Building the ROI Case for an AI Product

By Brian Tr · Updated Oct 6, 2026

An ROI case for an AI product has to answer a harder question than it does for ordinary software. The buyer isn't only asking "is the benefit bigger than the price?" They're asking whether the benefit is real, whether it will show up in a line the finance team can see, and whether the true cost, which moves with usage and needs people to supervise it, was counted honestly.

This page is written for the AI vendor and for the champion inside the buyer who has to defend the purchase. It covers the value levers, why "hours saved" alone usually fails, how to set a baseline, what total cost of ownership (TCO) looks like for AI, how to run conservative, expected and optimistic scenarios, and how the evidence should change from the pre-sale estimate to the pilot to the renewal. For the generic formula and the document structure, see ROI and business case creation. This page covers what's specific to AI.

The Six Value Levers

Most AI business cases draw on the same handful of levers. Naming them separately matters, because each one is measured differently and each one convinces a different person on the buying committee.

A credible case usually leads with one or two levers and treats the rest as upside it doesn't count. A case that claims all six at once reads as a wish list, and a finance reviewer will cut it down to whatever they can verify.

Why "Hours Saved" Alone Rarely Survives a CFO

Time saved is the most natural claim for an AI product and the easiest one to puncture. The reason is simple: time saved is not cost removed.

If a tool saves each person 30 minutes a day, nothing changes on the income statement unless something happens with that time. Either the company needs fewer people (a headcount or contractor reduction, or a hiring plan that no longer needs a role), or the same people handle more work (throughput), or they move to work that produces revenue. Until one of those happens, the saving is real for the employee and invisible to the CFO.

So the question to put in the case is not "how many hours are freed?" but "where do those hours go?" Good answers are specific:

• A backlog that currently needs two extra hires won't need them.
• A team that turns away work because it's at capacity will take it on.
• Reps get 30 more minutes for selling and the case states what that time is expected to produce.

A vendor who answers this before the CFO asks is usually the one who gets through the review. The same discipline shows up in value selling: value is the outcome the buyer can state in their own terms, not the activity your product speeds up.

One more caution. A time-saved figure that arrives with no population or task attached is the weakest evidence in the room. Which brings us to the published research.

What the Published Evidence Does and Doesn't Support

Three widely cited studies give vendors and champions real numbers to work with. Each measures a specific population on a specific kind of task, and the case should say so.

Customer support agents. Brynjolfsson, Li and Raymond studied the staggered introduction of a generative AI conversational assistant using data from 5,179 customer support agents. In their NBER working paper, access to the tool increased productivity, measured by issues resolved per hour, by 14% on average, including a 34% improvement for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers. This is one firm's support operation. The 14% applies to issues resolved per hour among those agents, not to other jobs.

Professional writing tasks. Noy and Zhang ran a preregistered online experiment with 444 college-educated professionals doing occupation-specific, incentivized writing tasks, with half randomly exposed to ChatGPT. In their working paper (March 2023, not peer reviewed in that version), time taken decreased by 0.8 standard deviations and output quality rose by 0.4 standard deviations, and the productivity distribution was compressed because low-ability workers benefited more. These are effects on short, mid-level writing tasks in an experiment, not on a full working week.

Consultants and the "jagged frontier." The Harvard Business School AI Institute summary of the study done with Boston Consulting Group covers 758 consultants, about 7% of the firm's individual contributor workforce. For tasks within the AI frontier, the tool boosted speed by over 25%, human-rated performance by over 40% and task completion by over 12%. The researchers describe a "jagged technological frontier" where AI excels in some tasks but falls short in others. The practical lesson is that the gain depends on whether the buyer's task sits inside the frontier, and nobody should assume it does.

Three things to take from these when building a case:

1. Match the population. A 14% result from support agents is evidence for support-style work, not for finance or engineering. Say which task your result applies to.
2. Expect uneven gains. In the support study the benefit went mostly to novices. If your buyer's team is mostly experienced, the average result may not transfer. The buyer's own pilot is the evidence that does.
3. Treat outside research as a prior, not a promise. It makes your estimate plausible. It doesn't replace measuring the buyer's workflow.

Establish a Baseline Before Anything Else

An ROI claim is a comparison, and a comparison needs a starting point. Without a baseline, "improved by 20%" is an assertion nobody can check.

Capture the current state of the metric before the product touches the workflow:

• Volume and mix: how many tasks, of what types, per period.
• Effort per task: handling time or cost per task, measured rather than guessed.
• Quality: error rate, rework rate or escalation rate on those same tasks.
• Fully loaded cost: salary plus benefits and overhead for the people doing the work.
• Seasonality: a before period that includes a quiet month and a busy one will mislead.

Use the buyer's data. Industry averages are fine for sanity checks, but a figure the buyer measured is the one their finance team will trust, as business case creation also stresses.

Then choose how you'll compare. There are two honest options:

A control group is the stronger design, and it's what the support study's staggered rollout gave its authors a version of. Where a control isn't possible, a before-and-after comparison works if you also record what else changed during the period, such as new hires, a pricing change or a seasonal peak.

