AI Adoption Is High. Why Is Enterprise ROI So Hard to Prove?

The gap is no longer anecdotal; it shows up in the benchmark data. In McKinsey's Global Survey on AI published in August 2026, 80 percent of respondents say AI has improved their individual productivity, yet only 37 percent report any positive EBIT contribution from AI, essentially unchanged from the prior year even as the share of organizations scaling AI enterprise-wide rose to 44 percent.[1] MIT's NANDA initiative reached a similar conclusion from a different angle: its 2025 report, The GenAI Divide: State of AI in Business, found that roughly 95 percent of enterprise generative AI pilots delivered little to no measurable impact on the P&L.[2]

Task speed is not business speed

The first reason is mechanical. An eight-hour report cut to two hours still waits days for review, approval and downstream use, so the end-to-end process barely moves. Individual task acceleration gets absorbed by handoffs, queues and decision points that AI never touched. As Howard Rubin argues in CIO, most celebrated AI metrics measure capability formation and activity, not realized enterprise outcomes, and the causal chain between the two is often indirect or delayed.[3]

The second reason is a missing decision. When hours are saved but nobody decides what happens to them, the savings dissipate into the same structures that produced the old baseline:

• The same headcount stays in place, because no redeployment decision was ever made
• The same backlog refills, because demand was never re-scoped
• The same external spend continues, because no contract or vendor volume was renegotiated
• The same cycle times persist, because the approval chain upstream and downstream was never redesigned

The third reason is the one this article addresses. None of this is a technology failure. It is a measurement-architecture failure: the company never built the explicit chain of evidence that connects AI usage to workflow change, to a named operational KPI, to a specific financial mechanism, and finally to EBIT. The rest of this article sets out how to build that chain.

We call the missing structure the AI Value Bridge. It is a five-rung measurement architecture that forces every AI use case to be evidenced in order, from first usage to financial statement. Each rung is a claim that must be proven before the next rung can even be asserted. The rungs are:

Contrast this with the typical AI dashboard, which jumps straight from rung two to rung six: it reports prompt volumes and active users on one side and a dollar figure on the other, with nothing in between. The middle rungs, where the value actually gets created or destroyed, are simply missing. That is why the numbers do not survive contact with a CFO review.

The bridge is the connective tissue leadership teams currently lack. McKinsey's 2026 survey shows enterprise-level financial impact staying flat even as scaling rose to 44 percent of respondents, which means the constraint is no longer deployment. It is the absence of a traceable path from deployment to earnings.[1]

The bridge only works if rung four is held to a hard standard, and most organizations fail it. Adoption metrics measure usage, and usage is not value. A thousand shallow users prompting occasionally can lose to fifty deep users embedded in a redesigned process. CIO's analysis makes the same distinction: counts of prompts executed, tokens consumed and employees using AI tools reveal adoption and activity, but say little about whether operational performance actually improved.[3]

A worked conversion

Take one common activity metric: the accounts receivable team reports 30 percent of hours saved on invoice-chasing through AI-drafted reminders. As stated, it is a slide number. Convert it: the hours are spent on faster, more consistent dunning, which shortens the average collection period. The operational KPI is days sales outstanding, with a frozen pre-deployment baseline of, say, the trailing twelve-month average. The delta between baseline and current DSO, multiplied by daily revenue, is what finance can actually book. If DSO does not move, the hours saved were activity, not value, and the use case gets re-scoped or killed.

The litmus test behind this example is simple: if a use case cannot be tied to a specific financial line item, it belongs in an R&D lab, not the operating budget. CFO.com's recent reporting makes the stakes clear, noting that despite widespread adoption, AI has rarely translated into measurable improvements in EBITDA, cash flow or return on invested capital.[4]

Rung five is where most value cases go quietly wrong, because hours saved is not a financial mechanism. Saved hours reach the P&L only through one of three specific routes, and each use case triggers exactly one of them first:

Choosing the mechanism is not an accounting afterthought; it determines what has to be decided operationally. Cost avoidance requires the discipline to leave a vacancy unfilled. Capacity release requires a named destination for the redeployed hours. Revenue per FTE requires a denominator both sides agree on. If the mechanism is not chosen in advance, the savings stay a slide number, because nobody was ever accountable for converting them.

Honest timelines matter as much as honest mechanisms. Deloitte's 2025 survey of 1,854 executives across Europe and the Middle East found that most respondents achieve satisfactory ROI on a typical AI use case within two to four years, against the seven to 12 months typically expected for technology investments, and only 6 percent reported payback in under a year.[5] A bridge built on a two-quarter payback promise will fail, not because the value is absent but because the expectation was set wrong.

