What to look for in an AI golf coach: a 12-point checklist

Every golf app added an AI feature in the last two years, and almost all of them describe it the same way: a coach in your pocket, personalised, always available. The descriptions are close to identical, which makes them useless for choosing between the tools.

So here is a grid instead. Twelve criteria you can run against any product in the category, phrased so that each one has an observable answer rather than a marketing answer. Most can be checked in a free trial in under ten minutes.

Full disclosure: this is published by FlyAway, which sells an AI golf coach. We've written the checklist we'd want someone to apply to us, and at the end we run it against our own product and mark where we lose points.

The 12 criteria that separate a real AI coach from a chatbot

1. Does it see your actual data, or only what you type?

This is the criterion that splits the category in half, and it is the one to check first.

A chatbot with a golf prompt knows only what you type into it. It can explain what causes a slice, and it will do that well — large language models are genuinely good at golf theory. What it cannot do is know that you have hit 38% of fairways over your last ten rounds while your putting is fine, because nobody told it.

A coach connected to your data starts from evidence. The test: does the tool's first message reference something you did not tell it in that conversation?

2. Does it produce a plan, or a list of tips?

Five suggestions are not a plan. A plan has an order, a duration, and a reason why the first item is first.

This matters more than it sounds. The hardest decision in practice is not what could I work on — a golfer with a handicap knows five things wrong with their game — it is what should I work on for the next three weeks. A tool that returns an unranked list has handed the only difficult part back to you.

3. Does it track whether the plan worked?

Advice without a follow-up measurement is entertainment. The question to ask is whether the tool measures the same thing again after the practice block and tells you plainly whether it moved.

Fewer do it than claim to, though some now do: GolfFix issues monthly progress reports on which faults improved, and Otto positions you over time against golfers on your handicap. It remains the most common gap in the category, because closing the loop requires a tool to risk being shown wrong.

4. Does it weight advice by where you actually lose strokes?

A swing analysis will find something. It always finds something — there are dozens of named faults and every amateur swing contains several. That does not make the most visible one the most expensive one.

If your scoring problem is three-putting and greenside contact, a steep downswing is real but it is not where your strokes are going. A coach that reads swing findings and scoring statistics together can make that distinction. One that only sees your swing video cannot.

5. Does it separate what it measured from what it estimated?

Not all numbers in a swing report deserve equal trust. Timing measurements from video are reliable; clubface angle at impact from the same video is an estimate, and often a rough one. We've written about where those error margins actually sit in detail.

A tool that presents tempo and face angle with the same confident decimal places is concealing the distinction that determines which numbers you should act on.

6. Does it decline to answer when it cannot measure something?

The corollary of criterion 5, and the strongest single quality signal in the category.

If you filmed face-on, certain metrics genuinely cannot be recovered — head rise and spine angle need a down-the-line view. A well-built tool returns those as unmeasurable and excludes them from any overall score. A weaker one produces a plausible number anyway.

A system that always fills every field is not more capable than one that sometimes declines. It has the same limits and is less honest about them.

7. Does it know the course you are actually playing?

Course strategy is worthless in the abstract. "Favour the left side off the tee" only means something on a specific hole with a specific hazard at a specific distance from your specific tee.

Check whether the tool knows the hole you are standing on, and whether its advice changes between two different holes. Plenty of course management advice is generic enough to apply to any golf hole ever built, which is another way of saying it applies to none of them.

8. Does it handle the Rules of Golf correctly, and show its source?

This is the cheapest test in the list and one of the most revealing, because rules questions have verifiable answers.

Ask about relief from a penalty area, or the procedure for a provisional ball. A grounded tool gives the correct procedure and can point at the rule. An ungrounded one produces something that reads correctly and is wrong in a detail that would cost you a stroke. Try two or three questions where you already know the answer.

9. Is the feedback measurable?

"Turn more through impact" cannot be verified. "Get your tempo ratio from 2.4 to 3.0" can be, and so can "raise fairways hit from 38% to 50% over ten rounds."

The difference is whether you can tell, in four weeks, if you did it. Advice you cannot pass or fail is advice you cannot act on with any confidence.

10. Does it keep enough history to show a trend?

Golf improvement is visible over months, not sessions. Any single round is mostly noise.

A tool that shows you the last five rounds and discards the rest cannot demonstrate whether anything it told you in March has worked. Check how far back the history goes and whether you can see a metric plotted over a season.

