AI ROI needs an acceptance-to-EBIT bridge
In short: Individual productivity is not enterprise ROI; connect accepted AI work to adoption, economic value, allocation and a finance-owned EBIT reconciliation.
McKinsey's 2026 State of AI survey puts three numbers beside one another: 80% of respondents said AI improved their individual productivity, 37% attributed some EBIT impact to AI, and only 6% met its high-performer definition of at least 5% EBIT impact plus “significant” value.
Those numbers do not prove that productivity gains failed. They show that individual productivity and enterprise value have different denominators. AI ROI needs an acceptance-to-EBIT bridge before time saved enters a financial claim.
McKinsey measured three different layers of value
The survey published on 25 August 2026 covered 1,719 participants in 97 nations between 4 May and 8 June. Its productivity, decision-quality and EBIT findings are respondent reports, weighted by each nation’s contribution to global GDP. They are not audited accounts or a causal estimate of profit created by AI.
That distinction is useful. An employee can finish a task faster while the business saves nothing because the freed capacity is not used, the output needs rework, adoption stays low, or the benefit disappears inside a shared cost centre. A valid gain at the work layer can vanish before it reaches the company ledger.
A workflow can also create value through faster response, lower risk or revenue growth without appearing as fewer labour hours. One universal “hours saved × salary” formula misses both cases.
The acceptance-to-EBIT bridge has five fields
This is editorial judgement, not a Trigent or client deployment I ran. Before an AI workflow enters a portfolio ROI number, I would require five linked fields:
| Field | Evidence required |
|---|---|
| Accepted work unit | The completed case, claim, change or decision that passed its owner’s acceptance rule |
| Adoption denominator | Eligible units, attempted units and accepted units for the same period |
| Economic effect | Displaced cost, avoided loss or added revenue against a declared counterfactual |
| Allocation rule | How shared model, platform, review and change costs reach this workflow |
| Finance reconciliation | The account, period and owner who can explain how the effect appears in EBIT |
Accepted units prevent activity from posing as value, while the adoption denominator catches a successful pilot that nobody uses. The economic effect forces the counterfactual into view. Allocation stops a cheap model call from hiding an expensive operating system. Finance gets the final right to reject double counting.
Workflow redesign is part of the financial control
McKinsey's 14 September agentic-AI blueprint describes one medical-technology project losing four to five weeks to data rework. Another platform mismatch left an estimated 30% of intended value outside scope. These are McKinsey-reported client examples, not independently reproduced results, but they expose where the bridge breaks: before model performance can become business performance.
So I would review the bridge monthly with the workflow owner and finance, not leave it inside the model team’s dashboard. If accepted volume rises but the economic effect does not, investigate adoption, rework, capacity use and allocation before celebrating more tokens or more users.
What's in it for you
- Keep individual productivity as evidence without mislabelling it as enterprise profit.
- Find exactly where accepted work stops turning into financial value.
- Give finance a traceable basis to approve, adjust or reject an AI ROI claim.
Count accepted work first; call it ROI only after finance can trace its path into the ledger.
