A joint CFO and CAIO framework for AI benefit realisation

IBM's 2025 CEO study found only twenty-five percent of AI initiatives delivered their expected return and sixteen percent scaled enterprise-wide. Gartner's 2026 finance research found sixty-three percent of finance organisations saw AI implementation run slower than expected. Ownership is part of the reason. IBM's 2026 studies found seventy-six percent of organisations now have a Chief AI Officer, up from twenty-six percent, and sixty-two percent of CFOs say their role has expanded into AI strategy. Two executives now hold half the problem each. The fix proposed here is four controls, each prepared by the CAIO and signed by the CFO: a signed baseline, an agreed attribution method, a cost-per-outcome line, and stage-gates with kill criteria written before the pilot starts.

A baseline taken after a pilot starts is a reconstruction, and it flatters the pilot. Drawing on UK Infrastructure and Projects Authority guidance, the baseline record has seven parts: the outcome defined the way finance would count it, an owned data source, a measurement period covering a full business cycle, the natural variance in the number, a quality measure alongside speed, the current cost to serve, and both the target and the kill threshold fixed in advance.

Attribution methods are ranked by confidence: a randomised holdout first, then a staggered rollout analysed with difference-in-differences, a matched comparison, a trend-adjusted before-and-after, and last, self-reported estimates, which should never be counted as realised benefit. The research shows why the grade matters. In Dell'Acqua and colleagues' study of 758 BCG consultants, AI users completed 12.2 percent more tasks and worked 25.1 percent faster, but were nineteen percent less likely to be correct on a task outside the AI's frontier. Benefits of different evidence grades should never be summed into one total, savings should never be double counted, and released time is capacity, not cash, until it is redeployed or removed.

Per-token prices are falling fast, more than 280-fold for GPT-3.5-level performance between late 2022 and late 2024 according to Stanford's AI Index, but Gartner notes agentic models use five to thirty times more tokens per task, so total cost can still rise. The cost-per-outcome line therefore covers model inference, retrieval, evaluation, platform share and human review, the component most often left out and frequently the largest. Funding then runs through four gates: Gate 0 funds the pilot once baseline, attribution, target and cost ceiling are signed; Gate 1, at eight to twelve weeks, decides to scale, extend once, or stop; Gate 2 approves scaling only if unit benefit exceeds unit cost at projected volume; Gate 3 reviews realised benefit at six and twelve months. The stop rate itself is worth reporting to the investment committee as a measure of whether the gates are working.

The Singapore context sharpens the timing. Budget 2026 announced AI Missions in four sectors including finance and extended the Enterprise Innovation Scheme to AI expenditure, and MAS's final AI risk management guidelines, issued on 7 October 2026, expect AI inventories and board oversight. There is still no Singapore government guidance on AI benefits realisation. An investment committee can start with five questions: which use cases had a signed baseline before they started, what attribution method supports each claimed benefit, what one outcome costs today including human review, when the organisation last stopped an AI pilot and what it learned, and whether the AI inventory records benefit cases and gate decisions.