AI ROI is one of the most contested topics in enterprise technology. Vendors overclaim it. Sceptics undercalculate it. Internal teams often measure the wrong things. Here's how to approach it with intellectual honesty.
Start With a Counterfactual
ROI is always relative to an alternative. "The AI saves us 10 hours a week" only means something if you're comparing it to a specific alternative — another tool, a manual process, hiring a person. Define your counterfactual before you start, not after.
Separate Hard Savings from Soft Benefits
Hard savings are directly measurable and attributable: reduced headcount, lower error rates, faster processing times with measurable cost consequences. Soft benefits — improved morale, better decision quality, reduced cognitive load — are real but harder to put a number on.
Don't ignore soft benefits, but don't use them to paper over a weak hard-savings case either. A good AI project should clear the bar on hard savings alone.
Account for Total Cost of Ownership
The biggest ROI measurement mistake we see is counting only the direct cost of the AI tool and ignoring the cost to build, integrate, maintain, and govern it. A $500/month API subscription might require $200,000 in engineering work to integrate properly. Include everything.
Measure Baseline First
You cannot claim savings you didn't measure before you started. Before every AI project, document your current-state metrics with rigour: how long does the process take, how many errors occur, what does it cost. This baseline becomes your before-and-after comparison.
