KPIs and ROI of GenAI projects: how to measure what really matters

“AI saves us time.” That phrase, with no number behind it, convinces no investment committee. Calculating the ROI of GenAI projects is one of the most difficult and most necessary conversations I have with management. Without clear metrics, a generative AI project stays an eternal pilot and ends up cancelled for lack of evidence of value.

Why measuring GenAI is different

Generative AI produces value in ways that are hard to capture with traditional metrics. It does not only save hours: it improves quality, reduces errors, speeds up decisions and frees up talent for higher-value tasks. The challenge is to translate those diffuse benefits into indicators that management understands and believes.

On top of this, the costs of GenAI are not obvious either. Consumption-based usage, process redesign and governance add expenses that many ROI calculations ignore, inflating the apparent return.

KPIs that really matter

Efficiency metrics

Cycle time, volume processed per person or reduction of manual tasks. They are the easiest to measure and the first ones I ask for, but on their own they tell half the story.

Quality metrics

Error rate, rework avoided or customer satisfaction. This is where GenAI usually shines, and it is worth measuring well because it often justifies the investment better than pure time saving. It connects directly with the way I advocate for aligning strategy and delivery with data.

Adoption metrics

A brilliant system that no one uses has zero ROI. Measuring real usage, frequency and drop-off is essential to know whether the theoretical value materialises.

How I calculate the ROI of GenAI projects

My method is simple to state: quantify the total benefit (efficiency plus quality plus new revenue) and subtract the real total cost, including the hidden expenses of FinOps and cost governance. The important thing is to be honest with both sides of the equation; an inflated ROI backfires when it is not met.

Common mistakes when measuring

The three I see most: counting only the hours saved, forgetting the cost of governance, and measuring in the pilot but not in production. Avoiding them is the difference between a solid business case and a promise no one will be able to verify.

From the KPI to the GenAI dashboard

An isolated metric tells you little. What really helps management is a dashboard that combines efficiency, quality and adoption in a single view, and that evolves with the project. At the start you want to measure whether the tool is used and liked; later, whether it reduces costs and errors; and in the mature phase, whether it generates revenue or opens new lines of business. Adapting the indicators to the stage avoids the classic mistake of demanding financial return from a pilot that is still validating basic usefulness.

When I build that dashboard, I try to make sure every metric has an owner and an associated decision. A number that no one looks at and that changes no decision is noise. That is why I prefer a few well-chosen indicators to a huge panel that gives a sense of control but that no one uses to act.

ROI and hidden costs: two sides of the same coin

You cannot calculate the ROI of a GenAI project well without understanding its real costs, and that is where many companies go wrong. The price per token or the model licence is only the tip of the iceberg: there is the cost of integration, of governance, of human review and of maintaining the context. If you count only the provider’s invoice, your ROI will come out artificially pretty and reality will catch up with you within a few months. That is why I recommend approaching measurement together with the budget, something I explain in detail in my article on how to budget an AI project and its hidden costs.

My way of calculating the return is deliberately conservative. I prefer to underestimate the benefit and overestimate the cost, because a project that comes out profitable even with prudent figures is a solid project. When the numbers only add up with the most optimistic assumptions, it is usually a sign that it is worth reviewing the use case before scaling.

Fundamentally, measuring GenAI well is an exercise in honesty. The indicators are not there to justify what we already decided, but to discover early what works and what does not. That culture of sincere measurement is what separates the companies that scale AI with judgement from those that pile up pilots that never take off.

A mistake I see repeated when measuring the return on AI is confusing activity with value. A team using a tool a lot does not mean it is generating benefit; the relevant figure is the impact on the business indicator that mattered before you started. That is why I insist on setting the baseline and the concrete target from day one: how much time is saved, how many errors are avoided or how much additional revenue is attributed. Without that initial reference, any ROI calculation ends up being an optimistic narrative impossible to defend to management.

Conclusion: without measurement there is no scale

No generative AI project scales without demonstrating its value. Defining honest KPIs and calculating a credible ROI of GenAI projects is what separates the pilots that die from those that become stable capabilities of the organisation.

Frequently asked questions about KPIs and ROI of GenAI projects