How to measure the ROI of an AI project

The ROI of an AI project is measured by comparing the business value the system generates once in production with its total cost of building and operating it, against a single metric agreed before you start. The hard part is not the calculation: it is fixing in advance what you will measure and having a real production system that produces it.

The metric is agreed before you start

Without a success criterion agreed in advance, any result looks justifiable and none is provable. That is why the first decision of an AI project is not technical but business: which number has to move for this to be worth it? Cycle time, volume processed, error rate, cost per case. A single metric, clear and measurable. If it cannot be defined, the project should not start: it is the most reliable signal that the use case is not yet mature.

The value is in production, not in the demo

A pilot that never crosses into the operation has zero ROI, no matter how brilliant the demo. Value only appears when the system works over real data, day after day —that is, in AI in production. Measuring the return of a prototype that nobody uses is measuring an intention, not a result. That is why the ROI conversation and the taking-the-system-to-production conversation are the same conversation.

Count the full cost, not just building

The most common mistake is comparing the value generated only against the cost of development. The real cost includes operating the system: compute, maintenance, human oversight and the corrections every living system needs. An agent that saves a hundred hours a month but demands another eighty in review has a very different ROI from what it appears to have. Measuring well means adding the cost of continuous operation, not just the initial invoice.

Honest attribution: was it the AI?

For the number to mean anything, you have to be able to attribute the change to the system and not to noise. Here AI traceability helps: knowing which cases went through the system, what it decided and with what result lets you compare against a real baseline. Without that record, ROI is a story you tell; with it, it is a figure you show. Measurement thus connects with AI governance: what is measured, who is accountable and how it is reviewed.

Measure over time, not in a snapshot

The ROI of AI is not a snapshot of the first month. Systems improve with tuning and degrade if nobody maintains them; the data changes and so does the context. A good measurement framework reviews the metric periodically and uses it to decide: scale what works, fix what does not and retire what adds nothing.

How we approach it at Codara

Our way of working starts from a business metric agreed per project —if it cannot be defined, we do not sign— and measures the result against that criterion in production, with the traceability needed to attribute the value to the system and not to chance.