11. Is it honest about what extra hardware would add?

Almost every AI coach in this category runs on a phone alone, and that is a legitimate design. The signal is whether the tool tells you what it would gain from more.

A launch monitor measures club and ball data properly. A dedicated 3D system such as Sportsbox 3DGolf resolves body positions far more precisely than a single ordinary camera can. A tool that acknowledges those gaps is more trustworthy than one implying a phone camera measures everything.

12. Can you get your data out?

Three seasons of rounds is real work. If there is no export, switching tools means starting from zero, and you should know that before you invest the seasons rather than after.

Quick comparison: what the main AI tools in this space actually do

The useful distinction between products here is scope — which part of the problem each one solves. Most are narrower than their marketing suggests, and narrow is not a criticism: a tool that does one thing properly beats one that does five badly. Listed alphabetically, ours included.

Tool	What it centres on	Strongest at	Where it stops
Arccos	Sensor-based shot tracking with a strokes-gained model	Data quality: automatically captured shot data is more reliable than self-reported entries	Requires sensors, and is analytics rather than technique instruction
FlyAway	Swing, scoring statistics and on-course strategy in one assistant, plus a tagged library of professional video	Cross-referencing swing findings against scoring data, so the drill you get is weighted by where strokes are actually lost	Breadth costs depth: the swing analysis is less precise than a dedicated 3D system, the shot data less reliable than a sensor set, and the video is curated from third parties rather than produced in-house
General assistants (ChatGPT and similar)	Golf theory on demand	Explaining concepts, and free	No access to your data, so nothing it says is specific to you — covered here
GolfFix	Swing video analysis	Focus — the whole product is pointed at one problem	Does not extend to scoring statistics, strategy or on-course use
Otto (Hole19)	AI analysis of your played rounds	Benchmarking you against golfers on the same handicap, which tells you whether a weakness is normal for your level or genuinely yours — no other tool here does that	No swing video; reports are bought in packs rather than running continuously, and it launched iOS-only
Teechgolf	A GPS scorecard and statistics feeding AI-built training plans, with a video drill library	Turning tracked round data into a structured plan of measurable drills, with the drill videos produced for the app and wired into those plans	Does not analyse video of your own swing

Read that table against criteria 1 to 4 and the pattern is clear: most tools do one column well. The question worth asking is which column your actual problem sits in.

Red flags: 5 signs an "AI coach" is a wrapper around a chatbot

1. It never references your data unprompted. If everything it knows came from your last message, it is a chat interface with a golf-themed system prompt.
2. The advice would be identical for any golfer. Ask it something specific about your game. If the answer would suit a 5 and a 25 handicap equally, it is not personalised.
3. Every field is always filled. See criterion 6. Perfect coverage means it is estimating where it cannot measure.
4. It is confidently wrong on the Rules. A rules answer that is nearly right is the clearest sign of ungrounded generation.
5. There is no follow-up. New advice arrives every week and nothing is ever measured again. The loop never closes.

Running the checklist on FlyAway

It would be a poor checklist if we did not submit to it, so here is where we land — including the parts we lose.

Where we do well: our AI coach reads your scorecards, statistics and swing reports rather than only your messages (1, 4), produces sequenced programmes rather than tip lists (2), and returns metrics as not applicable when the camera view you filmed cannot support them rather than estimating them (5, 6). Our swing analysis screens for 26 named faults with a severity and the exact frame range for each, which makes the feedback checkable (9). Ball-flight tendency is reported as a tendency with a confidence level, at roughly 75% accuracy, rather than dressed up as a measurement (5).

Where we lose points, plainly: if what you want is maximum measurement precision on the swing itself, a dedicated 3D system beats us and we have said so before (11). If you want automatically captured shot data with no manual entry, a sensor system is more reliable than any phone-based tool including ours (1).

On video we take a different route from Teechgolf, and the trade-off is real. Its drill library is produced for the app and wired into its own training plans, so the clip matches the instruction exactly. We curate and tag existing videos from professionals, which buys a much larger library at a cost in precision: retrieval works off those tags, so the clip attached to a drill is sometimes less on-point than it should be. If having video built to fit the lesson matters to you, that is a genuine advantage over what we do.

The category is young enough that no tool passes all twelve. The point of the grid is not to find a perfect product; it is to know which of the twelve you are giving up, and to make that a decision rather than a surprise.

---

Run the checklist on whatever you are using now. If it comes up short on criteria 1, 3 and 4 — the ones about your own data and closing the loop — that is the gap worth fixing, whichever tool you fix it with. Ours is here if you want to start with it.