AI Caddies and On-Course Decisions

A golf ball doesn’t care that your app suggested 7-iron. It cares about clubhead speed, strike location, spin loft, and the 12mph breeze quartering off your left shoulder. The gap between those two things is where most AI golf tools fall over, and it’s also where the good ones earn their subscription.

This page is about two connected problems: which AI tools actually read your phone footage correctly, and which ones give you a number on the tee that you can trust. They’re connected because a caddie recommendation is only as good as the yardage model behind it, and that model is built from your own strike data. Feed it rubbish, get rubbish back. Feed it forty honest 7-irons and it starts to outperform your gut.

What an AI Golf Caddie App Actually Computes

Strip away the branding and every ai golf caddie app on the market runs a version of the same pipeline:

1. Establish your dispersion pattern per club (carry mean, carry standard deviation, lateral standard deviation)
2. Model the hole geometry, including hazard polygons and green contour where available
3. Simulate a few thousand shots from your current position with each club in the bag
4. Score each outcome against an expected-strokes table
5. Return the club with the lowest expected strokes, not the highest probability of a good result

Point five is the one club golfers misread most often. Arccos Caddie will regularly tell a 14-handicap to hit 5-wood off a tight par 4 rather than driver, not because 5-wood is more likely to find the fairway in absolute terms, but because the driver’s miss distribution overlaps a penalty area that costs 1.4 strokes when found. The app isn’t being timid. It’s doing arithmetic you can’t do standing over the ball.

Here’s what that arithmetic looks like in practice. Take a 385-yard par 4 with out of bounds down the right, starting 240 yards from the tee and running the length of the hole.

The 5-wood wins by 0.13 strokes over driver. Across eighteen holes that’s roughly two shots a round if the same shape of decision recurs four or five times. It’s also close enough that the 4-iron and driver are nearly equivalent, which tells you something useful: the penalty for being aggressive is much smaller than the penalty for being timid to the point of leaving 200 in.

Most of these apps won’t show you that table. Arccos gives you the club and a one-line rationale. Shot Scope’s dashboard gets closer to the underlying numbers, though the strokes-gained breakdown sits in post-round analysis rather than on the course. Golfshot’s Caddie feature is essentially a yardage-to-club lookup dressed up with confidence language.

Which Tools Read Phone Footage Correctly

This is the part where marketing copy and reality diverge sharply. Pose estimation from a single phone camera is a genuinely hard computer vision problem, and the tools fall into three tiers.

Tier one: measures real 3D kinematics. Sportsbox 3D Golf is the only consumer phone app doing genuine 3D reconstruction from a 2D video, and it needs a specific setup to do it: 120fps or 240fps, camera at hip height, 8 to 10 feet away, full body and club in frame throughout. Under those conditions its pelvis rotation and thorax bend figures track within a few degrees of marker-based motion capture. Outside those conditions the numbers degrade fast and it doesn’t always tell you. If your phone is propped against a range bucket at knee height, the hip sway number is fiction.

Tier two: 2D angle measurement with good tracking. Onform, V1 Golf, Hudl Technique. These draw lines and angles on frames you select. The tracking is reliable because it’s essentially assisted manual annotation. No AI is inferring depth, which means no AI is inventing depth. For down-the-line shaft plane and face-on head movement these are perfectly adequate, and for 90% of what a club golfer needs to see, adequate is the correct standard.

Tier three: LLM-generated swing commentary. A growing number of apps take a few frames, run them through a vision model, and return paragraphs about your swing. The output reads plausibly. It is frequently wrong. Ask one of these to analyse a swing where the club is clearly shut at the top and there’s a reasonable chance you get back advice about strengthening your grip. Treat text-only output with no measured numbers attached as entertainment.

A practical test for any tool claiming to read your footage: film the same swing twice from the same position and run both through. If the reported hip turn differs by more than 4 or 5 degrees between two recordings of one swing, the measurement isn’t precise enough to track change over a season. Do this before you build practice around anything the tool says.

The frame rate problem nobody mentions

At 30fps, a 90mph clubhead travels about 4.4 feet between frames. You will not catch impact. You will catch a frame 2 inches before and a frame 2 inches after, and any “face angle at impact” figure derived from those is interpolated guesswork.

At 240fps that drops to roughly 6.6 inches per frame, which is enough to get a frame within an inch or two of impact. Every iPhone since the 8 shoots 240fps at 1080p. Most mid-range Android handsets do too. The setting is usually buried under Camera > Record Slo-mo, and it defaults to off.

