When the research disagrees with the roadmap.

Three months in, the line was flat

Perflection AI builds motion analysis for golf. We'd shipped SneakySwing on a B2B2C model — coaches on one side, their students on the other — with a promise to golfers we were proud of: a coach in your pocket. Distil a coach's expertise, build an AI digital twin, and let it analyse your swing whenever you practise.

Three months after launch, growth hadn't moved and investors were telling us the scale wasn't big enough. Inside the company, two readings were on the table: we hadn't marketed hard enough, or we'd built something people don't want.

Those two diagnoses point the company's resources in opposite directions — one spends on reach, the other on rebuilding the product. So I went to find out which one was true.

Half the product was working. Half of it wasn't.

“An AI coach is just a worse coach”

Across interviews, the same thing kept coming back — and it went directly against the product we had shipped.

The wallet agreed with the sentiment: when AI analysis wasn't meaningfully better than the free tools already on their phone, nobody paid for it.

But they weren't saying they didn't need help. What they described wanting was specific — and it wasn't a coach substitute:

Which meant the thing we'd built — a machine that tells you what's wrong — was answering a question nobody was stuck on.

What we built instead

We talked to 20+ golfers and 6 coaches, and read 5+ competing products against each other. What came back wasn't two wish-lists. It was one blind spot both sides shared: the practice nobody is there for.

A golfer sees their coach for an hour a week. For plenty of them it's an hour a month. So 99% of the time the two of them are not in the same place — and that 99% is where the swing actually gets built, or quietly broken.

A copilot for the relationship — less work for the coach, and a golfer who feels more seen.

That created a hard design problem: how do you use AI heavily without letting it replace the coach? The answer turned out not to be a better model. It was deciding what counts as agreed — and building the thing that decides it.

Three conversations feed the system: the golfer alone with the AI, the shared room where golfer and coach both talk, and the coach's own feedback. Nothing in them is adopted on the AI's say-so. Every statement is typed, logged against a shared vocabulary, and then it waits — a self-report becomes a goal only once the golfer confirms it, and the issue pool becomes a goal only once the coach has edited it and set one. The system never concludes anything by itself.

Where the two sides disagree, there is a rule rather than a guess: the coach's judgement takes precedence. Their instruction and feedback override the golfer's own goal in the current focus — the thing that decides how the AI reads the next session. The golfer's goals aren't deleted; they're outranked by the person the golfer came here to learn from.

That's where the trust comes from, and it isn't the model. Everything the AI says next traces back to something a real person wrote down and a real person approved. Neither human is being automated — they are the two who decide what the system treats as true, and the AI does the noticing, sorting and remembering in between.

What the coach is actually buying

Coaches don't divide by skill. They divide by how they hold a roster — and the two kinds end up wanting the same thing for opposite reasons.

High volume

Relationship as a cost

More students than a day has hours. Replying to everyone simply does not fit, so staying in touch is something to contain — and they only track who read what for their top students.

High touch

Relationship as an investment

A roster small enough to know by name, and they want to see all of it: did the feedback get opened, did the golfer practice after it. The same work compounds instead of draining.

Opposite economics, one shared ask — stay present with every student, not just the ones there was time for. That is what a coach is paying for. It is also why the thing being automated here is relationship maintenance, not swing analysis: the analysis was never the scarce part.

The scarce part is being understood. Drawing a line on a video is table stakes — every coaching app does it. What coaches and golfers both asked for, unprompted, was comparison: this swing against the one from six weeks ago, or against the coach's own. That is the difference between an annotation and an explanation, and it's the feature request I heard from both sides of the marketplace.

One dataset, many ways to see it

The research also killed the idea that there is one right report. One golfer told us skeleton points and joint-angle charts were unreadable — “I can't tell what I'm looking at.” A tour-level coach in the same study wanted the opposite: exact pelvis and chest-turn angles, side by side with a professional's swing. Same data, opposite needs. Several were also explicit that templated analysis is actively wrong for them — bodies differ, so a one-size-fits-all read misdiagnoses.

