AI ROI: The Incomplete Math

It's not fraud. It's selective accounting.

AI savings are measured, documented, and presented in a polished deck. The costs it generates stay off the same spreadsheet.

A few examples of what never shows up in the math:

- Time spent maintaining prompts every time the model is updated
- Hours reviewing exceptions the agent simply doesn't know how to handle
- Manual curation of critical outputs that "should" be automatic
- Rework caused by hallucinations at decision points
- Debugging integrations that break without warning

All of that has a cost. It takes time. It ties up people.

And it rarely makes it into the project's ROI.

Brazil is adopting AI at the speed of enthusiasm. According to ITDBr, the adoption rate jumped from 20% to 51% in a single year. At this pace, more and more projects get approved with inflated ROI — not out of bad faith, but due to structural omission: the numerator is presented clearly, the denominator is left out.

The point here isn't to question whether AI creates value — it does. The problem is that the math used to approve the project is rarely the same math used to sustain it. When the gap between the promised and actual ROI surfaces (and it does), the conclusion that circulates is "AI didn't work." The more accurate conclusion: the accounting was incomplete from the start.

Before approving the next project, it's worth adding these lines to the spreadsheet:

- Prompt maintenance cost (monthly)
- Exception curation hours per process
- Human review time for critical outputs
- Internal capacity for debugging integrations
- System adjustment or retraining cycle

If the ROI still holds with these lines included, the project is real. If it doesn't, better to find that out before the approval than after the deploy.

Tell me in the comments: which AI operational cost has never appeared on your project's ROI spreadsheet?