Both sides of the AI ROI debate run on soft numbers
The viral 94% AI cost-saving claim has no survey behind it. KPMG's skeptical 7% is also self-reported maturity, not calculated ROI. Build one workflow receipt instead.
The booster number is 94%. The skeptic number is 7%. Neither is calculated AI ROI.
The first claim says 94% of small businesses using AI agents cut operating costs by at least 30% in one quarter. I can trace the exact sentence to a March 28 tool-review page. I cannot trace it to a survey. The page says its authors tested 21 tools across three businesses. That work, even if accepted as described, cannot establish what 94% of small businesses experienced.
The second number comes from a real survey with a named instrument. KPMG asked 2,145 senior leaders across 20 countries and territories about AI between April 28 and May 25. Seventy-six percent agreed that AI was delivering meaningful business value. Seven percent selected “Established ROI” as the phase that best described their organisation's AI journey.
That is useful evidence. It is still a self-placement on a six-step maturity curve, not a return calculated from cash flows. The AI ROI argument has put an untraceable success rate on one slide and a self-reported maturity label on the next, then mistaken the distance between them for financial truth.
The bullish number has no denominator
There are attributable positive numbers. CrewAI's February survey covered 500 C-level executives and senior leaders at organisations with more than $100 million in revenue and 5,000 employees. Sixty-nine percent reported significant reductions in operating costs. The vendor also reports that every respondent planned to expand agent use.
That tells us what a selected group of large-enterprise leaders said. It does not supply the baseline cost, the size of the reduction, the deployment denominator, the cost of the agents, or a financial return. CrewAI sells agent infrastructure, which does not invalidate its survey; it does make the instrument and missing fields part of the evidence.
The 94% claim is weaker. The closest credible 94% I found is Nvidia's 2025 retail and consumer-goods survey: 94% of respondents said AI helped reduce annual operating costs. Nvidia did not attach a 30% threshold in its published summary, did not limit the result to agents, and studied one sector.
So “94%” exists. “30%” exists elsewhere as a threshold and forecast. The combined small-business-agent sentence still has no named population, instrument, sample size or fielding date. It should not enter a budget deck.
The skeptical number is one rung firmer
KPMG is much more inspectable. Its report names the sample, eligibility, geography, fielding window and questions. It also states what its categories mean. “Meaningful business value” includes productivity, cost, revenue and decision-making outcomes. “Established ROI” is the sixth phase of an AI-maturity journey: meaningful outcomes with tangible growth and opportunity.
The actual chart asks respondents which phase best describes their organisation. KPMG says the measure is intended to identify organisations that can prove outcomes justify investment. But the survey does not ask respondents to provide a baseline, cost ledger, benefit amount, time period, discount rate or calculation.
That distinction matters because “only 7% can prove ROI” sounds audited. The evidence supports a narrower sentence: 7% of surveyed leaders placed their organisation in KPMG's established-ROI maturity phase.
KPMG's own economics section reinforces the measurement gap. Only 35% reported that AI operating costs were fully visible and actively monitored; 42% had partial visibility. Firms with full visibility were five times as likely to report established ROI, 15% versus 3%. That is an association inside a self-reported survey, not proof that installing a dashboard causes returns.
rendering diagram…
The skeptic has climbed one rung higher than the booster. Neither has reached the calculation.
Replace the survey fight with one receipt
I have seen this movie in a client budget review. A champion arrived with a momentum percentage; a skeptic arrived with a failure percentage. Drawing the denominator changed the conversation, but it did not create the missing baseline. I wrote about that meeting in July. The harder question was whether the client could prove its own row.
For the next steering meeting, pick one workflow and bring a one-page receipt:
- Unit: name the accepted outcome—resolved case, collected invoice, qualified lead or merged change.
- Baseline: record volume, full labour and software cost, cycle time and defect or reversal rate before the agent.
- After: measure the same fields for a fixed period, including model, tool, integration, review, rework and exception costs.
- Net: subtract the full after-cost and one-time implementation cost from the monetised benefit. Do not annualise a one-time cleanup as recurring value.
- Guardrail: keep quality and risk beside the return. A cheaper workflow with more reversals is a different product.
This also catches feature activation disguised as agent ROI: switching on a reminder inside software you already own can be a real one-time win without proving a recurring intelligence premium.
The point is not that AI has no return. It is that survey sentiment cannot settle the amount in either direction. The first honest ROI number is usually smaller, local and less impressive than the headline. It is also the only one a finance team can act on.
Stop asking which survey won the AI ROI debate; bring one baseline, one accepted unit and one fully loaded before-and-after calculation.
