AI ROI Sensitivity Analysis: Find the Assumption That Breaks the Case

A proposed AI workflow can show a positive return while relying on one fragile assumption: that most staff will use it, most drafts will pass review, or review will take almost no time. Finance should be able to move each input and see when the case turns negative. That is sensitivity analysis: changing an assumption while holding the others steady to find what drives the decision. The U.S. Government Accountability Office’s cost guide recommends this method for credible estimates; the worked AI example below is our own illustration.

Start with accepted work, not generated work

Define one eligible case and its current active time. Then count how many cases staff will actually try with AI, how long drafting and review take, how often a case needs repair, and the cost of the tool and support. An attempted case that fails still consumes time and money. A draft is not a completed output until it reaches the old quality bar.

Separate touch time from elapsed time. If the business goal is a shorter customer wait, measure the whole handoff as well as human minutes. If the goal is capacity, name the queue or additional work that can use released time. Only record cash savings when an approved expense actually falls. Our ROI decision lab keeps those categories apart.

A small model you can challenge

The following numbers are hypothetical, not a customer result. A team receives 1,000 eligible briefs a month. The current process takes 12 human minutes per brief. In the AI process, drafting takes five minutes and review takes three. Twenty percent of attempted briefs need seven more repair minutes. Staff attempt 60% of eligible briefs. Monthly software and allocated setup cost total $1,200; a $60 loaded hourly rate is used only to value capacity.

Average AI-process time per attempted brief is five plus three plus 20% of seven: 9.4 minutes. The difference from 12 minutes is 2.6 minutes. Across 600 attempted briefs, that releases 26 hours a month. At the illustrative hourly rate, the capacity equivalent is $1,560. After the $1,200 monthly cost, the modeled net capacity value is $360. Payroll cash savings are zero unless a separate staffing or spending action occurs. The remaining 400 briefs stay in the baseline process.

Change one assumption at a time

• Adoption falls from 60% to 40%: 400 attempted briefs release about 17.3 hours. The capacity equivalent is $1,040, below the $1,200 cost.
• Repair rises from 20% to 40%: average AI-process time becomes 10.8 minutes. At 60% adoption, only 12 hours are released, worth $720 of capacity at the assumed rate.
• Review adds two more minutes to every attempt: average AI-process time becomes 11.4 minutes. Six hours are released at 60% adoption, worth $360 of capacity.

Those scenarios use the same monthly volume, hourly rate, and fixed cost. None predicts what will happen. They show which measurement can overturn the proposal. Here, review time and repair rate deserve stronger evidence before a wide purchase. The GAO cost guide also advises recording assumptions and updating estimates with actual costs.

Make the downside operational

Finance should ask for three cases: a measured base case, a plausible lower-adoption or higher-rework case, and a case where the workflow fails a quality gate. Use pilot observations for ranges when available; label guesses as guesses. Do not pile the worst possible value for every input into one theatrical number, and do not offset a real error risk with an invented revenue lift.

If an error reaches a customer, a regulated record, or an irreversible action, cost sensitivity alone is insufficient. Give that failure a separate stop rule and owner. Run the proposed workflow against the AI pilot scorecard, then update this model with accepted cases, actual review minutes, and paid invoices.

The approval question

At the funding meeting, show the smallest change in adoption, review, or repair that erases the projected benefit. If that change is within the range already seen in the pilot, fund a narrower test or fix the process first. If the estimate survives the observed range and an owner can use the released capacity, the case is ready for a controlled expansion.

Useful next steps

• Open the AI ROI decision lab — Run your measured inputs and keep capacity separate from cash.
• Build a finance-ready AI business case — Show the evidence behind each assumption.
• Use the AI pilot scorecard — Collect the acceptance and review data that the model needs.