Proving AI value is an infrastructure problem

Gartner calls 2026 the inflection year for enterprise AI spending, forecasts worldwide AI spend up 47%, and in the same release notes that CIOs “face challenges in proving the value from AI investments,” which is why they favor tactical initiatives over disruptive ones. Deloitte research puts a number on the discomfort: only 28% of global finance leaders report clear, measurable value from AI, and in Deloitte’s 2025 US Tech Value survey, nearly half of leaders expect up to three years before basic AI automation shows ROI. Money going up 47%, provable value visible to about a quarter of the people who sign for it: that tension is the defining budget conversation of this year, and it lands on engineering leaders as a question that sounds unfair. What’s the AI actually returning?

The question is answerable. But I’ve come to a specific view about why so few companies can answer it, the same view I hold about AI adoption generally: it’s an infrastructure problem. Not a model-quality problem, not a use-case problem, and not something a better slide fixes. ROI is a fraction, value over cost, and in most organizations the platform produces neither the numerator nor the denominator.

The denominator: what does the AI cost?

Start with the half engineering fully controls. AI cost is the fastest-moving line in the technology budget, and it’s denominated in tokens, which means it scales with usage nobody provisioned. In most shops it accumulates on shared line items: one inference endpoint, one GPU pool, one vendor bill. Tagged perfectly, those still attribute all the spend to “the AI platform,” which is true and useless, the same dead end as an unattributed EKS cluster.

The fix is per-request metering: which feature made the call, how many tokens, what model tier, whose customer. With it, “what does the AI cost” becomes a query, per feature and per month, and questions like “did the support copilot cost more than it saved this quarter” become arithmetic instead of a meeting. Without it, every ROI claim in the deck is an allocation guess wearing a decimal point.

The numerator: what did the AI change?

The value side is harder, because it usually requires admitting the “before” was never measured. A copilot that reduces ticket-handling time proves nothing if handling time wasn’t tracked before the rollout; an agent that accelerates infrastructure changes proves nothing without the prior cycle time on record. The workflows AI touches are exactly the ones that were rarely instrumented, because they were human workflows.

This is still platform work. Baselines, event logs, before-and-after measurement on the one or two workflows a given AI feature claims to improve. Not a company-wide value framework, one measured workflow per funded initiative, defined before the initiative ships. The discipline resembles what an AI-ready platform needs anyway; readiness and provability are mostly the same buildout, and the teams that skipped the layers are the same teams that stall in pilots and then can’t say what the pilot was worth.

Tactical with a meter beats transformational with a story

Gartner’s observation that CIOs favor tactical initiatives gets read as timidity. I read it as correct instinct under-executed. A tactical initiative with honest metering, this feature, this workflow, this cost per request, this change in cycle time, builds the credibility that funds the next, larger one. The three-year ROI expectation in Deloitte’s survey data says the funding conversation will recur many times before value is fully realized. What survives that many budget cycles is instrumentation. A narrative gets one meeting, and it spends trust; a meter spends evidence, and it compounds.

So when the CFO asks what the AI is returning, the strong position is a report generated from the platform: AI cost per feature per month from the metering layer, workflow deltas against recorded baselines, and an honest line dividing what’s measured from what’s still estimated. The companies that can produce that page are the 28%. Producing it is not finance work, and it is not storytelling. It’s a few weeks of platform engineering, pointed at the newest spend in the building, and it’s considerably cheaper than the AI itself.