Genuinely useful at the descriptive layer. But it can't do the math, it has no idea what its advice does to your body, it agrees with everything you say, it changes its mind every week, and it forgets you tomorrow.
That's not a coach. That's a very confident search engine wearing your data.