To prove the AI paid off, measure the baseline.

“So did the Artificial Intelligence (AI) pay off?” Every team that talks a client into an AI project gets the question eventually. The answer that holds up is a number you can defend: the work is faster, cheaper, or better by this much, measured against what came before, after you subtract what it cost to run. A vendor’s 700% case study doesn’t give you that.

Why is AI return so hard to prove?

The loud version of return on investment (ROI) is a slide with one enormous number on it. The measured version is quieter, and more useful. When Stanford’s AI Index looked at what companies report from AI, the gains were real and modest: firms that see savings mostly report cost savings of less than 10%¹, and where revenue rises, the most common level of revenue increases is less than 5%¹. A confident case study is easy to produce and easy to believe, the same way a fluent AI answer can be wrong².

How do you measure it honestly?

What's Real?

• •The gains are real, and usually smaller than the pitch. In Stanford’s AI Index, firms that report savings mostly report cost savings of less than 10%¹, and where revenue rises, the most common level of revenue increases is less than 5%¹.
• •Controlled tests show where the lift is real. In a Boston Consulting Group field experiment, consultants using AI on well-suited tasks finished 12.2% more tasks on average, completed tasks 25.1% more quickly³, and produced 40% higher quality results³, measured against a group who worked without it.
• •The same test shows where it backfires. On a problem built to sit outside the model’s competence, the consultants who leaned on AI did worse³ than those who did not, talked into a confident wrong answer. The tool helps on the right tasks and hurts on the wrong ones, and only measurement tells them apart.
• •Proving it is a named discipline. The United States National Institute of Standards and Technology (NIST) tells teams to run quantitative, qualitative, or mixed-method⁴ measurement, and that AI systems should be tested before their deployment and regularly while in operation⁴. Proof is measurement done on purpose, and on a schedule.

The “so what” — for anyone serving tech clients

• •Baseline on day zero, before you build. The most valuable move on an AI project costs nothing and happens at the start: write down what the task costs and how long it takes today. Skip it and all you can do is claim the lift.
• •Price the whole thing. The subscription is the visible cost. Integration, the human review step, and the reruns when the model is wrong are the rest, the same gap the demo-to-production brief⁵ named between a slick demo and a shipped product.
• •Prove it on their work. Whether the tool clears the bar on a client’s real cases is an evals⁶ question, and it is answered on their own examples.
• •Sell the measurement as part of the job. Offer a metric the client already watches and a date to read it. That is the line between a partner and a vendor with a slide.