Nearly nine in ten small businesses now use AI. About three in ten report major productivity gains. That’s a 59-point gap, and I keep coming back to it because it’s the honest version of the AI story for small businesses. Almost everyone has the tool. Most can’t show what it earned.
The obvious explanation is that the gains are just delayed. I don’t buy it. The gap has held long enough to be structural, not a timing problem.
Where the naive ROI math goes wrong. Say you price out an AI assistant at the raw API bill — the metered cost of the AI provider’s processing. Run it on one process, and it’s a rounding error. Then the other costs show up: monitoring so you can see what it’s doing and what it costs, hosting, integration work, maintenance, and retraining when the model starts behaving differently than it did in week one. Research puts those at 40–60% on top of visible API spend. A pilot that looked like a 3x win on the invoice is suddenly a 1.7x win — and if you never measured the human baseline, you don’t know whether 1.7x covers the pain of changing how your team works. Payback periods are the honest metric here. Not capability. Not enthusiasm.
What the ones that paid back had in common. When I went through the documented deployments, the mechanics weren’t clever. One bounded process, chosen for volume and repetition. A baseline measured before anything was switched on. A human still handling the exceptions — typically the AI takes 60–80% of the work, and a person owns the rest. And a proof window of 30 to 90 days, because a result that only appears in month seven never gets funded. The payback numbers on the good ones I saw clustered at under two months — one of them as quick as 1.4. Nobody wins on a smarter model. They win on a narrower question.
The Kolkata Test. I have a name for what kills the rest. Whether an AI tool survives contact with ground-level friction: an internet connection that drops, an accounting system that’s a pirated 2014 copy, an operator who speaks Hindi and an interface that only speaks English, a power cut in the middle of a process. Most tools assume reliable internet, cloud-first everything, English-native users, modern software, and stable electricity. For most small businesses here, none of that holds. This is where the gap actually lives. Not in the model’s benchmark score.
What I’d do with this. Measure the process before you touch it — hours, cost per transaction, error rate. Pick one process. Run the tool beside the person, not instead of them, for a full month. Watch the hidden costs as carefully as the API bill. Then decide, on numbers, whether to keep going. The 59-point gap isn’t an AI problem. It’s an accounting problem, and accounting problems are solvable.