How to Build an AI Business Case Your CFO Can Audit

The steering committee is looking at a slide that promises $1.8 million in annual value. Finance asks one question: “Show me the queue that gets smaller.” The room goes quiet. The estimate came from hours multiplied by salaries, but nobody measured current handling time, checked whether saved minutes become usable capacity, or included review and correction work. This is where most AI business cases stop being cases and become aspirations.

Start with a decision, not a technology

A useful AI business case asks whether a specific operating change should be funded. Name the workflow, the people who own it, the system boundary, the eligible volume, and the decision that follows the pilot. “Deploy generative AI” is not fundable scope. “Reduce the time and rework required to prepare 400 monthly renewal briefs without increasing material errors” is.

Call this the traceable-value chain: baseline, intervention, accepted output, operating consequence, financial consequence. Every number should move through all five links. If the tool drafts faster but reviewers spend the saved time correcting unsupported claims, the accepted-output link breaks. If accepted work is faster but staffing, throughput, or customer outcomes do not change, the financial link breaks.

Build the baseline finance can reproduce

Measure at least four weeks of current work before forecasting improvement. Use completed cases, not recollection. Capture eligible case volume, median and 90th-percentile completion time, active human time, wait time, rework rate, error severity, and cost per accepted output. Separate touch time from elapsed time: removing 20 minutes of drafting does not shorten a three-day approval queue unless the approval stage also changes.

• Volume: how many comparable cases enter the workflow each week, and what share is actually eligible for AI assistance?
• Labor: which roles touch a case, for how long, at what loaded hourly cost?
• Quality: what share requires correction, escalation, or customer recovery?
• Delay: where does work wait, and does delay affect revenue, service, or compliance?
• Constraint: which stage limits throughput today, and will the pilot relieve that stage or merely feed it faster?

Use four value buckets—and do not mix them

Hard savings remove an approved expense: a contract ends, overtime falls, or a vacancy is not backfilled. Capacity value releases time that can be reassigned to named work. Revenue value requires a causal route from the intervention to conversion, retention, price, or volume. Risk value reduces the expected cost of a defined failure. Finance will discount each bucket differently, so present them separately instead of adding every optimistic benefit into one return figure.

A worked example makes the distinction concrete. This is arithmetic, not a customer result. Twenty reviewers each spend six hours a week on eligible preparation work. At a $72 loaded hourly cost across 46 working weeks, the gross time pool is $397,440 a year. A pilot that removes 35% of that work appears to release $139,104. But if review, exception handling, licenses, and support cost $74,000, the net capacity value is $65,104—and it is not cash savings unless leaders remove expense or redeploy the hours to measured work.

Grade every claim before it reaches the CFO

Use a three-level evidence ledger. Grade A is your measured baseline or controlled pilot. Grade B is a comparable internal workflow or an external study with a visible denominator and method. Grade C is a vendor case, anecdote, or assumption. Grade C can help set a test range; it should not set the committed return. Record the source, date, population, definition, owner, and confidence beside every material claim.

Microsoft’s September 2026 account of its internal AI transformation is a useful example of both evidence and restraint. It reports selected workflow gains and explains that broad usage plateaued before teams redesigned work around business outcomes. The cases are vendor-reported and not universal benchmarks. Borrow the operating pattern—workflow ownership, shared data, measurement—not the percentages.

Price the system around the model

The license is often the smallest visible number. Include implementation, integration, data preparation, security review, evaluation, human review, exception handling, training, change management, monitoring, incident response, and exit. Model costs at base volume and at a plausible peak. If usage-based pricing or downstream API calls apply, calculate cost per accepted output rather than cost per generated response.

Pre-register the pilot and its stop rules

Before the tool touches live work, write the eligibility rule, baseline window, comparison method, minimum sample, owner, review requirement, success threshold, and stop condition. A sensible scorecard pairs one outcome measure with one quality measure, one risk measure, and one adoption measure. Do not change definitions halfway through because the first chart is disappointing.

• Outcome: accepted cases per week or end-to-end cycle time—not prompts, drafts, or logins.
• Quality: material correction rate, rework minutes, or first-pass acceptance.
• Risk: privacy incidents, policy exceptions, unsupported claims, or wrong actions by severity.
• Economics: total pilot cost and cost per accepted output compared with the baseline.
• Stop rule: a pre-agreed quality, safety, or cost boundary that pauses the pilot.

The one-page approval memo

Page one should state the workflow and owner, current baseline, proposed change, eligible population, 90-day cost, expected range by value bucket, evidence grades, success and stop rules, decision date, and what happens after a pass or failure. Put assumptions and sensitivity math in an appendix. A CFO should be able to change volume, adoption, accuracy, review time, and price without rebuilding the model.

The strongest AI business case is not the one with the largest return. It is the one the company can disprove quickly, learn from safely, and scale without changing the arithmetic. If finance can trace the value from a smaller queue to an operating consequence, the proposal is ready. If it cannot, another percentage will not rescue it.

Primary sources

• NIST AI Risk Management Framework — A primary framework for governing and measuring AI risk across the lifecycle.
• Microsoft: lessons from its own AI transformation — Vendor-reported internal cases, methodological notes, and the usage-plateau lesson.
• U.S. Census Bureau: business AI use — National adoption context and survey definitions.