Three Numbers That Tell You if AI Is Working

Eighty percent of small businesses in the QuickBooks July 2026 survey now use AI regularly. Two years earlier, that figure was 48%. Somewhere in that acceleration, the question of whether AI is actually working got skipped.

The measurement problem hiding inside good feelings

Thryv's 2026 small business survey found that 83% of AI users say AI saves time, and 70% say it has increased revenue. Those numbers look like confirmation. They are sentiment — consistent, multi-country, directionally aligned sentiment from over 500 respondents across multiple surveys, but sentiment nonetheless.

The OECD's 2025 Digital for SMEs survey across ten countries found something that sits awkwardly next to Thryv's figures: while two-thirds of AI-using SMEs report at least moderate impact, only 18% report significant improvements in efficiency, productivity, or decision-making. The OECD's firm-level data separately shows AI adopters outperforming non-adopters by 4% to 15% on productivity. That gap between "moderate" and "significant" is not explained by AI underperforming. It is better explained by founders who feel gains but cannot see them clearly enough to act on them.

An NBER working paper on AI, productivity, and the workforce names this directly: a productivity paradox where perceived gains exceed measured ones, attributed to delays in revenue realization rather than absent results. The gains are real. They are just arriving later and in different forms than founders expect at the moment they fill out a survey.

Why 83% satisfaction still leaves you flying blind

Here is the problem with feeling like AI is working: you cannot tell whether your gains are growing, shrinking, or concentrated in tasks that stopped mattering six months ago.

A founder who says AI saves time cannot identify, from that answer alone, whether the savings are compounding or plateauing. A founder who says AI increased revenue cannot tell whether the increase came from AI-assisted lead follow-up, reduced outsourcing spend, or a seasonal shift unrelated to any tool. The UK Business Data Survey 2026 found that only 17% of AI-using businesses have any policy or guidelines around AI use. That governance absence is the same informal, untracked adoption pattern that produces "AI feels useful" without producing a number you can act on.

The systematic review that examined 68 peer-reviewed articles on AI adoption in SMEs explicitly flags performance measurement gaps as a core deficiency. Not a minor oversight. A core deficiency across the research base.

The three numbers worth tracking

The OECD's generative AI and SME workforce survey found that 45% of generative AI users report saving money and 26% report increased revenue. Thryv found that most AI users save between 11 and 40 hours each month. Those ranges are not vague — they are trackable.

Three metrics, logged manually once a month, cover the territory where AI impact actually lands in a small operation.

Time saved per task: pick two or three recurring tasks you now do with AI help. Log how long each takes. Compare month over month. QuickBooks found that nearly one in four AI users report shorter workdays. That is a real signal, and you can replicate it at the task level without any analytics tool.

Revenue from AI-assisted leads: if you use AI for follow-up emails, lead qualification, or outreach, tag those leads separately in your CRM or even a spreadsheet. Track what closes. The NBER paper identifies revenue-based productivity as the primary channel through which AI gains show up at the firm level. This metric points directly at that channel.

Cost reduction in outsourced services: if AI has replaced or reduced work you previously paid a contractor or agency to do, log the before and after spend. Thryv found that 55% of AI users say AI has reduced costs. That figure is self-reported. Your version should have a dollar amount attached.

What this checklist does not do

The systematic review is clear that AI outcomes in SMEs depend on organizational readiness and strategic alignment. A time log does not measure either. If your team distrusts AI outputs, or if the tools you are using do not connect to where the business is heading, these three numbers will look flat and you will not know why.

That is a real limitation. The checklist is an argument for a better checklist, not a complete measurement system.

Start with 30 minutes at the end of the month. Three columns in a spreadsheet. The OECD's 4% to 15% productivity premium for AI adopters exists at the firm level. You will not see it until someone at your firm is counting.