There's a version of AI-powered fitness that's been overpromised for years. You answer ten questions, and the algorithm hands you a 12-week program. It's personalised — in the sense that your name is at the top and the weights are adjusted for your inputs. But any experienced coach will tell you that what's generated looks about right and will almost certainly miss the point.
The gap isn't a technology limitation. It's a category error.
What "personalisation" actually requires
When a strength coach sits down to program for an athlete, they're drawing on a layer of context that no intake form captures:
They know that this particular athlete lifts emotionally — that when they're stressed, they overreach and their technique breaks down at 80% of what they should be handling. They know the athlete had a back tweak eight months ago that the physio said resolved, but which still shows up in how they set up for deadlifts. They know the athlete's job involves long hours on their feet three days a week. They know the athlete responds better to volume than to intensity, that RPE-based loading confuses them, and that they need explicit coaching on posterior pelvic tilt every single warm-up.
None of this lives in a database. It lives in a coach's working memory, built over months of observation and relationship.
Generic AI programming starts from zero with every session. It has no access to any of that context — because none of that context was ever structured.
The two failure modes
Most training apps fail coaches in one of two ways.
The first is the automation trap: the app generates programs without requiring coach involvement. This works acceptably for beginners who don't yet know the difference between a program that fits them and one that doesn't. For experienced athletes and clinical populations, it produces programs that are fine on paper and miss the mark in practice. The coach is cut out of the loop — which means their expertise is also cut out of the loop.
The second is the customisation trap: the app gives coaches complete flexibility to program however they want, with no intelligent scaffolding. These tools are glorified word processors. They don't reduce the time it takes to program. They don't surface useful information at the right moment. They don't calculate progressions automatically. The coach is doing all the cognitive work, the tool is just storing the output.
Both failure modes share the same root problem: they treat the coach and the algorithm as alternatives, rather than as a team.
What coach-guarded automation looks like
The right model is one where the AI does the computationally expensive work — generating a program skeleton from assessment data, calculating load progressions, adjusting volume based on logged fatigue — while the coach retains full authority over every decision.
This isn't a subtle distinction. In practice, it means:
The AI generates a phase-structured program based on the athlete's goal, training age, movement screen, and history. The coach reviews it, adjusts what doesn't fit, and approves it. The athlete receives what their coach prescribes, not what an algorithm prescribed.
When session data comes in, the AI calculates next week's load recommendations and flags any anomalies — a session where bar speed dropped unexpectedly, an RPE that spiked relative to expected. The coach sees these flags and decides what to do with them. The auto-progression runs by default if the coach doesn't intervene; the coach's judgment overrides it if they do.
The athlete's context — injury history, psychological tendencies, life stressors — lives in structured notes that the AI can surface at the relevant moment. When the coach is programming a lower body block, the flagged back history appears. It doesn't disappear into a PDF that nobody reads.
Why the coach's philosophy matters
There's another layer that generic AI can't replicate: programming philosophy.
Two equally qualified coaches may program very differently. One runs linear periodisation almost exclusively. Another uses block periodisation. One front-loads volume and tapers for competition; another runs submaximal loads year-round. One believes in daily maxes; another thinks they're an injury risk for most athletes.
None of these approaches is wrong. They're philosophies — developed through years of practice, refined through watching what works for the athletes they coach.
A generic AI has no philosophy. It has defaults. And defaults produce average outcomes.
The right tool preserves and amplifies the coach's philosophy rather than replacing it with the algorithm's. When a coach sets up their template — their preferred periodisation structure, their go-to exercises for each pattern, their progression rules — those preferences should propagate through everything the system generates. The AI should feel like an extension of the coach's thinking, not a substitute for it.
Where AIMS fits
This is the design principle behind AIMS. The framework generates program structures from structured assessment data — but every program passes through coach review before it reaches an athlete. Auto-progressions run based on session data — but the coach sees the logic and can override it at any point. The system builds a longitudinal picture of each athlete — but the coach controls what that picture means and what to do with it.
The goal isn't to replace the expertise that makes good coaching good. The goal is to eliminate the administrative overhead that sits between a coach's expertise and the athletes who need it.
For most coaches, that overhead is 30–40% of their working time. Intake paperwork, program formatting, tracking progressions manually, compiling outcome reports. None of that requires expertise. All of it consumes it.
When the AI handles those tasks, coaches get that time back. Not to do less coaching — to do more of it, with more athletes, at a higher standard.
That's the version of AI-powered training that actually works.