The CFO-ready AI ROI report
April 2026 · Jason Lee
Every AI initiative eventually faces one meeting: the one where a CFO asks what the money bought. Most initiatives arrive at that meeting with screenshots and enthusiasm. This article is the report that survives it — the exact structure I deliver at day 76–90 of every engagement, published so you can hold anyone’s AI work to it, including mine.
The report is five sections and rarely more than four pages. Brevity is a feature: a CFO reads evidence, not narrative.
Section 1 — The baseline (set before anything was built)
What the target workflows cost before: hours per week, volume processed, error or rework rate, and cycle time, measured over at least two weeks. Loaded labor cost applied — salary plus benefits and overhead, typically salary × 1.25–1.4, with the multiplier stated.
This section is boring and non-negotiable. A report without a pre-intervention baseline is a story, not a measurement, and a CFO knows the difference in one glance. If a vendor or consultant presents results with no baseline, the conversation should end there.
Section 2 — Hours recovered
Per automated workflow: hours before, hours after, delta, annualized. Illustrative shape:
| Workflow | Staff-hours/week before | After | Recovered | At a $38/hr loaded rate |
|---|---|---|---|---|
| Referral intake summarization | 14.0 | 4.5 | 9.5 hours/week | ~$18,800/year |
Two disciplines keep this section honest. First, recovered hours are counted only where the after measurement is as rigorous as the before. Second, the report states what the hours were redirected to — because “we saved 9.5 hours” invites the question “then why is payroll unchanged?” The answer (capacity absorbed growth, overtime eliminated, backlog cleared, headcount avoided) is the difference between soft savings and a number finance will book.
Section 3 — Dollars saved and dollars avoided
Three lines, kept separate because a CFO will separate them anyway:
- Hard savings: eliminated spend — overtime, outsourced processing, a redundant tool license.
- Capacity value: the recovered-hours math from Section 2, explicitly labeled as capacity, not cash.
- Cost avoidance: the hire not made, the penalty not incurred. Stated conservatively and last, because it is the easiest line to inflate and the first line a skeptical reader attacks.
Then the other side of the ledger, unprompted: engagement fees, tooling and API costs, internal time spent. A report that presents its own full cost column earns the credibility every other number in it spends.
Section 4 — Risks retired
The section AI vendors never include, and in a regulated business often the largest value line: what exposure existed before that no longer does. Written as before/after facts, not adjectives:
Before: no inventory of AI tools in use; staff pasting client data into consumer tools with no agreements. After: AI register in place; enterprise agreements executed; consumer tools blocked; every prompt logged and attributable.
Risk retirement resists precise dollar figures, and the report should resist inventing them — a fabricated breach-cost projection undermines the real numbers around it. State the exposure closed; let the CFO, who prices risk professionally, do the pricing.
Section 5 — The decision
One page: results against the success criteria set in week one, and a recommendation — scale (which workflows next, at what projected return), continue (what needs another cycle and why), or stop. With the numbers above it, the recommendation writes itself, which is the point. In my engagements this page feeds a contractual decision at day 90; the report exists so that decision is made on evidence rather than momentum.
Why publish the template
Because the report is only trustworthy if you knew its structure before the work began. Baselines can’t be reconstructed after the fact, cost columns can’t be added retroactively without suspicion, and success criteria defined after results arrive aren’t criteria. Hold every AI initiative — internal, vendor, or consultant — to this structure from day one, and the stalled-pilot problem largely solves itself: what gets measured gets finished.