Build the AI Tooling ROI Case a Skeptical CFO Will Sign

Workflow area: Measurement & ROI · Tier: Advanced · Hands-on time: ~45 minutes

Engineering leaders lose the budget conversation two ways. Overclaiming: the "35% productivity gain" slide gets one hard question — "then why isn't output up 35%?" — and the whole case dies with the number. Underpreparing: real benefits presented as anecdotes get politely deferred to next cycle. The winning posture is the one CFOs use on each other: conservative arithmetic, visible assumptions, and pre-registered conditions for being wrong. Everything this playbook needs, pb-37's honest measurement system has already collected; this is where the discipline pays out in currency finance respects.

When to use — and when not to

• Use it at renewal and expansion decisionsthe pilot's grace period is over, seats are scaling, and "the engineers like it" has stopped being an accepted line item.
• Use it when finance builds the case without youif the alternative analysis is seat cost versus a vendor white paper, you want your version on the table first.
• Use it to kill or shrink your own spendan honest case sometimes concludes the expensive tier isn't earning its delta over the cheap one; making that call yourself, in writing, is the single best deposit into your credibility account.
• Skip it for trivial spenda case for a $2k/year tool costs more to build than it governs; batch small tools into one portfolio line and save the machinery for the six-figure decisions.
• Skip the vendor's numbers as inputsvendor case studies and calculators are marketing priors, admissible as context, never as arithmetic; the case runs on your pb-37 measurements or it isn't ready to run.

The recipe

1. 1.

   Count the full cost — especially the four everyone omits. Licenses and token/usage spend are the visible tier. Add: enablement time (training hours × loaded rate — a 200-engineer rollout with two days of ramp is a six-figure cost before the first PR); review overhead (the pb-27 standard's added review depth, estimable from review-time deltas in your telemetry); platform and governance work (the pb-28 gates, pb-30 guardrails, and telemetry pipelines are real engineering weeks); evaluation and switching reserve (the pb-35 bake-off cadence, and what migrating away would cost — finance will ask). A cost model missing these reads as advocacy the moment anyone competent inspects it.

2. 2.

   State benefits as measured ranges with their pairing intact. From pb-37: lead-time deltas with change-failure rate alongside, throughput with revert rates, survey-based friction reductions — each as a range reflecting the confound log ("8–14% lead-time improvement attributable within noise bounds"), never a point. Where a benefit is real but unmeasured (retention effect, hiring signal), list it in a separate unquantified section at zero value in the arithmetic — visibly counting something at zero buys more trust than optimistically pricing it.

3. 3.

   Build three scenarios in a sheet finance can audit. Conservative / central / optimistic, with every assumption a labeled cell, not a buried constant:

   Copy

                        CONSERVATIVE   CENTRAL   OPTIMISTIC
   Adoption (active/seats)     55%        70%        85%
   Lead-time gain               6%        10%        14%
   Quality cost (CFR delta)    +1pt        0        -1pt
   Full cost (₹/yr)           1.9Cr      1.7Cr      1.6Cr
   Net benefit (₹/yr)         0.4Cr      1.6Cr      3.1Cr
   

   The conservative column is the case: if it clears zero using measured lower bounds and full costs, you have a decision, and the other columns are upside, not evidence. Have Claude red-team the sheet before finance does:

   Copy

   Here is my AI-tooling ROI model [PASTE SHEET STRUCTURE AND
   ASSUMPTIONS]. Act as a skeptical CFO's analyst. List: (1) every
   assumption that most favors the case, (2) costs that appear
   missing or underweighted, (3) any benefit double-counted across
   lines, (4) the three hardest questions to ask in the review
   meeting. Do not soften. Do not invent numbers — question mine.
   

4. 4.

   Write the falsification section — before the meeting, not after. One box in the memo: this case is wrong if… — adoption stays under the conservative floor by Q2; the paired quality metrics degrade beyond the stated bound; the lead-time range's lower edge isn't sustained over two clean quarters; usage spend grows past the modeled ceiling. Each condition gets a check date and a pre-committed response (renegotiate tier, shrink seats, re-run pb-35, exit). This section is the case's engine: it converts "trust my projection" into "here is the tripwire, audit me on it" — the only argument that gets stronger under skepticism.

5. 5.

   Deliver as a one-page memo over the auditable sheet. Recommendation and net range up top; cost table; benefit ranges with the pb-36-style confound note; scenarios; falsification box with dates; the unquantified list at zero. Then honor the dates: the Q2 check happens, its result goes back to finance unprompted — including "conservative case holding" or "tripwire 2 fired, here is the response." The second budget cycle is where this playbook compounds; you arrive as the one engineering leader whose last case audited itself.

Make it yours

• Small team or startup: the sheet is twenty rows and the memo is half a page, but keep all five parts — especially falsification; investors and founder-CFOs respect a self-auditing case at any scale, and small-company spend decisions reverse faster, making the tripwires genuinely useful rather than ceremonial.
• Large org or GCC: split the model per business unit or client segment — adoption and benefit ranges differ by codebase age and engagement posture (pb-33 constraints cap benefits on Red accounts; say so); for GCCs, present the parent company the same conservative-case discipline, because the alternative is their finance team applying a group-wide vendor benchmark that fits nobody.
• Token-heavy agent workloads: usage-based spend breaks the per-seat model — model tokens as a variable cost with a per-workflow ceiling, wire the pb-30 spend caps in as the enforcement mechanism, and give runaway-cost its own falsification tripwire; agent economics can invert from accretive to dilutive inside a quarter without one.

Quality check

• The conservative case stands alonecover the central and optimistic columns and re-read the memo; if the recommendation no longer holds, the case is a hope with a spreadsheet.
• Every benefit traces to a pb-37 measurementwalk each benefit line back to its source metric and confound note; any line that traces to a vendor asset or an anecdote gets moved to the zero-valued list.
• The four hidden costs are pricedenablement, review overhead, governance engineering, switching reserve; their absence is the first thing a competent finance reviewer finds, and finding it costs you the room.
• Falsification dates are on calendarsyours and finance's; an unscheduled tripwire is a rhetorical device, and this section only works as an actual mechanism.
• The memo survives hostile summaryhave someone compress it to two sentences; if the compression is "AI makes us 30% faster," the ranges and pairings aren't prominent enough — the honest headline should be the only extractable one.