Enthusiasm, Not Evidence: EY's AI Chief Says Businesses Still Can't Show Returns

For roughly three years, the enterprise AI conversation has run on a simple promise: spend now, measure later. This week, one of the people responsible for the biggest AI deployments on Earth said the second half of that bargain still isn’t being kept.

Dan Diasio, Ernst & Young’s global consulting AI leader, told The Information’s AI Agenda Live conference that businesses are still not seeing substantial revenue gains or cost reductions from their AI programs. His framing was blunt: corporate spending on AI right now is “based on enthusiasm as opposed to that evidence.”

Coming from most commentators, that would be a hot take. Coming from EY’s top consulting AI executive — the person whose firm is paid to make enterprise AI actually work — it reads more like a field report. EY sits inside hundreds of large-scale transformations. When Diasio says the returns aren’t showing up in the numbers yet, he is describing what his own practitioners see at client after client, not a thesis drafted from afar.

What was actually said

Diasio’s core claim, as reported by The Information’s Laura Bratton and Kevin McLaughlin, is that the gap between AI spending and AI value remains stubbornly wide. Companies keep expanding budgets, but when EY looks for the two things that matter — meaningful revenue lift and meaningful cost reduction — the evidence is thin on the ground for the majority of adopters.

He also noted that EY runs extensive CEO surveys, and that leadership conviction about AI remains high even where the financial proof is weak. That combination — strong belief, weak measurement — is precisely what he means by enthusiasm outpacing evidence. It is not an argument that AI is useless; it is an argument that the average enterprise is still paying for capability it has not yet converted into outcomes.

The statement matters for timing. It lands after a stretch in which the biggest AI vendors have shifted their pitch from “experiment” to “deploy agents everywhere,” and after a year in which consulting firms — EY included — have quietly built new internal machinery to police their own AI spending.

EY is policing itself first

The most telling detail is that EY is behaving as if Diasio’s warning applies to itself.

In August, the firm announced an “AI Value Realization Office” — a dedicated function whose job is to centralize oversight of EY’s own AI investments, monitor adoption, and decide which internal initiatives deserve to scale and which should be killed. Bloomberg reported that the unit is designed to give leadership “end-to-end visibility” of where and how AI money flows inside the firm. Around the same time, EY began hiring for a head of “agent economics” — a role that exists specifically to scrutinize the cost of an increasingly AI-powered workforce.

Read together, these moves are an unusual admission for a consultancy: if AI value were easy to capture, you would not need a new office and a new C-suite-adjacent role to chase it. EY is institutionalizing the measurement problem inside its own walls before selling the fix to clients.

The pattern is not new — but it is persisting

Diasio’s comments extend a line of evidence that has been accumulating for over a year. MIT’s widely cited “State of AI in Business” analysis found that around 95% of enterprise generative AI pilots produced no measurable financial return, against an estimated $30–40 billion in enterprise GenAI spending. Later industry analyses put the share of AI pilots that never reach production at roughly 88%. EY’s own US AI Pulse Survey work has oscillated between optimism and caution — touting high levels of reported positive ROI in some waves while flagging that reinvestment, not headcount reduction, is where most gains currently go.

What changed in 2026 is the scale of the bet. Global AI spending estimates now run into the trillions of dollars when infrastructure is included, and hyperscalers are pouring capital into data centers on the assumption that enterprise demand will keep compounding. If the average corporate buyer still cannot demonstrate returns, that assumption rests on faith in future capability rather than current P&L impact — which is exactly the “enthusiasm versus evidence” gap Diasio named.

Why returns stay invisible

A few structural reasons keep showing up in the failure post-mortems:

• Pilots that never touch the core. Demos run at the edge of the organization, on friendly workflows, with hand-picked data. Production means integration with messy systems of record, and that is where most projects die.
• No redesigned workflow. Layering a model onto an unchanged process captures a fraction of the possible value. The organizations that do see returns tend to be the ones that restructure the work around the model, not the ones that buy licenses and wait.
• Measurement is an afterthought. Many adopters never defined a baseline before deploying, so even genuine improvements cannot be proven to a CFO. The “value realization” concept EY is formalizing internally exists to fix exactly this.
• Cost accounting hides the bill. Token spend, seat licenses, integration engineering, and remediation of agent errors land in different budget lines, making total cost — and therefore return — unusually hard to compute.

Why it still matters

The uncomfortable reading of Diasio’s remarks is that 2026’s agentic-AI wave raises the stakes of the same old problem. Agents don’t just generate text; they take actions in production systems. If enterprises struggled to get returns from a chatbot summarizing documents, the burden of proof for autonomous software that touches customers, code, and payments is far higher — and so is the cost of getting it wrong, as this year’s string of agent-related security incidents demonstrated.

The charitable reading is that this is what the middle of a general-purpose technology rollout looks like. Electrification took decades to show up in factory productivity statistics; the famous “productivity paradox” of computers in the 1980s resolved once businesses reorganized around them. Diasio himself does not dispute the capability curve — his complaint is about conversion, the discipline of turning capability into measured value.

Both readings can be true. The capability is real and compounding; the median enterprise’s ability to harvest it is lagging badly. Statements like Diasio’s are useful precisely because they mark the gap instead of papering over it — and because the firms best positioned to know are quietly reorganizing themselves as if the gap is their next big market.

For buyers, the practical takeaway is unglamorous but valuable: before the next budget cycle, insist on a baseline, pick workflows you are willing to actually redesign, and treat any AI initiative that cannot name its counterfactual as enthusiasm, not evidence.