Why three attractive AI statistics cannot answer a company’s investment question.
By Alexey Kosov
Evidence checked: 16 October 2026. Public-source analytical note; no verified client outcome.
Decision in brief
Widespread AI adoption does not establish business value for a particular company. The decision is not whether AI is popular, but which workflow deserves a controlled test and what evidence would justify further investment. Before scaling, choose an owned process, measure its current time, quality, rework and full costs, and run a bounded pilot with explicit stop conditions. Stanford HAI, Microsoft and McKinsey report different things: organizational use, AI-user experience and leaders’ assessments of transformation. Their percentages cannot be combined into an estimate of financial return. This is a method for evaluating a decision, not evidence that a KTC client has achieved returns.
Three figures, three different questions
Stanford HAI’s 2026 AI Index reports an 88% figure from McKinsey’s 2025 State of AI survey: respondents reporting that their organization used AI in at least one business function. The underlying survey covered 1,993 people across 105 countries. This tells a reader about reported organizational use within the survey base. The same chapter notes that AI-agent deployment is still in single digits across almost all business functions. Neither observation shows that 88% of companies gained a measurable financial return. [1, 1a]
Microsoft’s 2026 Work Trend Index reports that 66% of surveyed AI users say AI has let them spend more time on high-value work. Its online survey covered 20,000 full-time employed or self-employed knowledge workers who already used generative AI for work in ten countries (February 18–April 7, 2026). The report also uses product telemetry, but this particular 66% is a self-report, not objective proof of productivity, revenue or time saved for the wider workforce. [2]
McKinsey’s 2026 readiness study surveyed 750 English-speaking workers who had already incorporated AI into their jobs. It found that 11% of surveyed leaders described their organizations as reaching its ‘reinvention’ horizon. The authors explain that recruitment deliberately included more advanced organizations, and that results may not represent general market prevalence. Organizational readiness and enterprise-value questions were asked of a smaller leader subset (608), not all 750 participants. Eleven percent is neither a universal adoption measure nor an experiment demonstrating what caused economic value. [3]
These three percentages cannot be placed on a common ‘AI maturity’ chart without changing what they mean. Their units differ: survey respondents describing their organizations, surveyed AI users describing their experience, and leaders assessing organizational redesign. The questions differ too: use, personal experience and transformation. This does not mean AI creates no value. Stanford summarizes measured gains in certain tasks, while McKinsey’s 2025 survey reports some respondents attributing earnings improvements to AI. Neither result predicts the return from a particular company’s proposed workflow. [1, 1a]
What an investment decision needs
Before a founder buys an AI system, or a department asks for a larger license budget, a useful evidence chain has three stages.
First, find a workflow worth testing. Define the present work, who owns it and what happens when it fails. For example, a company might consider AI-assisted classification and first-response drafting for incoming service requests. That is only an illustrative candidate, not a recommendation for an unseen business.
Second, establish an observed baseline. Measure current time to an accepted response, correction and rework rates, human escalations and the fully loaded cost per accepted case before introducing AI. Separately estimate the proposed AI step’s incremental license, integration, training, monitoring, security, human-review and exception-handling costs. Check whether the data is reliable, whether permissions permit the proposed use and whether a human must approve outputs before they affect customers.
Third, test a narrow change. A short pilot might compare comparable incoming requests before and after an assisted step, controlling where feasible for case mix, workload and seasonal changes, and recording exceptions and reasons for correction. A before-and-after comparison alone does not prove that AI caused any observed improvement. A better average response time has little value if errors become harder to detect, if workers simply review more AI output, or if the cost of oversight exceeds the benefit.
A ‘GO’ decision would need a measurable improvement against the agreed baseline, credible net value after incremental costs, acceptable quality and privacy safeguards, and clear human ownership. ‘LEARN’ may be appropriate if the sample is too small or inputs were inconsistent. ‘HOLD’ is the responsible choice if no baseline exists, risks cannot be controlled, or no process owner is willing to accept the result.
These gates do not require every company to conduct a formal randomized trial. They require each company to avoid calling a vendor demonstration an outcome.
Limits of this analysis
These surveys describe reported use, personal experience or assessed organizational change; none tests the proposed workflow of an individual buyer. A short pilot can be affected by case selection, workload or temporary effects even when its metrics improve. This note does not choose a vendor or establish client-specific financial returns.
What a leader should ask next
Ask three questions before funding a larger rollout:
- Which decision would this pilot actually change?
- Which single observed workflow measure could falsify the promised value?
- Who would stop or redesign the deployment if its costs, exceptions or customer harm increased?
The answers matter more than choosing the most impressive percentage for a slide.
The same evidence discipline applies to market-entry and product decisions: a source must support the precise claim used in a recommendation. The KTC method note, ‘Keep evidence attached to decisions‘, explains how to maintain that connection.
AI investment may be justified. The next sensible expenditure, however, is evidence about the workflow, net benefit and risks that would make your own deployment decision defensible—not another impressive adoption percentage.
Method note: AI-assisted research and drafting supported source coverage and structure. The interpretation, claim boundaries and publication decision remain the author’s responsibility.
