Return on investment for AI initiatives is one of the most frequently cited metrics in AI business coverage, and one of the least consistently defined. Ask five companies reporting positive AI ROI how they actually calculated it, and the methodologies behind that single phrase turn out to vary enough that the numbers aren't really comparable to each other.
Where the definitions diverge
Some companies calculate AI ROI against the fully loaded cost of a deployment, including licensing, integration, and the ongoing human oversight a system still requires. Others calculate it against a narrower cost base that excludes the human oversight time, which can make a deployment look considerably more efficient than its true operating cost would support. Some measure the benefit side in hard, realized savings; others include projected or estimated savings that haven't yet been confirmed against actual outcomes.
Why this inconsistency persists
There's no standardized accounting framework specific to AI deployments the way there is for more established categories of capital expenditure, which leaves individual companies free to construct a methodology that reflects favorably on their own initiative. This isn't necessarily deceptive-many companies genuinely believe their own methodology is reasonable-but it does mean that comparing reported AI ROI figures across companies, or even across different initiatives within the same company, requires more scrutiny than the headline number alone provides.
Questions worth asking of any reported AI ROI figure
Does the cost side include the ongoing human oversight the system requires, or only the direct licensing and infrastructure cost? Is the benefit side based on realized, measured outcomes or projected ones? Over what time period is the calculation made, and does it account for the often-longer-than-expected ramp-up period before a deployment reaches its steady-state performance?
Toward more honest reporting
Business coverage would serve readers better by treating a reported AI ROI figure as a starting point for these questions rather than a directly comparable, standardized metric. Until a more consistent methodology becomes standard practice, the specific calculation behind any cited figure matters as much as the figure itself.