Most AI business cases fail at the finance meeting for the same reason. The return is a guess dressed up as a number, and the cost only includes the invoice.
A CFO does not reject AI because they distrust the technology. They reject a business case that would not survive the same scrutiny applied to any other capital request.
Short answer: real AI ROI compares a fully loaded cost, including build time, review time, and correction time, against a baseline captured before anything changed. Most business cases collapse because one side of that equation is missing, not because the technology underperformed.
Why the standard business case falls apart
The typical pitch estimates savings from a national average or a vendor case study, then compares it to a subscription fee. Both halves are wrong in the same direction. The savings figure is borrowed from a business that is not yours, and the cost figure ignores everything except the invoice.
Approve enough of these and the finance team stops trusting AI business cases specifically, which makes the next legitimate one harder to get through, regardless of how well it is built.
What the cost side actually includes
Four line items, not one.
License or usage cost. The number everyone already includes.
Build and integration time. Internal hours or contractor fees to connect the tool to your actual systems, which is almost always underestimated by a factor of two or three.
Review and correction time. Someone checks the output, at least at first, and that checking has a cost even when the system performs well.
Ownership time. Someone watches the system after launch. This does not disappear once it is live, and pretending it does is how projects quietly stop working without anyone noticing for months.
What the return side actually includes
Not a percentage borrowed from a case study. A number from your own operation, captured before you touch anything.
Time per unit of work, error rate, cost per transaction, or revenue per cycle, measured for two to four weeks before the project starts. This baseline is the single most commonly skipped step in the entire process, and its absence is why so many finished projects cannot answer whether they worked.
A worked example
| Line item | Weak business case | Fully loaded business case |
|---|---|---|
| Tool cost | $1,200 per month | $1,200 per month |
| Build and integration | Not included | 60 hours at internal rate |
| Review time, first 90 days | Not included | 4 hours per week |
| Ownership time, ongoing | Not included | 2 hours per week |
| Claimed savings | Industry average, 30 percent | Measured against a captured baseline |
The right-hand column is not more expensive to produce. It takes an afternoon of honest accounting instead of a vendor slide, and it is the version that survives a finance review.
Payback periods worth trusting
A payback period calculated from the weak column is a story. A payback period calculated from the fully loaded column is a forecast you can defend when someone asks about it in July.
Aim for a payback window under twelve months on a first project. Longer than that and you are asking leadership to fund conviction rather than evidence, which is a harder sell and a worse precedent.
What to do when the number is close
If the fully loaded case shows a marginal return, that is useful information rather than a reason to inflate the estimate. Either the workflow was the wrong first choice, or the scope needs to shrink until the ownership cost drops below the value it protects.
A project that clears the bar honestly is worth more to your credibility than one that clears it by omission.
FAQ
How precise does the baseline need to be?
Directionally accurate is enough. A baseline built from two weeks of real data beats a guess by a wide enough margin that further precision adds little.
Should we include the cost of doing nothing?
Yes, where it applies. A backlog that is growing or an error rate that is trending up has a cost even without an AI project attached to it.
What internal rate should we use for hours?
Fully loaded compensation, not just salary. Benefits and overhead are real costs even when they are not itemized on an invoice.
Does this apply to a small first pilot?
Especially there. A small pilot is exactly where teams skip the baseline, and it is the cheapest place to build the habit of capturing one.
What if the case looks bad once fully loaded?
Better to know in February than to explain it in November. Rescope and try again with a narrower workflow.
Where to go next: Before the next AI request goes to finance, run the fully loaded version of the numbers above. It is the difference between a business case and a hope with a subscription attached.
