Measuring AI ROI When You Can't Prove It's Working

Near 60% of US small businesses now use AI tools, according to the US Chamber of Commerce. AIWorldMeter's 2026 SMB data shows high reported waste and low feature utilization sitting alongside that adoption number. Both things are true at the same time, which means a lot of owners are paying monthly subscription fees for tools they cannot evaluate.

The waste is not a mystery

When Brynjolfsson, Li, and Raymond studied 5,179 customer support agents with access to a generative AI assistant, they measured a 14% average productivity increase. The key word is measured. They tracked calls handled, resolution times, and output per agent. The owners in the AIWorldMeter data did not do that. They adopted tools, noticed nothing definitive, and kept paying.

The problem is not that AI tools fail to produce returns. Noy and Zhang's MIT experiment on professional writing tasks showed large time savings and quality improvements, with stronger gains for lower-ability workers. The returns exist. They are just invisible without a unit of measurement attached to a specific task.

Pick one workflow. Not your whole business. One task your team does repeatedly: writing product descriptions, answering customer emails, generating reports. Log how long it takes today. Log how many errors or revision cycles it produces. Log how many units get completed per week. That is your baseline. You need two weeks of it before you touch any AI tool.

What you are actually measuring

The SansaTech framework and AI-First approach both anchor ROI to labor cost conversion. The mechanism is straightforward: take the net time saved per week, multiply by the hourly labor cost of the person doing the task, and subtract the weekly cost of the tool. Net time saved is gross time saved minus rework time. If your AI writing assistant cuts drafting from four hours to two but adds 45 minutes of editing out AI errors, your net saving is 75 minutes, not two hours.

Error rate matters separately from time. If your customer support team previously resolved tickets incorrectly at a rate requiring follow-up calls, and that rate drops after AI deployment, the avoided labor cost of those follow-up calls is a second line item. Volume is the third: if the same headcount now processes more tasks per week, the marginal output per dollar of labor cost is your productivity ratio.

These three numbers — net time saved, error reduction, volume per labor dollar — give you a before-and-after comparison a spreadsheet can hold. Business.com's 2026 Small Business AI Outlook documents role-level time savings across SMB workplaces. Those figures only became visible because someone tracked them at the task level.

When tracking time doesn't fix a broken tool

Otis and co-authors ran a randomized controlled trial giving Kenyan entrepreneurs access to a GPT-4 business assistant via WhatsApp. They collected revenue and profit data. Low-performing entrepreneurs saw profit declines. The researchers measured the outcomes directly, which means absent tracking was not the cause of the damage.

The OECD Digital for SMEs 2025 report names the structural version of this problem: skills gaps and integration constraints in smaller firms create barriers that a before-and-after spreadsheet cannot resolve. An owner with weak business fundamentals who adds an AI tool to a broken workflow gets a faster broken workflow.

This is a real constraint. The thesis that measurement recovers positive returns assumes there are positive returns to recover. For some owners, the Kenya result is the relevant data point, not the Brynjolfsson customer support study. The honest version of this argument is that workflow-level tracking will surface a negative ROI number for some tools in some businesses. That number is still more useful than no number. An owner who runs a two-week comparison and sees costs rising has evidence to stop. An owner with no data has no such signal and keeps paying.

The AIWorldMeter 2026 data makes this concrete: high waste and low feature utilization are the dominant SMB pattern. That waste is not entirely explained by weak fundamentals. A portion of it is owners who never checked.

Connecting savings to a number that survives a tax audit

Take your net weekly time saving in hours. Multiply by the fully-loaded hourly cost of the employee doing the task, including benefits and overhead. Subtract the weekly tool cost. If the result is positive over a 90-day window, the tool is paying for itself on that workflow alone.

Avoided outsourcing is a second calculation. If you previously paid a contractor to write ad copy and now your in-house team handles it with AI assistance, the delta between contractor invoices before and after is recoverable ROI. Business.com's 2026 data captures this pattern across SMB workplaces where AI reduced reliance on external vendors.

Quality improvements are harder to convert. Noy and Zhang documented measurable quality gains in writing tasks, but translating "fewer revision cycles" into a dollar figure requires you to know what your revision cycle costs. If you don't, start by counting revision cycles per document before deployment. After deployment, count again. The ratio is a number you can defend.

The 90-day window matters because it eliminates novelty effects. Productivity often rises in the first two weeks of any new tool simply because people are paying attention. Brynjolfsson et al. found that novice agents gained more from AI assistance than experienced ones, which suggests early gains can be real but also that they may compress over time as users plateau. Measuring across 90 days captures the post-novelty baseline.

If your three-month calculation produces a negative number, you have two options: diagnose whether the tool is misaligned with the workflow or misaligned with the skill level of the person using it, then adjust. Or stop paying for it. Both are better outcomes than month four of untracked spending.