Build the AI Business Case That Survives Finance Review

Workflow area: AI Opportunity & Feasibility · Tier: Advanced · Hands-on time: ~45 minutes

MIT's NANDA finding — roughly 95% of GenAI pilots showing no P&L impact — is not a technology indictment; it is a business-case indictment. Pilots launched against asserted baselines can never prove improvement, benefits stated as point estimates hide their own uncertainty, and cost stacks that stop at the license line understate spend by multiples. The analyst's differentiator is a case built like evidence: measured before, ranged benefits, complete costs, and kill criteria agreed while everyone is still friends. The LLM helps draft and structure; the arithmetic stays auditable because the spreadsheet, not the model, is the calculator of record.

When to use — and when not to

• Use it when an opportunity has survived triage and a feasibility signala business case before a spike result is a pitch; after, it is an investment analysis.
• Use it to rescue a sponsor's draft caserunning the cost-stack and baseline checklist against an existing case is the fastest honest review you can offer.
• Use it when finance will actually review the askthis format is designed for the CFO's three questions: where did the baseline come from, what does it really cost, and how do we know if it isn't working?
• Skip it for experiments under the approval thresholda one-day spike needs a hypothesis, not a benefits model; save the machinery for real asks.
• Skip letting the LLM do unaudited arithmeticmodels drop rows and invent subtotals; the AI drafts structure and prose, the spreadsheet computes, you reconcile.
• Skip cases built on baselines nobody measuredif the current cost or cycle time is a corridor estimate, the first project task is measurement, and the case says so.

The recipe

1. 1.

   Measure the baseline — or schedule its measurement. From operational data (or a two-week sampling exercise), establish the current state: volume, unit handling time, error/rework rate, cycle time, and the fully loaded cost per unit. Record the method and period. "We believe the team spends 30% of its time on this" is not a baseline; it is the reason pilots can't prove anything.

2. 2.

   Build the benefit model as ranges with shown math. In the spreadsheet, model each benefit line as low/expected/high with a stated confidence and an explicit driver chain:

   Copy

   B1 Triage time reduction
   Driver chain: 41,200 cases/yr × 14 min avg handling ×
   (30% | 45% | 60%) reduction × ₹9.2/min loaded rate
   = ₹15.9L | ₹23.9L | ₹31.9L /yr   Confidence: MEDIUM
   Basis: spike result 38-52% on 25 sampled cases; adoption
   drag not yet observed → low bound assumes 30%.
   B2 Error-rework reduction ... Confidence: LOW (no spike
   evidence; industry analogy only — flagged)
   

   Use the AI to critique the model, not to compute it: "List assumptions I haven't stated, benefit lines that double-count each other, and drivers that would not scale linearly." Double-counting between time-saved and error-reduced lines is the classic finding.

3. 3.

   Assemble the full cost stack. Draft the checklist with the AI, price it from real sources: licenses/subscriptions; token or consumption costs modeled from expected volume at provider list prices (with a usage-growth scenario — per-unit token math × annual volume, shown); integration and engineering; data preparation; security and compliance review; change management and training (the line most often zeroed, and the one adoption dies without); ongoing evaluation and monitoring (building and running the eval set is a permanent cost, not a project one); model-change regression work; and the human review time the augmented workflow still consumes. Every line has a basis note; "TBD" is acceptable, silence is not.

4. 4.

   Name the null result and the kill criteria. Write the sentence most cases omit: "If, by [date], [metric] has not improved by at least [threshold] against the measured baseline, this initiative stops, and the residual value is [what survives — the eval set, the cleaned data, the process map]." Pre-agree it with the sponsor. Kill criteria negotiated after launch are never invoked; that is how zombie pilots are born.

5. 5.

   Run the sensitivity and the red-team pass. Compute NPV or payback across the benefit range and a token-cost stress case, then have the AI attack the case as a hostile CFO: unsupported baselines, missing costs, optimistic adoption curves, benefits claimed twice. Fix or concede each hit in writing — the concession log goes in the appendix and reads as credibility, because it is.

6. 6.

   Ship the two-page case plus the auditable model. Page one: problem, measured baseline (method stated), recommendation, benefit range with confidence, total cost of ownership over three years, payback range, kill criteria. Page two: cost stack table, key assumptions with owners, risks, and the measurement plan (who measures what, how often, against which baseline). Attached: the spreadsheet where every figure traces to a cell with a basis note. The AI drafts the prose from the model — numbers flow model-to-prose, never the reverse.

Make it yours

• Spreadsheet + LLM as the standing pattern: the model in Excel/Sheets with a basis-note column; the LLM critiques structure, drafts narrative, and generates the sensitivity commentary — reconcile every number it quotes back to the sheet before shipping.
• TCO templates per organization: turn the cost-stack checklist into your org's standard template so cases are comparable across the AI portfolio and the "forgotten lines" get harder to forget.
• Provider pricing pages as primary sources: token math from current list prices with the date captured — pricing moves, and a dated source is what makes the model refreshable.
• Program mode: for a portfolio of cases, keep baselines, benefit ranges, and kill criteria in one register reviewed quarterly — the register is what lets an organization actually kill the losers and double down on the winners.

Quality check

• The baseline is measured, with method and period statedor the case explicitly schedules measurement as its first milestone.
• Benefits are ranges with confidence and basisno point estimates, no line without a driver chain a reviewer can recompute.
• The cost stack is completetokens, change management, and ongoing evals present with basis notes; any TBD is visible, not silent.
• Double-counting was huntedoverlapping benefit lines reconciled; the critique pass and its resolutions are on record.
• Kill criteria are pre-agreed and datedthreshold, date, decider, and residual value named, with the sponsor's sign-off.
• Every number is auditableprose traces to spreadsheet cells; the LLM computed nothing that ships.