Share with your CFO
Most enterprise AI ROI claims circulating right now are measuring the wrong thing, and the SAP/Oxford Economics survey finding of 21% average expected ROI from 2,600 business leaders is a useful illustration of the problem: expected is doing enormous work in that sentence. A detailed breakdown of enterprise AI ROI disclosures from 2025 and 2026 pulls apart eight named cases, from Lupin’s 42.8 crore rupee savings figure to JPMorganChase doubling transaction-screening volume, and finds that nearly every headline number collapses under scrutiny when you ask for matching investment costs and a clean baseline.
What this means for your business
The companies most exposed here are those whose AI business cases are built on avoided-cost logic without a disclosed denominator. Intercom’s $7.5 to $9 million in annual savings sounds compelling until you notice it rests on a counterfactual, specifically that the company would have hired 100 additional support staff, a number no auditor can verify. Finance organizations that approved budgets on similar avoided-cost reasoning should be asking right now whether the denominator was ever stress-tested, because the framing decides the ROI more than the AI does.
The cases that hold up share one structural trait: they report an operational outcome first, then trace a financial consequence from it. Bank of America’s nearly one minute saved per call across 18,000 representatives is auditable, you can multiply call volume by labor cost and compare it against deployment spend. DBS’s S$1 billion economic value figure, by contrast, covers 430 use cases and two decades of analytics investment bundled together. The discipline gap between those two disclosures is not a measurement problem, it is a governance problem, and the CFO is the one who owns it.
The leading indicator to watch is whether AI savings targets start getting disaggregated in earnings calls the way segment revenue is. AT&T’s $4 billion cost target by end of 2028 bundles AI with vendor rationalisation, network retirement, and workforce changes into a single program number. If that distinction never gets made, the AI contribution will be impossible to isolate, which means future investment decisions will be made on the same anecdote-dressed-as-ROI logic that got the budget approved in the first place. The CFO who demands a clean cost baseline before the next approval cycle forces that discipline sooner than the audit committee will.
Concept deep-dive: Avoided-cost calculation
An avoided-cost calculation measures savings against a hypothetical future spend rather than against actual historical costs, the equivalent of claiming you saved money on a car you never bought. It is legitimate when the counterfactual is well-defined and defensible, a staffing model with documented headcount drivers, for instance, but it becomes circular when the AI deployment itself changes the demand curve that determined the staffing forecast. For finance teams approving AI investments, the question is always: what exactly would have happened without this system, and can we show our work.
Based on reporting from Enterprise AI ROI in 2026: Business Examples of Savings, Revenue and Payback, originally published 2026-10-14 11:40:00.