So what we're building isn't a single interface. It's a preference model × modules × components:

Progressive profiling — earn the data, don't demand it

We ask for as little as possible at the door — four questions — because a long form is where new users leave. Everything after that accumulates on its own: implicit signals from what they open and act on; small questions surfaced in the drawer while their analysis is processing, turning dead time into signal instead of a spinner; the profile they choose to fill in as they get more invested; and their coach's annotations — expert labels that feed the preference model and the analysis model at the same time.

The same session then reaches three people three different ways: a body-feel cue for one golfer, a trajectory overlay for another, precise angles for their coach. Learning a sport was never a one-size-fits-all process — so the way the information arrives can't be either.

And because preference is learned rather than hard-coded, it travels with the person — which is what makes the next part possible.

What it cost to be right

Repositioning sounds clean written down. It wasn't.

Once “the coach cannot be replaced” became the premise, roughly 60% of the existing roadmap and data tagging had to be reworked. Months of model work aimed at out-performing a human coach stopped being the point — the AI no longer needed to be that powerful. The hard problem moved from model accuracy to relationship and data.

That's an expensive thing to say to a team that has been building the other thing — and it's the part of this project I'm proudest of handling. I came in as a UX researcher; I now set product direction and lead a 6-person product team across design, growth and research. The work that actually takes the time isn't the research. It's holding a direction steady across people who each see a different part of it: engineers who'd built for the old premise, designers who need decisions to be concrete, founders carrying investor pressure, and the 11 PGA coaches who'd signed on, whose trust is the product.

The reposition only stuck because each of those groups could see their own reason to believe it. Not being the person with the answer — getting everyone pointed at the same one.

Fast is only an advantage if you can tell when you're wrong

We ship every one to two weeks. Research that takes six weeks to answer a question is research nobody will wait for — so the real constraint on my work isn't rigour, it's keeping pace with the build.

That pace is our biggest advantage over a large company: we can be wrong far faster than they can be right. But being wrong is only cheap if you can tell that you were — quickly, and ideally before it costs a release. Two things make that possible.

One · Numbers tell you where. People tell you why.

At our scale, no number is significant on its own. That doesn't mean we don't look — we read everything; we just never let a number conclude anything by itself. Small samples don't lie, they under-determine. The risk isn't missing the signal, it's over-reading it. So nothing travels from a chart to a decision without passing through people on the way.

Two · Stop using releases as the test

A public release is an expensive way to find out a concept doesn't land — and at a two-week cadence, that's an expensive mistake you could repeat twenty-six times a year. Trial and error only works when the error is caught somewhere cheaper. So I built the research to run on two tracks at once.

This is dual-track research: discovery and delivery running in parallel rather than in sequence. The delivery track keeps the current product honest; the discovery track keeps us from betting a roadmap on a hunch. The moment the two become sequential, you're either polishing something nobody wants or shipping something you never checked — and at our speed, either one compounds fast.

What holds both halves together is a persona system built on behaviour, not attributes — level of investment × willingness to pay × core need — with sample size and confidence attached to every segment, so we always know how much weight a finding can bear. An n=1 segment stays provisional. Nobody gets force-fit into a category to make the deck look tidy.

In a company where the product changes month to month, personas built on attributes rot on contact with a pivot. Built on behaviour, they survive one — and the team is still using them.

Why this isn't a golf app

Early on, the sentence that kept us out of rooms was always the same: the scale isn't big enough. Golf coaching is a small market, and about that they were right.

What I found once the new direction was set is that the most valuable thing we were building was never the golf product. A sports-science record of how a body moves and learns doesn't stop at one sport — it carries across sports, and then out of sport altogether, into clinical use.

The answer came out of the same asset thinking that fixed the product. A coach can treat a student's learning history as an asset: who practises, who improves, who deserves their scarce time. My extension was the other half — the student owns one too. They should be able to export their learning and progress state (not the coach's instruction, which stays with the coach) and carry it with them.

Once a record of how a person moves and learns is portable, the market stops being golf:

How I know whether it's working

What this one taught me