AI & Data
How to measure the return on AI projects
Many AI pilots never produce financial results. A simple method to define, measure and decide before and after you invest.
In 30 seconds
- 01AI without a business metric defined before the pilot is a demo, not a project.
- 02Measure the process before changing it: time, cost and error per unit of work.
- 03Add up every cost, including human review and maintenance, and decide by an agreed date.
The most cited study on the subject in 2025, from MIT, concluded that the vast majority of corporate generative AI pilots produced no measurable financial return.1 The problem is rarely the technology. In most cases, nobody defined up front what success would look like.
Define the outcome before the tool
An AI project starts with a business sentence, not a model: “cut customer service response time”, “reduce errors in invoice checks”, “free up hours on the finance team”. If you cannot write that sentence, the project is not ready to start.
If the pilot has no number to beat, it can neither succeed nor fail.
Measure the process as it is today
Without a baseline, any result looks good. Before changing anything, record for a few weeks:
Add up every cost
The cost of an AI project is not just the license or the API. The math includes:
- development and integration with your current systems;
- model usage, which grows with volume;
- human review, which is almost always still needed;
- monitoring and tuning once the project is in production.
The return is the difference between the cost of the process before and after, minus these costs, over the agreed period. Simple to write, but it only works with the baseline from the previous step.
In practice · the four pilot questions
- Which business metric will this pilot move?
- What is that metric today?
- How much does it need to change for the project to pay off?
- By when do we decide to continue, adjust or stop?
Without all four answers, the pilot should not start.
Decide by a deadline
Pilots without a decision date tend to drag on. Agree up front on a short window, usually a few weeks of real use, and three possible outcomes: scale, adjust or stop. Stopping a pilot that missed its target is also a good result: the cost was small and the decision was made with data.