How to Calculate AI ROI When Your Gut Says It's Working but Your Books Don't Show It

Seven in ten small businesses now use AI regularly, according to the 2026 Intuit QuickBooks AI Impact Report. Forty-three percent of those owners say revenue went up. Almost none of them can show a lender exactly where.

That gap is not a technology problem. It is a measurement problem, and it starts the moment you ask the wrong question. "Has AI helped my business?" is a firm-level question. Too many variables move inside a business over any given quarter — client mix, pricing, seasonality, staff changes — for a before-and-after comparison at that level to mean anything auditable. The question that produces a number a bank will accept is narrower: has AI changed the time and error rate on this one recurring task, and by how much?

Pick one task before you touch any tool

The OECD's 2025 SME digitalisation report found that among small businesses already using AI, 43% point to reduced manual tasks as their primary observed gain. Only 6% describe the impact as transformational at the firm level. That spread is not a coincidence. Task-level signal is visible where firm-level signal disappears into noise.

So start there. Pick one recurring task you perform at least weekly. Invoice processing, social media content, customer inquiry responses, expense categorization. The task needs to be discrete enough that you know when it starts and when it ends. "Marketing" is not a task. "Writing and scheduling three Instagram posts per week" is.

Write down, before you open any AI tool: how long does this task take per session, how often do you catch errors in the output, and whether this task connects to a revenue line you track. That last point matters more than most owners expect.

Record the baseline like someone will audit it

Spend two to four weeks logging the pre-AI version of the task. Time each session with a stopwatch, not a guess. Note error catches. If the task produces a deliverable that feeds revenue — a proposal, a listing, a content post — note the conversion rate or the downstream outcome.

This is the part where most AI measurement fails before it starts. Owners skip the baseline because they are eager to run the tool, and then they have nothing to compare against. Adobe's Small Business Superpower Study reports that owners using AI for social media content save 175 hours per year in that single workflow. That figure is credible because it is scoped to one workflow. It is not "175 hours saved across everything." It is one task, measured specifically.

Your baseline does not need to be statistically perfect. It needs to be written down, dated, and consistent enough that someone else reading it could reproduce the measurement.

Run the experiment for ninety days, not ninety minutes

The CFIB's 2025 digital transformation reports document that Canadian SMEs see measurable ROI from digital tools in as little as two months when the scope is narrow. Ninety days gives you enough cycles on a recurring task to separate a genuine pattern from a good week.

During the experiment, log the same metrics you logged at baseline. Time per session. Errors caught. Downstream outcomes if applicable. Track what you spend on the AI tool, including any subscription tier you upgraded to for this purpose. Do not expand the experiment to a second task mid-run. The moment you add scope, you lose the clean comparison.

One thing worth naming: you will get faster at using the tool over the ninety days. That learning curve is real, and it means your week-twelve numbers will look better than your week-one numbers for reasons that have nothing to do with AI's ceiling. Log weekly, not just at the end, so you can see the curve and account for it when you present the results.

Why ninety days on one task still leaves the ROI number open to challenge

The productivity paradox literature cited throughout the research on this topic makes a pointed argument: faster work does not automatically produce better output or higher profit, and some apparent gains reflect reclassification of effort rather than new economic output. Applied to this method, the critique runs like this — any ninety-day window contains seasonal variation, a natural learning curve, and concurrent business changes, none of which pause for your experiment. The ROI number at the end attributes the full time difference to AI. The productivity paradox says that attribution is not safe.

This is a real problem. It is not a fatal one.

The single-task audit does not claim to eliminate confounding. It claims to produce a number specific enough to be tested again. A lender who doubts your ninety-day result can ask: did the same task, run for another ninety days, produce a similar result? A firm-level productivity claim cannot be tested that way at all. The CFIB data showing returns in as little as two months came from tracking tool-specific outcomes, not firm-level performance. If task-level confounding were severe enough to mask signal in ninety days, that figure would not exist.

Present your number as what it is: a task-specific observation over a defined period, not a proof of AI's value to your entire operation. That framing is more conservative and more credible than a broad claim, which is exactly what makes it useful with a lender.

Calculate net ROI with three inputs

At the end of ninety days, you need three numbers. First, hours recovered per week: your baseline average time per session minus your post-AI average, multiplied by weekly session frequency. Second, your effective hourly rate: what you would pay someone else to do this task, or what your own time costs if you bill by the hour. Third, total tool cost over ninety days, including any setup time you spent at your hourly rate.

Net ROI equals hours recovered per week, times thirteen weeks, times your hourly rate, minus total tool cost.

The CFIB reports put the average return from digital tools at $1.60 per $1.00 invested across Canadian SMEs, with digital leaders reaching $2.40 per $1.00. Your task-level number will land somewhere in a range. What matters is that it is yours, it is dated, and it is traceable to a log a lender can read.

If the task also connects to a revenue outcome — a proposal acceptance rate, a content post that drives inbound inquiries — add that figure separately. The Adobe study found that 47% of small business owners using generative AI report revenue increases averaging 21% since adoption. Those owners are not reporting firm-level magic. They are reporting what happened downstream of specific workflows. Your audit produces the same kind of evidence, one task at a time.