Build the AI-tooling ROI model from real budget artifacts — license quotes, inference projections, enablement time, measured benefit ranges — and write the one-page memo that counters a vendor's 4x point-estimate with defensible ranges, explicit assumptions, and the falsification criteria that make a CFO trust you.
Amara Osei is Head of Engineering at Brightline Systems, a fictional 70-engineer B2B SaaS company. Renewal season: the AI-tooling line item wants to triple, the CFO wants a business case, and the vendor has helpfully supplied one — a glossy one-pager claiming 4.2x return. Amara knows that if she forwards the vendor's number and it later collapses, she loses the budget and her credibility; and if she claims nothing, the line item dies. The winning move is the one CFOs actually respect: a range-based model with named assumptions, costs counted honestly (including the ones the vendor omits), and a written statement of what evidence would prove her wrong.
LICENSE QUOTES (annual, 70 seats)
Tool suite (IDE + chat): $39/seat/mo -> $32,760/yr
Agent platform add-on: $20/seat/mo -> $16,800/yr (30 seats
proposed, not 70 -> $7,200/yr)
Enterprise controls tier: +$8/seat/mo -> $6,720/yr
INFERENCE / USAGE PROJECTIONS (finance-reviewed)
Metered agent usage: $1,400/mo current -> projected
$2,600/mo at scale = $31,200/yr
Overage variance band: +/- 40% (usage-driven)
ENABLEMENT ESTIMATES (from the rollout plan)
Training time: 70 eng x 12 hrs onboarding + 2 hrs/mo
ongoing = 2,520 hrs/yr
Champion time: 4 champions x 4 hrs/wk = 832 hrs/yr
Loaded cost per eng-hour: $85
Review-standard setup: 120 hrs one-time
MEASURED BENEFIT RANGES (from Brightline's own 2-quarter pilot,
paired metrics per Lab 08 discipline)
Time reallocated to higher-value work: 3.1-5.4 hrs/eng/wk
(surveyed + task-sampled; midpoint 4.2)
Quality delta: change-failure rate flat; churn +1.1pt (cost!)
Est. rework cost of churn delta: 0.4-0.9 hrs/eng/wk
Onboarding: time-to-10th-PR down 6.2 -> 4.9 wks (12 hires/yr)
VENDOR ONE-PAGER (excerpt)
"Customers see 4.2x ROI. Basis: 55% of engineers report saving
8+ hrs/wk; at $85/hr fully loaded, 70 engineers save $2.47M/yr
against a $46K license investment. ROI = 4,200%. Assumes full
seat utilization and productivity gains flowing to delivery."
Total the real cost stack first — no AI, ten minutes. Licenses ($32,760 + $7,200 + $6,720), inference ($31,200 ± 40%), enablement (2,520 + 832 + 120 hrs × $85). Write the number down before modeling benefits: most engineering ROI cases die because the CFO finds a cost the author didn't count — and enablement hours at loaded cost is the one they always find. Sanity total: roughly $375K/yr, of which licenses are barely a fifth.
Structure the model as ranges, not points. Three scenarios — conservative / central / optimistic — driven by the measured bands: time reallocated (3.1 / 4.2 / 5.4 hrs/wk), churn rework cost subtracted (0.9 / 0.65 / 0.4 hrs/wk), inference at +40% / projected / −20%. Prompt your AI for the skeleton: "Build a 3-scenario annual ROI table: benefit = 70 engineers × (net hours/wk) × 46 wks × $85; costs = licenses + inference + enablement as given; output net value and ratio per scenario." Then verify its arithmetic by hand on the central case — models transpose numbers under exactly this kind of load.
Value the time honestly. The model's most attackable line is hours-saved × loaded rate = dollars. Pre-empt the CFO's objection yourself: reallocated hours are only worth $85 if they flow to work of at least that value. Handle it the credible way — count net hours (reallocated minus churn rework), apply a realization discount (e.g., 60–80% of net hours produce delivery value; make it a named, sensitivity-tested assumption), and separately list the two benefits that don't route through the hours logic: onboarding acceleration (12 hires × 1.3 wks saved × loaded cost) and flat CFR at higher throughput.
Run the sensitivity. Identify which assumption moves the answer most: realization discount, hours band, or inference variance. One tornado-style ranking in the sheet. The finding you should reach: the case survives the worst plausible inference bill easily, and lives or dies on the realization discount — which is precisely where the memo's honesty must concentrate.
Dissect the vendor's 4.2x — line by line. Take each load-bearing assumption in the one-pager and test it against your own artifacts. Log the three inflations (they are findable and specific), then reprice the vendor claim using your measured data: what multiple survives?
Write the memo. One page: the ask (total cost, honestly stacked); the defensible range (net value and ratio per scenario, central case first); the three named assumptions doing the work (realization discount, hours band, churn cost) with your basis for each; one paragraph on why the vendor's 4.2x is not your number, without theatrics; and the closer that separates you from every other budget request the CFO reads this month — falsification criteria.
Make the falsifiers real. Three to four written commitments: "If task-sampled reallocated time drops below 2.5 hrs/wk for two consecutive quarters, this case is void and we renegotiate seats." "If churn-attributed rework exceeds 1.5 hrs/wk, the quality cost exceeds our conservative bound and we pause expansion." "If onboarding time regresses to baseline, that benefit line is removed." Each falsifier names the metric, the threshold, the measurement source, and the action.
Red-team as the CFO. Prompt: "You are a skeptical CFO. Attack this memo's weakest three lines." Patch or pre-empt each in the memo. If the AI's attack list doesn't include the hours-to-dollars conversion, attack it yourself — it is always the weakest line.
A CFO who sees falsification criteria in a budget memo learns the author expects to be audited — which is exactly what makes the numbers believable before any audit happens.
A complete lab has all of the following:
The ROI case a CFO can't dismiss is the one that counts every cost, converts hours to dollars through a named discount, reports a range it can defend at both ends — and volunteers the evidence that would kill it.