Reading Deloitte's "State of AI in the Enterprise 2026"

On January 21, Deloitte published State of AI in the Enterprise: The Untapped Edge, its annual enterprise AI survey, drawn from 3,235 director- and C-suite-level leaders across 24 countries and six industries. Deloitte's key finding is that AI is delivering productivity to most organizations but business transformation to very few, and the gap between those outcomes — what Deloitte calls "the untapped edge" — defines the strategic moment.

The report hits the familiar beats of industry research in this space: a near-universal desire for AI to grow revenue, an absence of meaningful proof that revenue is arriving, and self-reported productivity gains that don't quite map to success. 74% of surveyed organizations want their AI initiatives to grow revenue, but only 20% report that they have.

66% report productivity and efficiency gains overall, and the report separately notes that all three transformation tiers — deep transformation, process redesign, and surface-level use — are capturing those gains. Having produced reports of this kind, it stood out to me that the report has the data to test whether those gains differ meaningfully across tiers. I am virtually certain that test was run. And I know how a finding like "deep transformers see X% higher productivity gains" would be presented if it existed — it would headline the recommendation, not surface as a flat qualitative observation that "each tier is capturing gains." That phrasing is what reports produce when the test was run and the differential wasn't there to feature. If that's true, the recommendation to transform deeply has no support in the data. If no difference was found across tiers of transformational depth, the rational read of Deloitte's findings is to do nothing. Whatever you are already doing is evidently getting you the same gain you would get from rebuilding the company around AI.

This kind of selective framing threads through the entire report. We are told AI is making organizations more productive but isn't generating revenue. We are told that sovereign AI is a national-security imperative, but building critical enterprise infrastructure on top of foundation models owned by third-party providers is sound long-term strategy. These contradictions are quietly waved past, and they cohere only if you accept the tacit premise that the AI status quo is, on net, good. That premise is the investor class's article of faith, and the industry report genre's task has largely become to find ways to keep believing it, even when the data won't quite cooperate.

But the tension between Wall Street's belief in AI's transformative potential and the absence of data to justify that belief isn't strictly delusion. I see it as another instance in which AI's transformative nature requires a ground-up rethinking of basic premises — including the premises behind the research meant to evaluate it. In this case, the industry-standard, survey-style "State of X" report is partly culpable for the lack of meaningful guidance, because surveys are the wrong tool for the work this moment requires.

Survey science is well-suited to aggregating stable preferences over known territory: whether 85% of customers liked yesterday's menu item tells you something about how they will feel about it tomorrow. It is not well-suited to charting direction in a moment of structural transformation, where the territory is genuinely new and the respondents are themselves inside the disruption they are being asked to describe. The data are historical — by the time they are collected, coded, and published, the conditions they described have already shifted. And survey science depends on respondents understanding what they are being asked about, which, in a genuinely new domain, is the very thing under contestation. When the report finds that 21% of organizations have a "mature model for governance of autonomous agents," the 21% rests on 3,235 private definitions of mature governance and autonomous agents, in a field that has not yet settled the meaning of either. Asking 3,235 leaders what they think about a fundamentally transformative technology isn't an escape from the ambiguity. They are inside the ambiguity with the rest of us. If the data lack directional clarity, it is because the respondents do.

What the moment needs is the opposite of what has long been advocated in tech-sector decision-making: expertise-driven direction in the absence of data-driven direction. Data are confirmatory by nature. They can describe what is and test claims about what is, but they cannot generate what should be. Direction-setting through a structural transformation of labor is generative work, and generative work belongs to expertise and theory, not to aggregation. What we have instead are people who have spent careers thinking carefully about how work, organizations, and labor actually function — researchers in the humanities, critical scholarship, organizational studies, and philosophy of technology, to name a few.

Directional questions of this magnitude cannot be answered by aggregating opinions of generic respondents. Industry reports keep failing to find evidence that AI is generating value, and the risk is that to a casual reader the failure reads as evidence of absence. The debate frozen between "AI is good" and "the data shows AI is bad" has missed what the moment actually is. AI is reality; the work in front of us is figuring out how to make it good, and knowing who to ask about that path.