There’s a reason this question is searched so much: most AI projects can’t answer it. MIT’s much-cited finding that 95% of enterprise AI pilots show no measurable return says less about AI than it does about measurement, because you can’t show a return on a baseline you never took.
Here’s the method that survives a CFO.
Measure hours, not magic
The honest unit of AI ROI in most businesses is the hour. Before anything gets built, baseline the task:
- Who does it, and what does their time cost, fully loaded?
- How long does it take, each time?
- How often does it happen?
- What else stalls while it’s happening? (Quotes that go out late, leads that wait, reports that slip.)
That’s your “before.” It takes an afternoon to gather, and it’s the single highest-leverage afternoon in the whole project, because everything afterward gets compared to it.
Then, after the system ships and settles in, measure the same task the same way. Hours returned, times loaded cost, minus what the system costs to run and maintain. That’s the number. Real examples from our client work: weekly social content went from six hours to 45 minutes; weekly ad reporting went from two or three hours to about five minutes. Nobody has to argue about numbers like that.
Define success before you build, or you’ll never prove it after
Every project we scope answers four questions up front, and the fourth is the one that makes ROI provable later: what does success look like? What does the system do, how does it interact with people, and what changes in the week when it works?
Write that down before building. When the definition of success is agreed in advance, ROI is a comparison, not a debate.
Count the second-order returns, but honestly
Hours are the floor, not the ceiling. Real systems also return:
- Speed: the quote that goes out same-day instead of Thursday closes more often
- Consistency: fewer dropped balls, fewer errors to fix twice
- Capacity: the expert doing expert work instead of admin, which is the whole point
Count these when you can attach evidence. Resist counting them when you can’t, because one inflated line item makes the whole business case smell.
The cheapest proof is a working prototype
If you’re being asked to justify an AI investment before it exists, projections are the weakest tool you have. Where the problem allows it, build a working prototype after discovery and measure that. A stopwatch on a real task beats a spreadsheet of assumptions, and greenlighting something you’ve already seen run is an easier decision to defend to anyone.
If you want help baselining your highest-cost workflow and building the case, book a free readiness call. Twenty minutes, no pitch, and if it’s not a fit you still leave with clarity.