Every OnCourtAI analysis comes with an AI coach you can talk to. You film a stroke, get your scores, and then ask whatever you like about them.

Since we launched it earlier this year, the patterns in how players use it have become clear. What they ask, when they ask it, and how they react to the answers says a lot about how people actually learn tennis — and about which parts of coaching an AI can take on, and which it cannot.

Who Uses It

Just over a third of players who upload a video go on to ask the AI coach something. That share climbs sharply with commitment: among players who have analysed the same stroke at least five times, three in four have used the coach.

Usage is uneven. Around 30% of players who try it ask a single question and leave it there. At the other end, roughly 7% have asked more than ten. It is a tool that a minority of players lean on heavily.

It is also global. The largest groups are players in the US, the UK, Australia and Germany, and around 7% of players ask in another language — Chinese, French, Portuguese, Spanish, Japanese, Russian and Vietnamese among them.

What Players Ask

Grouping questions by topic gives a clear picture. Questions can touch more than one area, so these overlap:

  • “Analyse my shot” — 29%. The single biggest category is simply asking the coach to walk through a video: “Analyse my latest forehand”, “How's my serve looking?”
  • How to improve, and what to practise — 27%. Drills, exercises and what to focus on next.
  • Timing and contact point — 12%. Hitting late, takeback timing, contact too close to the body.
  • What a score or metric means — 8%. “Why is my spin generation score so low?” “What is X-factor?”
  • Spin and grips — 7%. Topspin, flat hitting, eastern versus semi-western.
  • Body rotation and kinetic chain — 7%. Hips, shoulders, separation.
  • Serve mechanics — around 6%. Racket drop, trophy position, toss, pronation.
  • Equipment — 5%. Rackets, strings, tension.

Just as revealing is what players barely ask about. Questions about level or rating (“What's my NTRP?”) make up around 1%. Mental game and match preparation: about 1%. Injury and pain: under 0.5%.

Players treat the AI coach as a technical coach. They bring it their swing. They do not, so far, bring it their nerves, their match plans or their sore shoulders.

The Questions Match the Faults

The topics players ask about line up closely with the faults we measure most often. Around 58% of forehands we analyse are not struck clearly in front of the body — and timing and contact is the most-asked technique topic. Spin generation is the lowest-scoring forehand component on the platform, and spin is among the most-discussed. Most serves we analyse show little or no racket drop, and racket drop questions come up again and again.

We have written up each of those in our “How do I” series: hitting the ball late, getting more topspin, the X-factor and racket drop. The point here is broader. Players are not asking random questions. Once they can see their numbers, they go straight to the weakest ones.

Conversation, Not Lookup

If the AI coach were just a faster search engine, most conversations would be one question long. They are not. Nearly two in three conversations go beyond a single question, and some run to ten or more.

The follow-ups have a recognisable shape. Players ask what something means, then why they are scoring low on it, then how to fix it, then how they will know the fix is working. “How will I know if my timing is improving?” is one of the most common late-conversation questions we see.

There is also a clear shift over time. In the first couple of months, only around one in five questions was typed by the player; the rest were taps on suggested questions. By late summer that had risen to around one in two. Players are moving from accepting prompts to asking their own questions, in their own words — usually short ones of a sentence or so, answered in a short paragraph.

When They Ask

The AI coach typically replies within a few seconds, at any hour. Players use that.

Looking at UK players in local time, 23% of questions come between 6pm and 10pm, and 14% between 10pm and 7am — hours when no coach is taking calls. Questions are spread evenly across the week, too: weekends account for 28% of them, almost exactly their share of the days. There is no weekend rush; people ask when the thought occurs to them, usually just after they have looked at a video.

Seasonality is visible as well. Question volume rose roughly sevenfold from spring to a peak in July, then roughly halved by October as the outdoor season ended — the same curve as uploads.

Where Players Push Back

Not every exchange is smooth. Around 1% of questions are confusion or disagreement: “I don't understand what you mean by flat swing.” “I think that speed is lower than you wrote.” “My stance can't be that wide.” “I don't think I'm hitting too late — it's too early.”