Total Cost of Ownership for an AI Product

The most common weakness in a vendor's ROI case is a cost side that stops at the subscription. A finance reviewer will build the fuller picture themselves, and a vendor who shows it first looks more credible.

Two items deserve emphasis.

Variable fees move the answer. If part of the cost depends on usage, then success raises the bill. A case should show cost at the expected volume and at a higher volume, so the buyer sees what happens if adoption beats the plan. Usage-based versus outcome-based pricing explains how these structures change the buyer's exposure.

Human review is a real cost. If a person still checks a share of outputs, that time belongs in the model. Leaving it out is the quickest way to have a CFO distrust the rest of the figures. The review rate is also something the pilot can measure, which is why a pilot is the right place to replace guesses.

A Worked Example (Illustrative Only)

The numbers below are invented to show the arithmetic. They don't describe a real company or a real product, and your figures will differ.

Suppose a support team handles 20,000 tickets a month at 0.25 hours each, which is 5,000 handling hours a month. Fully loaded labor costs $30 per hour, so handling costs $150,000 a month.

Costs. Platform fee of $48,000 a year. Usage fees of $0.15 per ticket handled with the tool, which is 20,000 x 12 x $0.15 = $36,000 a year at this volume. Human review of 100 hours a month at $30 is $3,000 a month, or $36,000 a year. One-time costs are $20,000 for integration and data preparation and $12,000 for change management and governance. That gives ongoing costs of $120,000 a year ($10,000 a month), one-time costs of $32,000, and a year-one total of $152,000.

Benefits. The tool reduces handling time by 10%, 20% or 30% depending on the scenario. But only part of the freed time becomes cost removed or extra work absorbed: 40%, 60% or 80% respectively.

Results, year one.

Payback is the one-time $32,000 divided by the monthly benefit minus the $10,000 monthly ongoing cost: $32,000 / $8,000 = 4 months in the expected case, and $32,000 / $26,000 is about 1.2 months in the optimistic one. In the conservative case the monthly value ($6,000) is below the ongoing cost ($10,000), so the investment never recovers.

That last row is the useful one. It tells both sides which assumptions the whole case rests on: the size of the handling-time cut and the share of freed time that becomes real value. Those are the two numbers a pilot should measure.

Sensitivity Analysis: Show the Range, Not a Point

A single ROI figure invites a single argument. Three scenarios invite a conversation about which assumptions are believable.

• Conservative: the low end of what you'd defend if the pilot disappoints.
• Expected: the figure the evidence best supports.
• Optimistic: what happens if adoption and results both beat the plan.

Vary the inputs that matter most and say which they are. In the example, the realization rate moved the result as much as the productivity gain did. If a buyer's own finance team picks the conservative assumptions, that's a good sign: they now own the case. A case that still works at their numbers is much easier to approve than one that only works at yours.

How the Evidence Changes Across the Deal

The same ROI case matures through three stages, and each stage should replace guesses with measurements.

At the pre-sale stage the honest framing is "here's the range and the assumptions behind it." At the pilot stage the aim is to replace the two or three riskiest assumptions with measurements, as described in why demos matter more for AI products, including a written decision rule. At renewal the buyer asks whether it paid off, so the report back is a comparison of actual results with the original case, including the places where reality was lower. A vendor who reports that straight is far more likely to be believed next year. The pilots-to-contracts page covers turning that evidence into a commercial commitment, and selling AI to skeptical buyers covers how to present it to people who expect to be disappointed. Finance is also usually part of the AI buying committee, so build the case for that reader.

A Checklist for the ROI Case

• Does the case lead with one or two levers and not claim all six?
• Does it say where freed time goes, and who confirms it?
• Is there a measured baseline from the buyer's data?
• Is the comparison method stated (control group or before and after)?
• Does the cost side include usage fees, integration, data preparation, review time, change management and governance?
• Are there conservative, expected and optimistic scenarios with the assumptions shown?
• Is each outside figure matched to the population and task it came from?
• Does the pilot plan measure the two or three assumptions the case depends on?

Key Facts

• Customer support agents: across 5,179 agents at one firm, access to a generative AI assistant raised issues resolved per hour by 14% on average and 34% for novice and low-skilled workers, with minimal impact on experienced workers (Brynjolfsson, Li and Raymond, NBER w31161).
• Professional writing tasks: in a preregistered experiment with 444 college-educated professionals, ChatGPT cut time taken by 0.8 SDs and raised quality by 0.4 SDs, and compressed the productivity distribution (Noy and Zhang, March 2023 working paper).
• Consultants within the AI frontier: across 758 consultants, tasks inside the frontier saw speed up by over 25%, human-rated performance up by over 40% and task completion up by over 12%, while AI falls short on other tasks (HBS AI Institute summary).
• Time saved is not cost removed: a saving counts financially only when capacity is redeployed, work is absorbed or a hire is avoided.
• Cost of ownership goes beyond the subscription: usage fees, integration, data preparation, human review, change management and governance all belong in the model.

Related Reading

• ROI
• Business case creation
• Value selling framework
• Why demos matter more for AI products
• Usage-based vs outcome-based pricing for AI