How to evaluate the return on AI investments: questions before the decision
ROI of AI projects is measured differently than ROI of traditional automation. I break down the right questions to ask before money is spent.
When a business owner asks me how to evaluate the return on an AI project, I usually start with a different question: "What exactly do you want to change?" Because AI is not equipment with a known throughput and it is not a licence with a fixed price. It is a capability that changes how people and systems perform tasks. That is harder to measure.
Difficulty does not mean measurement is unnecessary. It means you have to measure correctly.
Why the standard ROI approach does not work
The classic formula: count costs, count benefits, take the ratio. The problem is that in AI projects, benefits are often non-linear, delayed, and dependent on how the system is actually used.
The typical mistake is counting savings as "AI does what a person used to do, multiply by the hourly cost". Sometimes this is right. But more often AI changes the task itself - it makes practical what was previously impractical at scale, or it shifts the quality of a decision where speed was not the bottleneck.
Example: an AI assistant in customer support. It does not replace agents one-to-one. It changes what agents do, allows handling higher volumes without proportional headcount growth, and improves the quality of answers to routine questions. Counting ROI as "one bot equals one agent" is counting nothing.
What to look at when evaluating
I group effects into three categories:
Direct savings - tasks that used to require resources are now automated. This is measurable, though usually smaller than expected.
Scaling without proportional cost growth - volume of tasks increases, headcount does not. This is a non-linear effect, harder to calculate, but often larger than the first category.
Improvement in decision quality - fewer errors, faster responses, better anomaly detection. This is the hardest to measure but sometimes the most important for the business.
What you need to define before launch
Before starting a project, three things are required: a baseline, a success metric, and a measurement horizon.
Baseline - what the process looks like today: how much time, how many people, what quality of output. Without this, it is impossible to measure change.
Success metric - a specific indicator that should change. "Things got better" is not a metric. "Response time dropped from 24 hours to 4" is a metric.
Measurement horizon - when to expect results. AI projects often show modest results in the first months and meaningful results after six months to a year, once the system has been run in and the team has adapted.
A practical filter
Before signing off on an AI project budget, ask yourself five questions:
- What specifically will change about the work, and how will we measure it?
- What is the baseline cost of the current process?
- What result makes the project a success, and what makes it a failure?
- Who inside the company owns the success of this project?
- If nothing changes in a year, what exactly went wrong?
Clear answers to these questions are worth more than any ROI estimate written at the pitch stage.