What an AI Caddie Can Decide For You, and What It Can’t
Ninth hole at your place, medal round, and you’ve got 148 to a front-right pin. The phone in your pocket has already done the sums. Arccos says the shot plays 156 because you’re into a two-club breeze and the green sits eight feet above you, and it wants your 8-iron, which carries 152 on a good day. All of that is correct. None of it knows that your ball is sitting down in the first cut with a clump of wet rye behind it, that you’ve made one par in the last five holes, and that your last three 8-irons went 141, 138 and 149.
That gap is the whole story of the ai golf caddie app in 2026. The maths layer is genuinely excellent and getting better. The judgement layer is a stranger who has never seen you play.
The part the software actually wins
Yardage and dispersion are arithmetic problems with good data behind them, and computers eat those for breakfast.
Arccos Caddie is the strongest example. Screw fourteen sensors into your grips (about £200, plus a yearly subscription that has sat somewhere around £100 to £130 depending on the bundle), play twenty rounds, and it has a distribution for every club rather than the single fantasy number you quote in the bar. Its “Plays Like” figure folds in wind, elevation, temperature, humidity and altitude, which matters far more in Britain than people admit. A 150-yard 7-iron in July becomes a 143-yard 7-iron on a still 4°C January morning, because carry falls roughly two yards for every 10°F you lose. Elevation is worth about a yard for every three feet of climb. You are not doing that in your head on the tenth tee with a four-ball waiting.
Shot Scope goes at the same problem from a cheaper angle: passive tags, an X5 or V5 watch, and crucially no subscription at all. Its club averages screen is the single most useful piece of AI-adjacent kit most club golfers can own, and it looks something like this once you’ve banked a season:
Club Shots Avg carry Longest Shortest L/R spread
Driver 198 231y 258y 186y 58y
5 hybrid 104 196y 211y 163y 41y
7 iron 146 152y 164y 131y 24y
8 iron 131 142y 153y 120y 22y
PW 97 128y 139y 104y 19yLook at the 7-iron line. The average is 152. The shortest is 131. Your true two-thirds range is roughly 144 to 158, and the tail runs long on the short side because mishits lose distance and good ones don’t gain much. That asymmetry is exactly what a human caddie means by “take enough club”, and the app quantifies it better than any pro in a hut ever could.
Garmin’s Virtual Caddie on the S70 and the R10 ecosystem does a decent version of the same job, and 18Birdies and Hole19 will hand you a target line. Where the better tools have moved recently is aim point rather than club: Arccos will nudge you 11 yards left of a right-hand pin because your dispersion cone, not your intention, is what actually decides where the ball finishes. That advice is correct, boring, and worth about a shot and a half a round to a 15-handicapper who follows it. Arccos’s own marketing has claimed something over two shots of improvement in year one, which is a marketing number rather than a study, but the direction is real.
If you want the longer version of how these systems build those recommendations and where they fit into course strategy, the pillar on AI caddies and on-course decisions covers the decision architecture in more depth.
What it cannot see: the ball, sitting in something
Here is the hole in every caddie app on sale. Not one of them knows your lie.
Arccos assumes a fairway strike unless you manually tell it otherwise, and even then “rough” is a single bucket. A ball sitting up on top of trimmed semi is a flyer that might go 8% further with 1,200 rpm less spin. The same ball nestled down in wet autumn growth, grain against you, loses 10 to 15 yards of carry and comes out with no spin at all, so it releases 20 feet past the flag. Those two lies are eleven yards apart in the same 148-yard shot, and the app calls them both 156.
Bunkers are worse. A plugged lie in a rain-soaked greenside bunker at a links course in February is not a 22-yard shot, it’s an extraction. Hardpan under a tree, ball above your feet on a sidehill, a fescue lie where you can see half the ball: the software has no sensor for any of it. Uphill and downhill lies change your effective loft by several degrees and your low point by inches, and the only input the app has is a GPS elevation model of the green, not the four-degree slope you’re standing on.
Practical version: treat every distance recommendation as a fairway-lie number, then apply your own adjustment before you pull the club. Semi-rough sitting up, take one less and expect release. Sitting down in anything wet, take one more and accept you cannot stop it. Write your own correction factors down after a winter of paying attention. That’s the bit of the caddie job that stays yours.
What it cannot see: you, with a card in your pocket
Adrenaline is worth a club. Everyone knows this and nobody models it.
On the 72nd hole of your club championship your heart rate is 30 beats above your practice-round baseline, your tempo has quickened, and your 8-iron is now a 154-yard club rather than a 148-yard one. The app is still working from an average built mostly out of Wednesday afternoon roll-ups. It will hand you exactly the club that flies the green.
Then there’s the shot you cannot hit. Arccos will sometimes tell a 19-handicapper to hit driver over the corner of a dogleg because the expected-value maths says the reward beats the penalty at their dispersion. Fine, except the maths is built on 190 recorded drives, and it does not know that eleven of the last twelve times you stood on that tee with a hood-and-hit-it mindset you blocked it into the trees. The model averages your tee shots. It does not model your tee shots on that tee, in that wind, after a double.
Nerves, tempo, the shot you’ve lost confidence in, whether you can genuinely stop a ball on a firm green, whether you’ll chunk the 60-degree from a tight lie: all of it sits outside the input layer.
Which AI actually reads your footage correctly
This is where the money gets wasted, so be specific about what each type of tool measures versus infers.
Sportsbox AI 3DGolf builds a 3D model of your body from a single phone video, and on pelvis sway, thrust, lift and hand depth it is genuinely good. Its blind spot is the club. It is not measuring your face angle at impact, so any “diagnosis” it offers about ball flight is inference dressed as measurement. Frame it properly or the numbers are fiction: down the line, camera at hand height, tripod, full body in shot, 8 to 10 feet back, 60fps minimum.
Frame rate is the thing nobody checks. Your clubhead is travelling around 100mph through the ball, which is roughly 1.5 metres of travel between frames at 30fps. There is no impact frame. There is a blur and two adjacent lies. Shoot at 240fps on any recent iPhone or Galaxy and the auto-tracing in V1 Golf or Onform suddenly means something.
HackMotion sits in a different category because it measures rather than estimates: a wrist sensor reading flexion and extension directly, which is the single most reliable AI-adjacent read on face control that a club golfer can buy. Budget launch monitors split the same way. A Garmin R10 at around £500 gives you carry numbers on driver and mid-iron you can plan with, while its wedge spin figures can wander by well over 1,000 rpm because spin is being modelled from radar, not seen. Rapsodo’s MLM2 PRO photographs a marked ball, so its spin is closer to real, and camera-based units like SkyTrak+ are better again at three or four times the price. Any app that reads your phone video and confidently tells you your club path is 4.2 degrees out to the right is guessing.
The three things worth practising because of all this
Calibrate your rough. Take twenty balls, drop them in your own club’s semi from 140 yards, hit 8-iron, and log carry with whatever monitor you own. Now you have the number the app doesn’t.
Gap properly in cold weather. Most UK golfers’ distance charts were built in August and applied in January, which is a two-club error hiding in plain sight.
Rehearse the shot the app keeps asking for. If Arccos wants a 40-yard bunt off the tee on your par-5 fifth every single week and you don’t own that shot, the recommendation is worthless until you build it. Fifty balls with a 5-iron off a tee peg, aiming at a specific tree, is the most valuable practice session available to a 14-handicapper.
Next Saturday, try this: take the number, then say out loud what the lie is doing to it and whether you’d back yourself to hit the shot twice out of three. If the answer’s no, play the one you’d hit six times out of ten and let the app sulk.