Even a well-chosen KPI and mechanism prove nothing if the delta cannot be attributed. AI is rarely deployed into a static environment. Deloitte's interviewed executives are explicit about this: AI is typically introduced alongside efforts to improve data quality, reconfigure teams or streamline operations, which makes its value hard to separate. One energy-sector executive described only managing a ballpark estimate of the benefits because the gains from AI could not be separated from operational excellence programs and reorganizations happening at the same time.[5]

The practical answer is to freeze the baseline before deployment and pick an attribution design deliberately. Three designs cover most situations:

1. Staggered rollout with a control group: deploy the use case in some units or regions first and treat the not-yet-deployed units as the comparison, so secular trends cancel out
2. Before-and-after on a stable process: where the process itself is genuinely unchanged apart from AI, a frozen baseline with a defined observation window is defensible
3. Difference-in-differences across comparable units: compare the change in the KPI in AI-enabled units against the change in comparable non-enabled units over the same period

Whichever design you choose, the baseline must be frozen before deployment, not reconstructed afterwards from the memory of the team that owns the tool. A baseline built after the fact is a negotiation, not a measurement.

The counterfactual also needs a named financial owner. In most organizations, pilots are run by whoever commissioned the technology, with no one from finance holding the baseline or the attribution standard. When no one owns the counterfactual, success is undeterminable by construction: any reported gain can be challenged, and any shortfall can be explained away. That is precisely the position a CFO cannot defend in a board meeting, and it is an ownership problem, not an analytics problem.

No single function can build the bridge alone. The business owner knows the workflow: which steps actually disappeared, which handoffs were cut, where the redeployed hours went. Finance knows the mechanism and the counterfactual standard: which line item the value lands in, what evidence survives an audit, what baseline is defensible. A bridge built only by the business side produces hours-saved arithmetic; a bridge built only by finance produces skepticism. The value is created in the joint work.

The evidence supports the workflow side of that partnership. McKinsey's research finds that workflow redesign has the biggest effect on whether organizations see EBIT impact from generative AI, and its 2026 survey shows the high performers, the roughly 6 percent of organizations attributing at least 5 percent of EBIT to AI, are far more likely than others to have fundamentally redesigned workflows rather than inserting AI into existing ones, with nearly three-quarters having done so.[6]

Joint ownership, in practice, comes down to three disciplines:

1. One value-bridge sheet per use case, signed by both the business owner and the finance owner, with each rung evidenced in order before the next is claimed
2. A quarterly re-scoring cadence: rungs are re-verified, not assumed, because workflows and baselines drift
3. Kill-or-scale decisions taken only at rungs four and five, when the KPI delta and the mechanism are evidenced, never on adoption data alone

The cadence matters as much as the sheet. Quarterly re-scoring keeps the bridge honest as conditions change, and restricting kill-or-scale decisions to the top rungs prevents the most common governance failure: cancelling use cases on weak adoption numbers while scaling use cases that show usage but no KPI movement.

Armed with the bridge, look back at the typical AI ROI deck and the failure modes become easy to name:

• Unattributed benchmark numbers: industry survey figures presented as if they were your company's results
• Hours-saved arithmetic with no mechanism: 40,000 hours times a loaded rate, with no answer to which income statement line the money lands in
• No counterfactual: a delta claimed against a baseline that was never frozen, in an environment where five other change programs ran simultaneously
• No named financial line item: value described in the language of productivity, never in the language of the chart of accounts

The CFO test compresses all of this into one sentence: name the KPI, name the mechanism, name the counterfactual, or the number does not count. A deck built on the bridge answers each objection in order. To the question of whether the tool is used, it shows rung-two evidence tied to the specific workflow. To whether anything changed, it shows the before-and-after process map at rung three. To whether the KPI moved, it shows the frozen baseline and the measured delta at rung four. To where the money is, it names the mechanism and the line item at rung five. And to whether AI caused it, it shows the attribution design at the counterfactual. Every claim is one rung above the question that prompts it, and nothing rests on a benchmark borrowed from another company's survey.

This is also where the analysis work itself needs structure. Decisity is an AI-native strategy platform used by CFO and CSO offices to build exactly this kind of value-bridge analysis on the company's own documents: scoping briefs, process documentation, baselines and financials go in, and the value bridge per use case comes out as a board-ready deck with every claim linked to its source. When the CFO test is applied in the meeting, the evidence behind each rung is one click away, which is the difference between a number that survives scrutiny and one that does not. You can see how the platform structures this work at Decisity platform[7].

Related: the Decisity AI strategy consulting platform.