Ask a room of executives whether their AI investments are paying off and most will say yes. Ask them to show the number and the room goes quiet. The dirty secret of enterprise AI is that an enormous amount of reported ROI is theatre — plausible-sounding figures that wouldn't survive a serious finance review.

This matters because the credibility gap is starting to bite. After an initial wave of enthusiasm-funded pilots, budget holders increasingly want to see returns that look like real returns. Teams that can't produce them are watching their AI budgets get questioned.

Why 'time saved' is the weakest number

The most common AI ROI claim is some version of this tool saves each person two hours a week. It's the weakest possible evidence, for two reasons.

First, it's almost always self-reported and estimated, which makes it trivially inflatable. People are bad at estimating their own time and motivated to justify a tool they like.

Second, and more importantly, saved time is not saved money unless something downstream actually changes. If you save 200 people two hours a week and then everyone fills those hours with other work, you have not reduced cost — you've increased slack. Time saved only converts to cash if it converts to capacity used (more output from the same people) or headcount changed (fewer people for the same output). Most ROI decks quietly skip this conversion step, which is where the number evaporates.

What a credible measurement needs

A defensible AI ROI claim has three components that the theatre version lacks.

A baseline. You cannot claim improvement without knowing the before. This sounds obvious and is routinely skipped, because measuring the pre-AI state is boring and nobody thinks to do it until after they've deployed. Capture the baseline first or you're guessing forever.

A control. Ideally, some comparable group or time period without the AI intervention, so you can separate the tool's effect from everything else that changed. A clean A/B isn't always possible, but even a rough before-and-after with a plausible counterfactual beats a single number floating in space.

A line to something finance already counts. The outcome has to connect to a metric the business cared about before AI showed up: revenue, cost, risk exposure, cycle time, error rate, customer retention. "AI usage went up" is an activity metric, not an outcome. The translation from activity to outcome is the whole job.

The returns are often in the wrong column

Here's the part teams get backwards. The headline pitch for AI is usually labour reduction — do the same work with fewer people. But in practice, some of the largest and most durable returns show up as quality and capacity, not headcount.

An AI system that catches more errors, handles a surge without a hiring spike, lets a junior person produce senior-level output, or compresses a two-week process into two days is creating real value — but you won't see it in a headcount line. You'll see it in defect rates, in the ability to take on more volume without scaling cost linearly, in faster cycle times that win deals. These need different metrics, and they're often more defensible than headcount claims precisely because they're harder to fake and tied to outcomes the business already tracks.

The honest version

A CFO-proof AI ROI story looks unglamorous: here's the baseline, here's the cohort with the tool versus without, here's the specific business metric that moved, here's the dollar value of that movement net of what the tooling and the people maintaining it cost. It's slower to produce and the numbers are usually smaller than the breathless pilot decks. But they're real, they survive scrutiny, and they're the only kind of number that keeps a budget alive into year two. The teams that learn to measure honestly will still have AI programs when the theatre-only teams have had theirs cancelled.