How to Calculate ROI for an AI Project
A practical, defensible method for estimating the return on an AI initiative — and the assumptions that decide whether the number holds.
Start with the decision, not the model
ROI is a decision-support number, not a forecast. Its purpose is to tell you whether an initiative is worth committing scarce capital, engineering time and organizational attention. Treat it as a structured argument you can defend in a room, not a single figure you present as certainty.
The core equation
At its simplest, ROI over a period is (total realized benefit − total cost) ÷ total cost. For AI initiatives the difficulty is never the arithmetic — it is defining each term honestly. Costs are usually understated because recurring model, infrastructure and oversight costs are missed. Benefits are usually overstated because theoretical productivity is treated as realized financial value.
Separate cost into one-time and recurring
One-time costs include software or development, integration, data preparation, infrastructure setup, security and governance work, and change management. Recurring costs include model or API consumption, cloud infrastructure, licenses, monitoring, support and ongoing governance. A three-year view multiplies recurring costs and often changes the conclusion.
Model benefit in four categories
Productivity (time reclaimed, valued at fully loaded cost), direct cost reduction, revenue contribution (margin-adjusted, never gross revenue), and risk or loss avoidance. Keeping these separate makes the business case auditable and shows reviewers exactly where value is assumed to come from.
Discount productivity for adoption and realization
Theoretical productivity assumes every intended user adopts the tool and converts every reclaimed hour into measurable value. Neither is true. Apply an adoption rate (share of users who meaningfully use it) and a realization rate (share of theoretical benefit that becomes real business value). This single step is what separates a credible business case from a spreadsheet fantasy.
Adjust the first year for ramp-up
Value does not begin on day one. Account for an implementation period with no benefit and a ramp-up period where benefit scales toward steady state. Year-one ROI is almost always the constraint that determines whether a project gets funded.
Present a range, not a point
Because the result is assumption-sensitive, show conservative, expected and optimistic scenarios. The honest message is not 'ROI is 184%' — it is 'ROI is attractive under a broad range of adoption and realization assumptions.'
Put this into practice.
Run the numbers for your own initiative in the free AI ROI calculator.