How AI Caddies Actually Work (And Their Limitations)
Every golf app now ships an "AI caddie" that tells you which club to hit and where to aim. Underneath the branding sits a fairly simple piece of math. Understanding it tells you exactly when to trust the recommendation and when to overrule it.
An AI caddie is not reading your swing or watching your ball flight. When a golf app tells you to hit 7-iron and aim eight yards left of the pin, that recommendation came out of three models stacked on top of each other, and none of them involve artificial intelligence in the way the marketing implies. It is expected-value math running over a map.
That is not a criticism. The math is sound and it beats how most amateurs choose a club. But knowing what the calculation actually contains tells you which recommendations to follow and which ones to ignore, and that distinction matters more than the feature itself.
The three ingredients behind every recommendation
Strip the branding off any AI caddie and you find the same three components.
A course model. Satellite imagery and GPS mapping turn each hole into a set of polygons: fairway, rough, bunkers, water, green, out of bounds. Better implementations add elevation change and green contours. This part is accurate, and it is where most of the engineering money goes.
A dispersion model for you. This is a probability cloud, usually an ellipse, describing where your ball tends to end up with a given club. It has a distance component (you carry your 7-iron 155 yards on average, with a spread) and a directional component (your shots scatter maybe 6% of carry distance side to side).
An expected-strokes table. For every combination of distance and lie, a number describing how many strokes it typically takes to hole out from there. A scratch golfer needs about 2.8 strokes from 150 yards in the fairway and about 3.0 from 150 in the rough. These tables come from the same body of work that produced strokes gained: Mark Broadie’s analysis of millions of shots, published in Every Shot Counts.
The algorithm, in plain terms
With those three pieces in place, the recommendation engine does something almost mechanical.
It picks a set of candidate plays: 7-iron at the pin, 7-iron at the center, 6-iron long-left, and so on. For each candidate, it drops your dispersion ellipse onto the course map at that aim point. Then it asks what fraction of that cloud lands in each region: 54% green, 21% front fringe, 14% right rough, 8% greenside bunker, 3% water.
Each region gets scored with the expected-strokes table. Green at 22 feet is 1.85 strokes. Greenside bunker is 2.55. Water is the drop plus the recovery, so maybe 3.6. Multiply each outcome by its probability, add them up, and the candidate has a single number: expected strokes to hole out.
Whichever candidate produces the lowest total wins, and that is the club and target you see on screen. Every AI caddie on the market is running some version of this loop.
What AI caddies get right
The approach has real strengths, and they are the strengths of arithmetic over instinct.
It never gets seduced by a tucked pin. The math weights the bad outcome by how bad it is, which is why these tools nearly always steer you toward the center of the green when the flag sits near an edge. Amateurs know this is correct and ignore it anyway.
It prices penalties properly. Human golfers dramatically underweight a 5% chance of water because 95% of the time nothing happens. An expected-value calculation does not have that blind spot, and this is where an AI caddie earns most of its strokes.
It handles lay-up distances well. Deciding whether to leave 95 yards in the fairway or 40 yards in the rough is a hard comparison to do in your head. It is trivial for a lookup table.
Where the model breaks down
The limitations all come from the same place: two of the three ingredients are guesses about you, and only the course map is solid.
Your dispersion is usually assumed, not measured
Most apps seed your dispersion from your handicap. Tell the app you are a 14 and it assigns you the scatter pattern of a generic 14-handicap. But two 14s can arrive at the same index by completely different routes. One hits it 250 and sprays it; the other hits it 205 down the middle. The correct club for those players from 150 yards is frequently not the same club, and a handicap-seeded model cannot tell them apart.
Even apps that measure your distances often measure the wrong number. Your average carry with a 7-iron is not the number to plan around. What matters is the carry you produce on an ordinary, slightly-thin, slightly-rushed swing, which sits closer to your 70th-percentile result than your mean. Club selection built on averages leaves you short about half the time.
It cannot see the shot in front of you
The model knows you are 152 yards out in the rough. It does not know the ball is sitting down, that you are eight inches below your feet, that the lie will kill your spin, or that the green is running firm this week. A human caddie reads all of that in two seconds. The app is working from coordinates.
Wind and elevation adjustments are estimates layered on estimates. An app reading a weather API knows the wind at the clubhouse, not the swirl in the chute you are hitting into.
It recommends the shot you should hit, not the one you can hit today
An expected-strokes model assumes you execute at your baseline. On a day when you have double-crossed three drives, your real dispersion is nothing like the ellipse in the app, and the aggressive line it recommends is simply wrong for that afternoon.
It optimizes an average, which is not always what you want
Minimizing expected strokes is the right objective across a season. It is not always right on the hole in front of you. One down with two to play, you need variance, not the lowest average. The model has no concept of match state.
There is no feedback loop
This is the deepest limitation. An AI caddie makes a recommendation and then never finds out what happened. It does not know whether you hit the green, whether its club was long, whether its aim point kept you out of trouble over fifty attempts. Advice without measurement cannot improve, and most golfers using these tools have no idea whether the recommendations have gained them a single stroke.
The part that actually makes recommendations good
Every weakness above traces back to the same missing input: a real record of what you do on a golf course. Not a range session, not a handicap index, but shot-by-shot outcomes from rounds you actually played.
That is the layer golfity is built on. Every shot you log has a start and end position, so your distances come from swings you made on the course with your bad ones included, and your strokes gained is calculated from those positions rather than inferred from your index. Nothing is estimated from your handicap and nothing needs extra hardware.
The practical difference shows up when you check a decision against your own data. From 150 yards, an AI caddie might tell you the center of the green is worth 0.12 strokes over the pin. Your own logged rounds can tell you something more useful: that you hit 41% of greens from 150, that your miss is short-right eleven times out of twenty, and that you are losing 0.3 strokes per approach from that bracket compared with your benchmark. The first is a model’s opinion. The second is your scorecard.
How to use an AI caddie well
Take the penalty math seriously, because that is the part these tools do better than you. Take the center-of-green bias seriously too, since the research behind it is solid.
Overrule the club selection when the lie, the wind, or your form that day is not something the app can see. And check the recommendations against outcomes: after twenty tracked rounds, your strokes gained by approach distance will tell you whether the advice is working or whether you have been playing to somebody else’s dispersion pattern all season.
A model that knows the course can tell you the hole is 152 yards. Only your own data can tell you what you actually do from 152. Start tracking your rounds and give the math something real to work with.
