Ninety-seven percent of middle market leaders say they are satisfied with their AI investments. Fifty-six percent of CEOs worldwide say AI has delivered neither higher revenue nor lower costs. Both numbers are real. Satisfaction measures how a tool feels to use. Return measures what changed in the business, and that second measurement has to be set up before the tool goes in.
In March 2026, RSM surveyed 1,030 executives at companies between $30M and $10B in revenue. Ninety-seven percent reported satisfaction with their AI investments, and 54% said those investments had exceeded their ROI expectations (RSM Middle Market AI Survey 2026). A few months earlier, PwC put a harder version of the question to 4,454 chief executives in 95 countries. Fifty-six percent said AI had produced neither a meaningful revenue increase nor a meaningful cost decrease. Twelve percent said it had produced both (PwC 29th Global CEO Survey).
Neither survey is wrong. They are measuring two different things, and the distance between them is where the money goes.
Ask whether a tool is good and people answer from experience. Drafts come back faster. The proposal took an hour instead of four. The quoting desk stopped dreading Mondays. All of that is real and worth having.
Return asks something else. Which line moved, by how much, against what it was doing before. That last clause is the hard one, because it requires a before, and a before only exists if somebody wrote it down while it was still true. Once the tool is in and the workflow has reshaped itself around it, the old baseline is gone. You cannot go back and take the measurement later.
In the same RSM survey where satisfaction ran at 97%, a third of respondents named unclear ROI as a barrier to scaling. Those are not contradictory answers from confused people. They are the correct answers to two different questions.
Freed hours get quietly reabsorbed. A tool saves the estimating team six hours a week. Nothing on the P&L moves, because six hours went back into the day. Hours only become money when a decision converts them, and the decision is a management one, not a technical one. More quotes out the door. Fewer temps through the peak. One open role not backfilled. If no one makes that call, the saving is real and permanently invisible.
The pilot gets picked for how well it demos. The impressive use case and the profitable use case are rarely the same. Return concentrates in the unglamorous, high-volume work: quoting, order intake, claims coding, dispatch notes, invoice matching, first-pass document review, warranty claims. A three percent improvement on four thousand transactions a month beats a forty percent improvement on thirty.
Nobody owns the number. The vendor owns the deployment. IT owns the integration. Finance owns the invoice. In many companies no single person has been asked to come back in ninety days and say what it returned. Without that name and that date, the question simply never gets asked again.
That is roughly twenty minutes of work, and it is the difference between a satisfaction answer and a return answer a year from now.
Two thirds of the RSM respondents said they apply governance controls before the pilot or production stage, which means a third are running without them. RSM's separate cybersecurity report found only 35% of middle market executives use a formal AI governance framework (RSM Cybersecurity Special Report). The link to the money is direct, not philosophical. A tool that has to be pulled after a data exposure, a confident wrong answer that reaches a customer, or a contract clause nobody read before the data started flowing does not produce a poor return. It produces a written-off one, plus the cost of the cleanup and the year it takes before anyone is willing to try again.
Ground rules written early are cheap. Written after an incident, they cost the incident too. What actually belongs in them is shorter than most people expect.
Nobody is going to run a controlled trial on their own quoting desk, and they should not try. What is achievable is this: pick the work where the volume is, write down what it costs today, name the person who converts the saving, and set the date the question gets asked again. Do that and you will be able to answer the return question in twelve months. Skip it and you will have a satisfied team, a real improvement, and no way to prove either one to a board, a lender or a buyer.
The full ranked picture, both what AI is already exposing and where it is worth the effort in a specific business, is what the Exposure & Opportunity Review produces. If the question is narrower than that, where AI actually pays in a company this size covers it directly.