AI Tools Look Cheap Until You Count the Work They Create

McKinsey surveyed companies across industries and found that 80 percent had adopted the latest generation of AI tools, yet the same 80 percent reported no significant gains in either revenue or cost reduction. That is not a rounding error. That is a pattern. The OECD's 2025 survey of small and mid-sized businesses across ten countries found that only 6 percent of SME AI users reported transformational impact, despite 39 percent of those businesses now using some form of AI, up from 26 percent the year before. Adoption is climbing. Financial returns are not following.

The obvious explanation is that the tools are overhyped. The more accurate explanation is that owners are measuring the wrong things.

What the subscription fee does not include

When a small business owner signs up for an AI writing tool, a customer service bot, or a generative model to handle internal queries, the visible cost is the monthly fee. What does not appear on the invoice is the time spent checking outputs, rewriting what the model got wrong, rebuilding prompts when the tool drifts, and switching between platforms when one tool fails to do what another does better. None of that labour shows up in a time-saved calculation. All of it is real work.

The research names these costs explicitly: error correction, model oversight, data preparation, prompt refinement, tool switching, and psychological strain from working alongside systems that produce confident-sounding wrong answers. Each category is a line item that most small business accounting systems do not track, because the work looks like normal employee time rather than AI-related overhead.

This is where the ROI calculation breaks. You subtract the subscription fee from the hours saved and call it a win. The hours spent on correction and oversight never enter the equation.

The experiment that shows both sides at once

A field experiment with 758 consultants at Boston Consulting Group tested GPT-4 on a set of realistic consulting tasks. For 18 tasks within the model's capability range, consultants using the tool finished 12.2 percent more tasks, worked 25.1 percent faster, and produced output rated more than 40 percent higher in quality than a control group working without AI. Those are large, real gains.

The same experiment included one task outside that capability range. On that task, consultants using AI were 19 percentage points less likely to produce correct solutions than those who worked without it. The consultants using AI did not know they were underperforming. They produced worse answers while the tool presented them with apparent confidence.

This is the asymmetry that makes standard ROI tracking dangerous. You measure the 25.1 percent speed gain. You do not measure the downstream correction work generated by confident wrong outputs. The gains appear in your productivity log. The losses appear in someone's afternoon, unrecorded.

When faster workers and flat profits coexist for years

The reasonable counterargument is that these hidden costs are temporary. Workers learn the tools, figure out where AI helps and where it does not, and the overhead fades as the learning curve flattens. Task-level gains eventually aggregate into firm-level returns. The numbers will catch up.

McKinsey's 2026 adoption survey tests this directly. By that point, 54 percent of large organisations had scaled AI across the enterprise, meaning the learning curve had had real time to resolve. The share of organisations attributing any EBIT impact to AI stayed flat year over year despite that higher adoption rate. The lag explanation requires the enterprise-level numbers to move as adoption matures. They did not move.

The BCG experiment adds a specific reason why the learning curve argument struggles here. The consultants who underperformed on the out-of-frontier task were not novices who would improve with practice. They were experienced professionals who received no signal from the tool that they were going wrong. A learning curve works when workers get feedback. When the tool produces confident wrong outputs and the worker cannot distinguish them from correct ones, the correction burden falls on whoever reviews the work downstream, not on the person who will eventually learn to use the tool better.

What accurate tracking requires

The 6 percent of SMEs reporting transformational impact from AI in the OECD survey are not using better tools than the other 94 percent. The difference, based on what the research attributes to real ROI, is measurement and workflow design. Firms that reach genuine financial returns from AI track indirect costs alongside direct savings and redesign processes around where AI performs well rather than applying it broadly and hoping the aggregate numbers resolve.

For a small business owner, this does not require an analytics team. It requires adding two columns to whatever time-tracking or project management system you already use: one for correction work on AI outputs, one for oversight time spent reviewing before outputs go anywhere. Run that for 60 days across the tools you pay for. The number in those columns is the figure your current ROI calculation is missing.

McKinsey's unit-level data shows that teams using AI do report cost decreases and productivity improvements within their own operations. Those gains are real. The problem is not that AI delivers nothing. The problem is that the gains concentrate inside the AI's capability boundary, and the losses concentrate outside it, and most tracking systems measure only one side of that line.

If your AI subscription costs $200 a month and your team saves eight hours, the math looks obvious. Add the hours spent fixing outputs, re-prompting after failures, and switching tools when the model hits its limits, and the math changes. Whether it still favours the tool depends on your specific numbers. The point is not that AI is a bad investment. The point is that you do not know yet, because you are only counting half the ledger.