Enterprise AI ROI: How to Calculate the Business Case Before You Build

Most enterprise AI ROI cases are written after the budget is approved. Somebody has already decided to build something. The spreadsheet arrives later to justify it.

That order is backwards, and it explains a lot of what has gone wrong across the industry. McKinsey's latest State of AI research found that while the large majority of organizations now use AI somewhere, only a small minority can point to a measurable effect on earnings.

So here is the method we use with clients before anyone writes code. It takes about a week, and it kills roughly one project in four. That is the point.

Why most enterprise AI ROI cases fall apart

Three failures show up again and again.

The first is a baseline nobody measured. If you cannot say what the process costs today, you cannot prove it got cheaper. Yet teams routinely skip this because measuring feels slower than building.

The second is counting saved hours as saved money. Freeing up 200 hours a month is only a return if those hours go somewhere valuable, or if headcount actually changes. Otherwise it is a productivity story, not a financial one.

The third is a cost side that only includes the build. It leaves out inference. It leaves out maintenance, and change management too. A business case missing those understates the real number badly.

Step one: measure the baseline

Pick one process. Then find out what it genuinely costs right now.

Count the hours people spend on it, at loaded salary rates. Add the errors, and what each one costs to fix. Then include the delays, the rework, and the revenue that leaks while somebody waits. Where a number is unavailable, ask three people who do the work and take the middle answer.

Write the figure down and get it agreed before the project starts. A baseline established afterwards is always disputed, and a disputed baseline destroys an otherwise sound enterprise AI ROI case.

Step two: find where the value actually sits

Returns come from four places, and they are not equally credible.

• Labor reallocation. Hours freed for higher-value work. Real, but only if you can name what those people will do instead.
• Error reduction. Fewer mistakes, less rework, fewer penalties. Usually the easiest bucket to evidence, since finance already tracks the cost of failures.
• Speed. Faster cycle time that converts into revenue. Strong when a quote turned around in one hour rather than two days wins measurably more deals.
• Capacity. Handling more volume without adding people. This one tends to convince CFOs, because it is visible in a headcount plan.

Weight these honestly. Error reduction and capacity survive scrutiny. Vague productivity claims rarely do.

Step three: put the entire cost in

Your enterprise AI ROI model needs the full picture, not the invoice.

Start with the build. Fixed-price enterprise projects typically run $40,000 to $150,000 depending on scope and integration count. Add inference, which is ongoing and volume-dependent, from a few hundred to a few thousand dollars monthly.

Then add maintenance at 15% to 20% of build cost each year. Add internal time, because your people will spend real hours on discovery, testing, and rollout. Finally, add change management, which is the line most often left out and most often decisive.

Step four: do the payback math

The calculation itself is simple once the inputs are honest.

Annual benefit divided by total first-year cost gives you payback. A well-scoped build usually pays back in 12 to 18 months. Anything promising three months deserves suspicion, and anything past 24 months will struggle to hold executive attention long enough to matter.

Run three scenarios rather than one. Take your realistic case, then model a version where benefits land 40% lower and costs run 20% higher. If the pessimistic case still pays back inside two years, you have a project worth funding.

A worked example

Numbers make this concrete, so here is one drawn from a fairly ordinary document workflow.

Say a claims team processes 4,000 documents a month. Each takes 20 minutes of manual review, which works out to about 1,333 hours a month, or 16,000 hours a year. At a loaded rate of $45 an hour, that process costs about $720,000 a year.

An AI system handles 70% of documents without review and flags the rest. Labor drops by roughly $500,000. Error-related rework falls by another $60,000.

Against that, the build costs $90,000, inference runs $18,000 a year, and maintenance adds $16,000. First-year cost lands near $124,000, with about $34,000 recurring afterwards.

Payback therefore arrives inside four months. Figures like these are illustrative, obviously, and yours will differ. But the structure is what you are borrowing, not the numbers.

Notice how much of the case rests on volume. Halve the document count and the same build takes eight months to pay back, which is still fine. Halve it again and the project stops making sense. That sensitivity is worth testing before you commit.

What CFOs actually push back on

Expect three questions.

"Where does the headcount go?" If your case rests on labor savings, be specific about whether roles change, shift, or disappear. Vagueness here sinks proposals.

"What if it does not work?" Answer with structure rather than confidence. Phased engagements with a stop point after each stage make this question much easier to handle.

"Who maintains it?" A system nobody owns degrades quietly. Name the owner in the business case itself.

When the honest answer is no

Sometimes the enterprise AI ROI simply is not there, and finding that out in week one is a win.

Low-volume processes rarely justify a build. Neither do workflows about to be replaced by an ERP migration, or processes so full of exceptions that automation would cover barely half the cases. Fixing the underlying data problem often returns more than any model would.

We talk clients out of projects for these reasons regularly. Our AI consulting work exists partly for that, and the 4-Month AI Framework walks through the same scoping questions if you want to run the exercise yourself first.

Review your business case with us

A strong enterprise AI ROI case is specific, conservative, and agreed with finance before the first sprint.

Book a 30-minute call with our founders and bring your numbers. You will leave with a timeline, an investment range, and an ROI estimate you can take to your CFO. No sales pitch.