Lighting matters more at high frame rate because exposure time per frame drops. Indoors under fluorescent lighting at 240fps you’ll get motion blur that defeats tracking, and you may also get banding from the mains flicker. Outdoors on an overcast day is fine. Outdoors in direct low sun with the camera facing into it is not.

Budget Launch Monitors: What They Measure and What They Guess

If you own a Garmin R10, a Rapsodo MLM2PRO, or a FlightScope Mevo, the crucial thing to understand is which numbers are measured and which are modelled. This directly determines what an AI caddie built on that data can be trusted to say.

The Garmin R10 is a doppler radar unit. It genuinely measures ball speed, launch angle, launch direction, and clubhead speed. Spin rate is estimated from a model, not measured, and that estimate is weakest on wedges and on shots hit with unusual strike locations. Carry distance is therefore derived from one measured number and one modelled number. Indoors into a net, the R10 has too little ball flight to observe and leans harder on its model still.

The Rapsodo MLM2PRO uses dual cameras plus radar. With the marked balls it measures spin optically, which is a genuine step up over the R10 for wedge work. Without the marked balls it estimates, and the difference in reported spin on a 56-degree wedge between marked and unmarked can be 1,500rpm or more.

Mevo Plus measures spin optically with the metal dots applied. Without dots, modelled.

Why does this matter for on-course decisions? Because your carry distance table is the input to every caddie recommendation you’ll ever receive, and a 7-iron that the monitor says carries 162 but actually carries 154 because the spin model is running low will put you short of every pin for a season. The fix is straightforward: validate one club outdoors against a laser measurement. Hit ten 7-irons on a flat range with marked distances, laser the pitch marks, compare the mean against what the monitor reported. If there’s a systematic 6-yard gap, apply that offset to your whole iron set and re-check with a second club.

Validation log, 7-iron, 14 Sep, 8mph crosswind
Monitor reported carry (yds):  164 158 161 166 159 163 157 162 160 165
Lasered pitch mark (yds):      156 151 154 158 152 156 149 155 153 157
                               ---------------------------------------
Mean reported: 161.5   Mean actual: 154.1   Offset: -7.4 yds (-4.6%)

Re-check, 5-iron:
Mean reported: 183.2   Mean actual: 175.0   Offset: -8.2 yds (-4.5%)

Consistent percentage offset across two clubs means the model is running hot, not that one club is mis-measured. Apply the correction globally.

Building a Yardage Book Your Caddie App Can Use

Arccos, Shot Scope V5, and Garmin Approach all build distance profiles from on-course data automatically. That’s better than range data in one important respect: it includes your bad strikes. Range data from a launch monitor session usually gets mentally filtered, and the app has no way to know you topped three.

But automatic on-course capture has its own failure mode: it doesn’t know about elevation, lie, or wind on any given shot, so your recorded “7-iron: 158 yards” bucket contains uphill shots from the rough into a breeze alongside downhill shots off a tight lie downwind. The mean comes out roughly right. The standard deviation comes out inflated, which makes the caddie more conservative than it should be.

Shot Scope handles this better than most because its shot classification lets you exclude specific shots from your profile after the round. If you shanked a wedge into a car park, mark it. Twenty seconds of admin keeps your dispersion honest.

The number to target: at least 30 recorded full shots per club before you take a caddie recommendation seriously for that club. Below 20, the carry standard deviation estimate is so noisy it’s barely better than a guess. Arccos will show recommendations from fewer, and it shouldn’t.

Here’s how a real profile looks after a season for a 12-handicap:

Club    Shots   Mean carry   SD carry   SD lateral   80% band
7i        67       152         8.1        11.4       142-162
8i        54       141         7.3        10.8       132-150
9i        71       129         6.9         9.6       120-138
PW        48       116         6.1         8.9       108-124

Read the 80% band, not the mean. The mean tells you where a good one goes. The band tells you what actually happens, and it’s the band the app is simulating against. To a 148-yard pin with water 8 yards short, a 7-iron whose bottom band edge is 142 puts you in trouble more often than the mean suggests.

The Wind and Elevation Question

Arccos Caddie pulls live wind from weather APIs and adjusts. This is better than nothing and worse than most golfers assume, because the wind at 10 metres above a weather station four miles away is not the wind at 30 metres above the 14th fairway at your club.

The adjustment models generally use something close to: carry change of roughly 1% per 1mph of headwind, and 0.5% per 1mph of tailwind. Headwinds hurt more than tailwinds help, because a headwind increases the ball’s spin-induced lift and balloons the flight.

