You have been asked for a number. Your CFO wants it. Your board wants it. A vendor helpfully provided one. It says your AI program is producing a 3.7x productivity multiplier on a $90 per-employee monthly spend. You put it in a slide. The board nodded. Six months later, headcount hasn’t moved, contractor spend hasn’t moved, and the same CFO is asking why the line item is bigger than last year.
You no longer want to talk about the 3.7x. Nobody does. The number was never going to survive contact with a P&L.
Someone dressed up a decision problem as a measurement problem. The ROI framework produced a number nobody acted on. Theater. Seventy-four percent of organizations still can’t demonstrate measurable AI ROI after two or more years of spend.1 Not because they’re bad at measurement. Because the instrument doesn’t measure what AI actually does.
The ROI instrument was built for a different shape
The ROI framework you’re being sold assumes AI returns look like SaaS returns: uniform, per-seat, reducible to a cost-per-unit calculation. SaaS returns do look like that. A CRM produces a small, steady gain across every seat. AI doesn’t.
AI returns are lopsided. A small fraction of your users produce most of the output. The Federal Reserve’s labor data puts average productivity gain at around 5.4 percent of work hours. Under that average, engagement data shows a 6x gap between power users and everyone else.2 The average erases the only thing worth measuring.
AI returns are lagged. The champion who automates the monthly close in February doesn’t show up in the budget until the next hiring cycle decides not to backfill the role. The marketing team that ships campaigns without an agency doesn’t show up until the agency contract comes up for renewal nine months later. Quarterly ROI snapshots photograph a process whose financial signal arrives a year late.
AI returns are invisible to P&L lines. Anthropic’s own internal data shows 27 percent of AI-assisted work consists of tasks that wouldn’t have happened without the tool.3 Not cost saved. Not time reduced. Work that didn’t exist before and now does. A P&L line designed for cost savings can’t register that. Neither can a vendor-supplied productivity index. (The same asymmetry explains why some workflows compress and others don’t — the gains cluster where verification is cheap.)
Lopsided, lagged, invisible. The 3.7x embarrasses you because the instrument can’t see where the gains landed.
Three layers that produce decisions, not numbers
Give up the single number. You get three layers of measurement instead, each on its own clock, each producing a decision you can act on. (Measuring Returns is the full framework.)
Seat-level, monthly. Every seat in your AI budget gets one question every thirty days: did this person’s work observably change? Not “did they log in.” Not “how many prompts did they send.” Did something they shipped get faster, get cheaper, or happen that wouldn’t have otherwise. The data sources are vendor telemetry and five minutes with the person’s manager. If the answer is no for sixty consecutive days, reallocate the seat to someone on the waitlist who wants one. (Evaluating Spend is the mechanics.)
Workflow-level, quarterly. List the five to ten recurring workflows where AI is most likely landing. Month-end close. Contract markup. Campaign briefing. First-pass code review. For each, ask whether cycle time or staffing has moved by half or more in the last ninety days. If yes, invest more. If no, leave it for another quarter. If the workflow runs on a champion’s personal account, formalize it before they leave.
Org-level, annually. Before headcount planning, look for the asks that didn’t come in. The team that was supposed to grow from eight to twelve and asked for ten. The contractor renewal that came in lower. The backfill that got absorbed. These are the largest financial returns your AI program will produce, and they appear nowhere on a vendor dashboard. They appear in the diff between this year’s workforce plan and last year’s.
Three layers. Three decisions. You lose the aggregated productivity number. You gain a program you can steer.
Three of four signals mean it’s landing
When an AI program is actually landing, you see at least three of these four changes inside twelve months. Fewer than three means the program isn’t working, regardless of what the dashboard says.
A recurring process takes less than half the time it used to. Month-end close. Board prep. Pipeline review. Someone can name the change and point at the before and after.
A headcount request didn’t come in. A team scoped to grow asked for less. A backfill didn’t get filed. A contractor scope came in smaller. Cleanest financial signal in the entire program. Nobody writes a slide about a thing that didn’t happen, which is why your ROI framework missed it.
A vendor or contractor line item dropped. Agency spend down. Outside counsel on routine matters down. A SaaS tool canceled because the workflow now runs inside a chat window.
Internal artifacts are spreading across teams. A prompt template that started in one team and spread to three. A script someone built that others depend on. An internal tool that replaced a request to engineering. Strongest leading indicator. It means the gains are no longer trapped inside one person.
Three of four inside a year. That’s the test. Everything else is noise.
Your CFO wants a paragraph, not a multiplier
The CFO doesn’t want a multiplier. They want a defensible paragraph for the budget conversation. Here it is.
AI program spend this quarter was $X. Three workflows show measurable cycle-time reductions of 50% or more: [name them]. Two headcount requests that were on the workforce plan didn’t come in this cycle, representing approximately $Y in avoided cost. We removed N idle seats and reallocated them to the waitlist. Next quarter we’re increasing investment in [workflow] and piloting [tool change].
That’s the report. Spend, observable changes, headcount avoided, decisions made. The absence of a fake number is the credibility signal. Any CFO who has lived through vendor-supplied productivity claims will notice the difference.
The Defensible AI Spend argument pairs with this one. That piece covers defending the shape of the spend. This one covers proving the spend is producing something. Together they’re the two paragraphs that close the budget meeting.
Something to carry
Stop trying to calculate AI ROI. The instrument doesn’t fit the thing you’re measuring. Run the seat audit this month. Ask five managers whether anything observably changed. Pull the workforce plan diff from last year. Those three moves will tell you more about your AI program’s returns than any framework a vendor hands you.
If you can’t point at three of four after twelve months of meaningful spend, the program isn’t landing. More training won’t fix it. Neither will a better spreadsheet. The work in front of you is in Driving Adoption.
Footnotes
-
Multiple sources converge on this figure. McKinsey’s 2025 State of AI report: only 39% of enterprises can attribute any EBIT impact to AI investments. Grant Thornton’s 2026 AI Impact Survey: 74% of organizations report inability to demonstrate measurable ROI. Gartner 2026 data: fewer than 30% of CIOs report confidence in their AI ROI measurement. ↩
-
OpenAI engagement data, 6x gap between power users and median users. Federal Reserve labor analysis, 2025: average AI productivity gain ~5.4% of work hours across all users. See Recognizing Leverage. ↩
-
Anthropic internal analysis of Claude usage patterns across enterprise deployments, cited in Anthropic’s Q1 2026 enterprise report. ↩