Only 25% of AI initiatives delivered the ROI expected of them (IBM, 2025), and 56% of CEOs report no financial benefit from AI at all (PwC, 2026). Sit with the second number: it does not say the benefit was negative. It says nobody can find it, and a benefit nobody can find gets defunded in the next budget cycle regardless of whether it exists. Most AI programs do not die from lack of value. They die from unmeasurable value.
Why AI spend evades normal measurement
Three habits create the invisibility. Nobody sets a baseline, so improvement has nothing to be measured against. Value gets claimed in anecdotes and hours-saved estimates that finance correctly refuses to book. And the spend scatters across twenty tools with no owner, so no one is accountable for the aggregate. Each habit is fixable, and the fix compresses into one discipline.
The composite: one number, quarterly
Run your AI program against a single ratio: confidence-weighted value produced over AI cost, reported quarterly. The mechanics are deliberately unglamorous.
- Baseline before the build. Before any workflow is rebuilt, record its current cost, cycle time, and error rate. No baseline, no claim, ever.
- Label every value figure measured, estimated, or projected, and weight it accordingly. Measured dollars count in full; projections are discounted for honesty. Finance stops rolling its eyes the day the labels appear.
- Count all the cost: licenses, build spend, and the internal hours the program consumes. An honest denominator is what makes the numerator credible.
- Report it quarterly, same format, next to the other numbers the business runs on. One composite for the program, decomposable by workflow when someone asks.
The stop rule is the credibility
State up front the reading that kills the program: if the composite stays below the bar for two consecutive quarters after the first workflow ships, the program stops and the remaining budget returns. A program that can be stopped is a program a board can fund, which is why the stop rule accelerates approval rather than endangering it. It also changes behavior inside the building: teams pick workflows that can actually clear the bar instead of the ones that demo well.
The full scoreboard, with the worksheet and the funding offsets that improve the denominator, is in The 2026 AI Strategy, free. One number, honestly weighted, quarterly, with a stop rule. Every AI conversation in your company gets easier the quarter after it exists.
