How to write the AI ROI report your CFO and board will both sign off on

The morning two AI stories landed at the same outlet

On the morning of May 6, 2026, Fortune published two pieces by the same business editor, Nick Lichtenberg, a few hours apart. One quoted a fresh Goldman Sachs research note arguing that the AI capex arms race is being driven by FOMO not ROI, citing MIT data showing 95% of organizations are getting zero return on their pilots, and a 2025 EY survey finding 99% of companies reported financial losses from AI-related risks averaging $4.4 million per company. The other walked through how ServiceNow had just raised its full-year AI ACV target from $1 billion to $1.5 billion, with a customer math story that turned a 65% cost cut into a 5x revenue expansion at the same account.

Same day. Same publication. Same business editor. Two completely different stories about whether AI is paying off.

The reason both were true at once is the reason most Series B AI ROI reports keep getting bounced back from the board. They are written about the wrong number.

What most Series B AI ROI reports actually look like right now

I read a lot of these. The structure is almost always identical. Three slides on tools adopted. Six on hours saved. Two on a “productivity uplift” calculated as average seat-hours-saved times average loaded labor cost. Sometimes a chart of LLM tokens consumed. Rarely a number tied to revenue.

The Snowflake “ROI of Gen AI and Agents 2026” report from March gave the gentle version of the problem. Its global survey of 2,050 leaders found that 92% of early adopters say they see positive ROI, with quantifiers reporting $1.49 returned per $1 invested. Yet executive surveys from the same quarter put significant ROI from generative AI at only 29%, and from AI agents at 23%. The gap is not really a measurement gap. It is two different things being measured. “Saw productivity” is not “moved the P&L.”

This was tolerable as a Series A report. At Series B, with a CFO and a board ratifying material AI spend on a quarterly cycle, it is not. The Goldman note this week made that explicit by quoting it back to the room.

Where the report has to break differently now

The clean version is the one ServiceNow gave Wall Street on May 6.

That single line is the entire model. The customer cuts total cost by 65%. Then the seats freed up are reinvested into AI agent consumption at 6.5 times the value, and total ServiceNow spend at the account grows over 5x. Now Assist customers who renewed in 2025 expanded their ACV by an average of 3x. The internal validation was just as clean: $500 million annualized AI-driven value generated inside ServiceNow in 2025 with $100 million in OpEx savings, and an expected $200 million more in 2026.

What that report is doing that the typical Series B report is not: tracing where the freed capacity went. The Series B equivalent question is not “how many hours did the AI save us.” It is “where did those hours show up as revenue or margin, by name.”

The week’s other in-window signal made that point clearer for buyers as well. The Jitterbit 2026 AI Automation Benchmark Report, summarized on GlobeNewswire on May 6, found that 47% of IT leaders now name “AI accountability” (security, auditability, guardrails) as the single most important factor when evaluating new AI tools, with only 15% citing budget as a meaningful challenge. Jitterbit’s own framing was that the bottleneck has moved from the CFO to the CISO. The more useful read for an operator is that enterprise buyers are now demanding a kind of evidence the old productivity-pitch report was never written to provide.

The CFO’s May 7 piece by Nikita Alexander pinned the next mover. Citing Morgan Stanley, she noted that 80% of AI benefits today derive from cost efficiency and margin expansion rather than top-line revenue, and argued that data ownership has moved from a technical concern to a leadership mandate landing squarely in the CFO’s lap. The Gartner number in the same piece (59% of finance leaders are now using AI, but nearly 25% are still unsure how to move from planning to piloting) is the exact group most Series B CFOs are in this quarter.

The pattern the board is now looking for

The four-question board test, all of it derived from the past 72 hours of evidence:

1. Why is this not FOMO spend? Goldman just publicly named the cycle. The board has the question on a tab now.
2. Where is the freed capacity reinvested? The ServiceNow 6.5x story is the proof template. Without it, the savings number reads as a capacity-reduction story, not a value story.
3. What accountability proof comes with the dollars? The Jitterbit 47% number is the same signal a Series B board reads from its portfolio’s customers, just routed differently.
4. Who owns the math? The CFO piece on May 7 settled it. Data integrity is a Finance-led artifact now.

A Series B AI ROI report that answers all four reads almost completely differently from the one most operators are still writing. Six moves close the gap.

1. Replace the productivity-uplift slide with a capacity-reallocation table.

   For each AI deployment, name the team, the hours released per week, the loaded labor cost, and the specific work or pipeline those hours were redirected toward. Tie it to a revenue or margin line. A productivity figure with no destination is the line the board now reads as FOMO.

2. Anchor the report in named accounts and named workflows, not portfolio averages.

   Boards trust four named scenarios with timelines over twenty unnamed averages. Borrow the ServiceNow disclosure model from Fortune's May 6 piece: customer scenario, before-state cost, after-state cost, and the dollar that grew somewhere else.

3. Build a single-page ROI scorecard with five lines, not thirty.

   Total AI spend run-rate, gross savings captured, net savings after stack-cost overruns (Snowflake found 95% of organizations exceed initial budget on at least one part of their gen AI stack), revenue or margin attributable to the AI workflow, and a forward 12-month commitment number. If a metric does not fit on the page, it is not a board metric.

4. Attach a one-page accountability appendix.

   Name the data owner per system, the audit trail, the kill criteria, and the incident-response path. The Jitterbit 47% buyer signal is the same signal a Series B board reads from its enterprise customers, just from a different angle. Skipping this section in 2026 now reads as missing rather than economical.

5. Have the CFO present the report, not the head of AI or the CTO.

   The May 7 CFO accountability framing makes Finance the natural owner. Finance presenting changes the type of question that follows the slides. Boards ask different questions of Finance than of Product, and the AI ROI conversation now lives in the first set.

6. Write the FOMO answer at the top of the deck, in two sentences.

   Why this spend is not Goldman's FOMO story. What is being decided differently because of the AI deployment. Two sentences. The board reads them first and grades the rest against them.

What I would tell you over coffee

Most Series B AI ROI reports I see are written by the team that built the thing they are reporting on. That is the original sin. Builders cannot write a report whose primary audience is a board member who has been reading the Goldman note in the morning, the BCG May 5 paper on lunch break, and watching ServiceNow’s earnings call after dinner.

Capacity reinvestment is the story. Capacity reduction was just the warm-up.

The fix is not a better dashboard. It is letting Finance own the math. Hand the CFO the same kind of customer-level proof story ServiceNow handed Wall Street, scoped to two or three real deployments. Let Finance write the five-line scorecard. Let Finance present it. Then watch the next board meeting end with the next funding question, instead of the last one.

The week ahead is not about producing a bigger AI ROI report. It is about producing a smaller one that survives both rooms.