For a 150-yard shot in a 12mph headwind: 150 × 0.12 = 18 yards, so you’re playing about 168. In a 12mph tailwind: 150 × 0.06 = 9 yards, playing about 141. Most club golfers underclub into wind and overclub downwind, and the asymmetry is exactly why.

Elevation is simpler and more reliable. The standard model is 1 yard of effective distance change per 1 foot of elevation change, accurate enough within about ±30 feet of elevation. A green 24 feet above the tee plays 24 yards longer. Garmin’s PlaysLike, Shot Scope’s elevation feature, and Arccos all implement this. Where they diverge is whether they combine it with wind correctly, and at least one of them applies elevation to the total distance rather than the carry, which systematically overclubs you on uphill approaches where the ball won’t run out anyway.

Whether the distance number feeding all of this comes from a GPS app or a laser matters more than the caddie logic on top of it, and it matters differently depending on the shot. That trade-off is worked through in detail on GPS App vs Laser Rangefinder When an AI Caddie Is Involved, including why a laser’s ±1 yard accuracy is sometimes worse in practice than a GPS app’s ±3 yards.

A Worked Round: Where the App Was Right and Where It Wasn’t

Rather than theory, here’s a decision-by-decision on four holes from a round played off 11.

Hole 3, par 4, 402 yards, drive finished 168 from the flag, back pin, bunkers front-left and front-right, ball slightly above feet in first cut.

Arccos said 6-iron. Player’s 6-iron mean carry is 163, 80% band 153 to 173. Pin was 168 with 12 yards of green behind it. The recommendation is right on the mean but the band means a good third of outcomes are short of the flag and into the bunker complex if they leak. Ball above the feet promotes a draw, which the app doesn’t know about. Player hit 6-iron, pulled it 8 yards left, short-sided in the left bunker, made 5.

The app was arithmetically correct and situationally incomplete. The lie adjustment is the thing no consumer AI caddie currently handles, and it’s a bigger factor than wind on approach shots.

Hole 7, par 5, 498 yards, 232 to the front after a good drive, green fronted by a burn at 218.

App said lay up to 95 yards. Player wanted to go. The 3-wood carry mean is 219, which is one yard past the burn. Read that again: the mean carry lands in the water or a yard beyond it. The app’s recommendation wasn’t conservative, it was the only defensible option, and the honest version of that advice would have been a single sentence showing the 219 against the 218.

Laid up, wedge to 14 feet, two-putt for 5. Fine.

Hole 12, par 3, 141 yards, 11 feet uphill, 9mph off the right.

Playing distance after elevation: 152. The crosswind at 9mph adds perhaps 2 yards of effective distance and roughly 6 yards of drift. App said 7-iron and aim 5 yards right of the pin. Player’s 7-iron mean is 152. Hit it to 22 feet right of the flag, which is what happens when you aim right and the wind doesn’t push it as much as modelled. Two-putt, par.

The club was right. The aim adjustment was a fraction aggressive, which is typical: wind drift models tend to assume a higher ball flight than most mid-handicap iron shots actually produce.

Hole 17, par 4, 339 yards, driveable-ish, OB left from 250, bunkers at 270 right.

App said 4-iron to 120. Player’s driver mean carry is 241 with a lateral SD of 26. Simulate that and roughly 11% of drives find OB, 9% find the fairway bunkers, and about 18% end up in a position where you’re chipping from 40 yards with a chance of birdie. Expected strokes for driver came out 4.11 versus 4.02 for the 4-iron. Genuinely marginal, and this is the category of decision where the app should defer to how you’re swinging that day rather than the season-long distribution.

What to Practise Based on What the Tools Tell You

The reason to go through all of this is that dispersion data points at practice priorities far more precisely than “hit more balls.”

If your carry SD on mid-irons is above 9 yards, your strike location is the problem, not your swing shape. Face-sole spray applied to an 8-iron over ten shots will tell you in four minutes what a season of video won’t. Heel and toe strikes cost ball speed via gear effect and smash factor; a 7-iron struck half an inch toward the toe typically loses 4 to 6 yards. Practise with a foot spray pattern and a target of 8 of 10 within the middle third.

If your lateral SD is above 14 yards on mid-irons but carry SD is under 7, your strike is fine and your face control isn’t. This is where the 2D video tools earn their keep. Face-on at 240fps, look at where the hands are at impact relative to address. If the handle has moved back or the left wrist has cupped, the face is open regardless of what the top of the backswing looked like.

If your driver’s lateral SD is above 28 yards, the caddie app will spend the entire round steering you away from driver, and it’ll be right to. Gaining 15 yards off the tee is worth roughly 0.3 strokes per round. Cutting lateral SD from 30 to 22 is worth about 0.9. Nobody sells lessons on the second one and it’s three times the value.