These matter more than their share suggests. They come from players who know their own game, and they arise when a number from a phone camera conflicts with what they feel or know. Sometimes the player is right and the measurement is off. Sometimes the measurement is right and the feel is misleading — late and early contact genuinely feel similar from the inside.

Either way, the lesson is that players do not want a verdict; they want to understand where it came from. That is why we now flag measurements as unreliable when the camera angle cannot support them, rather than guessing. Trust is earned by showing your working.

Players rarely rate answers — only around 1% of replies get a thumbs up or down — though about 70% of those ratings are positive. The more useful signal is behavioural: most conversations continue past the first answer.

Does Talking to the AI Make You Better?

We would love to say yes. The honest answer is that we cannot show it yet.

We compared players who used the AI coach with those who did not, looking at strokes a player had analysed at least five times, and asked whether their latest scores beat their earliest. Players who had chatted improved 67% of the time; those who had not, 64%. Then we ran the test that matters: we shuffled the “chatted” label randomly between players with similar numbers of uploads, hundreds of times. A gap that big turned up by pure chance in roughly half of the shuffles. In other words, it is noise.

That is not evidence that the coach does nothing. The sample of players who analyse one stroke that often is still small, and players who chat differ from those who do not in ways that affect improvement in both directions. But it would be dishonest to claim the AI coach makes players improve faster on the data we have. What we can show, from earlier analysis, is that persistence — filming regularly over time — is the strongest predictor of improvement we have found.

How This Compares With Traditional Coaching

A lesson with a good coach and a conversation with an AI coach are not the same thing, and the data makes the differences fairly clear.

What the AI does well. It is available at 10pm on a Tuesday, replies in seconds, in the player's own language, and never tires of the fifth follow-up question about the same metric. It has the numbers from every shot to hand. And it is happy to explain the jargon — questions asking what a score means are exactly the kind people often feel awkward asking a coach twice.

What the coach does that the AI does not. A coach sees the ball, the court, the opponent and the score. They read frustration and confidence. They manage injuries, build match plans and keep players motivated through flat patches. The near-absence of mental, match and injury questions to the AI is telling: players already know where human judgement is needed.

The bandwidth gap. Coaches who use our platform can leave notes on their players' videos. Only around 2% of videos uploaded by players linked to a human coach have received a written note in the app. That is not a criticism — coaches give feedback on court, where it belongs, and their time is finite. But it shows the gap an AI can fill: the questions a player has between lessons, about a video filmed on a Thursday, when the next session is not until the following week.

How This Will Evolve

Based on what we are seeing, three shifts seem likely.

From answers to programmes. The most common follow-up is some version of “what should I practise, and how will I know it is working?” That points away from one-off answers and towards structured practice: a training plan built from your weakest scores, re-measured each time you film.

From prompted to self-directed. The rise in typed questions, from around one in five to around one in two, suggests players are getting more comfortable treating the AI as a conversation partner rather than a menu. Expect questions to get more specific and more personal over time.

From either/or to both. The most useful model is not AI instead of a coach, but AI between sessions with one. The coach sets the direction on court; the AI keeps the player on it during the week, answers the 10pm question, and gives the coach a measured record of what changed. The parts of coaching that are about the person — motivation, tactics, injury, confidence — stay human.

What We Have Learned

  • Players want their own data explained, not generic tips. “Analyse my shot” is the most common request by far.
  • They go straight to their weakest scores — and those match the faults we measure most often.
  • The value is in the conversation: nearly two in three exchanges go beyond the first question.
  • They use it when coaches are not available: evenings, late nights and in their own language.
  • They keep the human questions for humans.
  • Trust depends on transparency. When a number conflicts with what a player knows, they push back, and they are right to.
  • We cannot yet show that chatting with the AI speeds up improvement. Filming consistently still matters most.

If you want to try it, film a stroke — our filming guide covers the setup — and ask the coach the question you have been meaning to ask.