How to measure ROI of generative AI: start with the problem, not the tool

“What can we build with AI?” is the wrong question

The numbers show how common this mistake is. MIT’s NANDA initiative found that roughly 95% of enterprise generative AI pilots delivered no measurable return, despite $30 to $40 billion in corporate spending. IBM’s 2025 study of 2,000 CEOs found that only 25% of AI initiatives delivered the ROI they expected, and 64% of CEOs admitted that fear of falling behind drives them to invest before they clearly understand the value.

That last number explains the first two. When the tool comes first, teams look for somewhere to plug it in, and they usually pick whatever process is closest. If that process is broken, AI does not fix it. It just makes a broken process run faster. The result is activity without impact, and a generative AI ROI that nobody can prove because nobody defined what it was supposed to improve.

The better approach flips the order. Start with the business issue, find the best way to solve it, and only then ask whether AI is part of the answer. That shift comes down to three questions.

Question 1: What business problem are we solving?

Every credible AI business case starts with a problem that is specific, measurable, actionable, relevant, and time-bound. “Use AI in customer service” is not a problem. “Quote turnaround takes four days and we lose one in five deals while customers wait” is.

Once the problem is defined, measure its current state: how big it is, what it costs, how urgent it is, and what workarounds people already use. That measurement is your ROI baseline. Without it, any result you report later is an opinion. Expect the definition to sharpen as you learn more. That is a feature of good problem-solving, not a sign that the first version was wrong.

Question 2: Who lives with the problem?

Problems do not exist in spreadsheets. They exist in people’s workdays. Two groups matter most. Leaders and decision-makers set the priorities and know the business context that never shows up in the data. The people who deal with the problem every day decide whether any solution actually gets used.

Those two groups are often not the same, and that gap is where many projects fail. The person who requests an AI tool is rarely the person who will use it every day. If the solution ignores the daily user’s workflow, comfort with technology, and incentives, adoption stalls, and the projected ROI never becomes real. Measuring AI productivity gains means measuring whether the right people changed how they work.

Question 3: What is the best solution, and is AI part of it?

Only now does technology enter the conversation. Sometimes the best answer is fixing the process itself: removing a handoff, clarifying ownership, or standardizing an input. Sometimes it is simple AI workflow automation. Sometimes generative AI is exactly the right fit. Starting with the problem lets you choose honestly instead of forcing a tool into a gap it was never meant to fill.

This is also where ROI measurement becomes straightforward. Measure the solution against the baseline from Question 1, not against tool metrics. Prompts per week and licenses activated tell you people opened the software. Shorter quote turnaround, fewer errors, and more deals closed tell you the business got better.

Why this matters

Measuring the ROI of generative AI is not a reporting exercise at the end of a project. It is a decision made at the beginning. The companies that can prove their return tend to do three things:

• Define the business problem and its baseline before choosing a tool.
• Design around the people who will use the solution every day.
• Measure outcomes against the original problem, not against usage.