Wedge distance control from 60 to 100 yards is measurable with any of the budget monitors and is where mid-handicaps leak most. Build a three-length system (arms to hip, arms to chest, arms to shoulder) with one wedge, record ten shots at each, and write down the means. Typical result: a 12-handicap’s three gaps come out at something like 52, 71, and 88 yards, with a SD around 5 yards at each. Then repeat with the gap wedge. Six reliable numbers between 50 and 110 yards removes more strokes than any driver work.

Something worth noting about how the caddie apps treat wedges: most of them use your recorded mean and assume a symmetric distribution. Wedge distributions are not symmetric. Your misses are almost all short, because fat and thin both lose distance while only a flush strike gets the full number. If your app says 54 degrees and the pin is 8 yards from the front edge, you should be taking one more club’s worth of commitment than the app suggests.

Setting Up a Filming Routine That Produces Comparable Data

Video you can’t compare across weeks is video you’ve wasted. Four things to fix:

Camera position gets marked and reused. A strip of tape on the range mat and a tripod at a measured height. Down-the-line: camera on the target line extended behind the ball, lens height at hand height at address, 9 to 11 feet back. Face-on: perpendicular to target line, lens at sternum height, same distance. Write the tripod leg setting down.

Frame rate set to 240fps and left there. Check it every session because iOS camera settings reset after some updates.

Same club, same ball, same shot every time you’re collecting a baseline. A 7-iron off a flat lie to a target 150 out. Three swings, keep the middle one.

Clothing that makes the tracking work. Sportsbox and every pose estimation tool struggle with baggy waterproofs and struggle badly with dark clothing against a dark background. A fitted shirt in a colour that contrasts with the range netting genuinely improves the numbers you get back.

One more thing about interpreting what comes out. When Sportsbox reports your pelvis has 42 degrees of rotation at the top and the tour average is 45, that gap is not a swing fault. Tour averages are the output of a population of people who hit it 300 yards, not a target for someone playing off 14 in a mid-morning fourball. The useful comparison is against your own footage from March, on the same club, from the same tripod position.

Where AI Caddie Recommendations Fail Predictably

Four scenarios where you should override the app without hesitation:

Lie quality. No consumer app knows you’re in a divot, sitting down in the second cut, or on a perfect flyer lie in light rough. A flyer from 140 with a 9-iron can go 155. The app will say 8-iron for 140 and you’ll be over the back.

First tee and first approach of the round. Your season-long distribution assumes a warmed-up golfer. If you’ve come straight from the car park, take one more club on the first three approaches and expect the strike to be low-face.

Greens that are running fast and firm. Caddie apps optimise for proximity to the hole, not for leaving the correct putt. On a firm green with a front pin and a back-to-front slope, the app’s “hit it to 15 feet” recommendation may be pointing you at a 15-foot downhill slider when a 25-foot uphill putt from below the hole is a better outcome. Expected-strokes tables are built on average green conditions.

Matchplay. Every bit of the maths above optimises for stroke average. In matchplay, hole-by-hole win probability sometimes calls for a shot with a worse expected score and a fatter right tail. Two down with four to play is exactly when you ignore the 5-wood recommendation.

A Four-Week Block That Uses All Of This

Week one: collect. Ten shots each with 7-iron, 9-iron, pitching wedge and driver on the monitor, plus a down-the-line and face-on video of the 7-iron at 240fps. Validate one club against a laser outdoors. Record the offset.

Week two: diagnose. Calculate carry SD and lateral SD per club from the week one data. Whichever club has the worst ratio of SD to mean carry, that’s your priority. Apply face spray and shoot ten more with that club, counting centre strikes.

Week three: one intervention only. Pick a single change that addresses the diagnosis, work on it for three sessions, and film the 7-iron from the marked tripod position at the end of each. If the tool you’re using reports a kinematic number, check whether that number has moved more than its own repeatability threshold.

Week four: re-measure with the exact protocol from week one, same clubs, same count. Compare carry SD, not mean carry. A drop from 9.2 to 7.1 on your 7-iron is worth more than a 4-yard gain in mean distance, and it’s the number your caddie app will quietly reward you for on every approach shot for the rest of the season.

Play three rounds with the app recording and the caddie feature on, and take its recommendation on every tee shot even when you disagree. Three rounds isn’t enough to prove anything statistically, but it will show you the specific holes where your instincts and the arithmetic diverge, and those holes are where the strokes live.