The $30–40 Billion ROI Gap: Why Enterprise AI Spend Isn't Producing Returns

Enterprise generative AI has absorbed an extraordinary amount of committed capital in a compressed period. Estimates from MIT's NANDA initiative and reporting by Fortune, Yahoo Finance, and others place cumulative enterprise commitments in the range of $30 to $40 billion. The same body of research places the share of pilots that produced no measurable impact on revenue, cost, or margin at roughly ninety-five percent.

That figure is not a rounding error. It is the working baseline of a bubble. The relevant question for finance, audit, and restructuring functions is no longer whether returns are materializing at the promised pace. The relevant question is what, exactly, has been bought — and whether the record can support any subsequent decision to retain, restructure, impair, terminate, or sell.

The gap is not a technology problem

It is tempting to treat the ROI gap as a maturity problem: models will improve, adoption will deepen, gains will follow. That framing has been rehearsed continuously since late 2022. It also happens to be the framing that discourages any interim accounting of what has already been spent, on what, to what documented effect.

The pattern the reporting describes is different. It is not that AI systems fail to work in a technical sense. Many of them work. The pattern is that enterprises cannot reconstruct the causal chain between the spend and any operational outcome, because the underlying record was never designed to support that reconstruction.

What the record usually looks like

In the engagements that begin as an ROI question and end as a workout question, the findings are consistent. The evidence is fragmented across procurement, cloud billing, identity systems, model providers, business-unit budgets, and personal productivity tools. Ownership is contested. Contract terms have drifted. The system in production is not the system that was approved.

• Cloud and compute charges booked to shared infrastructure lines with no allocation back to the AI initiative that generated them.
• License and seat sprawl accumulated through departmental purchases, expense reports, and pilot conversions that were never formally reconciled.
• Model version changes made unilaterally by vendors, with no internal record of the switch or its output implications.
• Human review, correction, and rework labor absorbed by operations teams and never captured as a cost of the AI system it supports.
• Business cases that were approved on ROI assumptions no one has revisited since deployment.

Why the gap is a reconciliation problem

If the record cannot answer basic questions — what was acquired, from whom, at what cost, producing what documented outcome, owned by which decision-maker — then no responsible workout action can be taken. An impairment cannot be defended. A vendor renegotiation cannot be sized. A termination cannot be executed cleanly. A sale cannot be structured. The capital sits, unreconcilable, on the balance sheet.

This is the condition post-bubble portfolios inherit. Not necessarily fraud, not necessarily failure in the operational sense, but an absence of the record required to convert the investment into any defensible decision. The ROI gap is the visible symptom. The reconciliation gap is the operative problem.

The precedent

The closest institutional precedent is the workout of thrift and bank portfolios in the early 1990s. The Resolution Trust Corporation did not invent value. It reconstructed records that had been permitted to decay during the boom, then made the retain-restructure-impair-sell decisions the reconstructed record supported. The enterprise AI cleanup follows the same sequence in a different medium. First the record, then the decision. The pattern the publicly reported failures share is not model failure — it is missing record. And the record is usually incomplete because shadow AI spend was never captured against the initiative it belonged to.