AI project ROI: the metric that looks healthy until the operation breaks
ROI is gain minus cost, divided by cost. The formula is trivial. The three lies are not in the math; they are in the inputs that feed it.
The number shows up on an approval slide, in green, with a decimal place that fakes precision. Almost nobody asks where the gain and the cost came from.
What it measures well
When gain and cost are measured honestly, ROI answers the right question: the project paid back. It is the final judge of any investment, and it should be.
The problem is never the formula. It is the quality of the two inputs.
What it hides
Three things, and each one topples the number.
The first is gain attribution. Shelf-out fell, and the project takes full credit, ignoring that a supplier changed and seasonality helped. Attributing to the model what the whole operation moved is the most common way to inflate ROI.
The second is the best case. ROI is computed on the pilot’s peak, not on sustained performance six months later, when the model has aged and nobody recalibrated.
The third is the hidden cost. The math sums the license and the project, and forgets the permanent cost of keeping it running: the team, the recalibration, the integration that breaks when the legacy system changes.
The pair that is missing
ROI declared once lies by freezing. The pair that makes it honest is a sustained operational indicator, measured every week, not a snapshot of the best month.
Read together, they separate the project that paid back from the project that looked like it did. High ROI with a stable sustained indicator is real return. High ROI with an indicator that decays after the pilot is victory declared too early, the kind most projects end up confessing.
The implementation note
The rule is to define gain attribution and total cost of ownership before declaring ROI, not after. And to measure the indicator in week one, compare in week twelve, keep it in the review. It is the same discipline as writing the criterion before the code.
Take the ROI that approved your last AI project. Who attributed the gain? Did the cost include maintenance? If the answers are loose, the number is healthy and blind.
Send me the ROI on the slide and how gain and cost were measured. In one hour I will send back a one-page audit: where attribution inflates, which cost was left out, and the sustained indicator that should sit alongside. If it confirms the gap, the two-week Diagnóstico scopes the fix.