How to find the AI use cases that actually pay off

The reason most companies can't point to a return on AI isn't the technology; it's which work they pointed it at. A widely cited MIT study found 95% of enterprise generative-AI efforts produced no measurable impact on the P&L, and the most useful detail in it is where the money actually showed up. The returns concentrated in the back office, automating document review, support, and financial monitoring, while most budgets went to sales and marketing, where the demos look best. The popular AI use cases and the paying AI use cases were not the same list. That gap is a selection problem, not a capability one, and it's the one to solve first. Picking the right use case is the difference, and once you've picked it, measuring the AI ROI is the next problem, not this one.

Most operators are at the start of this, not scaling it: US Census data put business AI adoption at roughly 17 to 20% through early 2026. So the practical question for almost everyone is the same, where do you point AI first. Here's a way to answer it that survives a finance review.

Don't start with what AI can do; start with what costs you. The fastest way to a paying use case is to begin from line-of-business pain, not from a feature list. Ask each team three questions: what are your top priorities this year, where is the biggest gap between current and target performance, and which processes create the most friction for staff or customers. That surfaces a wide pool of candidates rooted in real cost, which is exactly the instinct you'd apply to any other part of the business. Cast the net wide here, because the best AI use cases are often unglamorous back-office work nobody would put in a pitch deck, and narrow in the next step.

Once you have candidates that pass the filter, score each on two axes and plot them. Impact is the dollars of cost cut or revenue moved, tied to the baseline KPI you named. Feasibility is data readiness, how cleanly it fits the real workflow, and whether the team will actually adopt it, not whether the model is capable, because model capability is rarely the blocker. High-impact and high-feasibility use cases get done now; the rest wait or get cut. Then sequence deliberately: pick a 60-to-90-day win that proves value and earns budget before the multi-quarter bets. Speed builds the executive confidence that a pile of scattered, never-ending pilots destroys. What counts as high-volume, repetitive work also differs by role, so score within each function rather than across the whole company.

There's one more trait the winners share, and it's the one most selection processes skip: they redesign the workflow instead of layering AI on top of it. McKinsey's data points to workflow redesign as the change most associated with real EBIT impact, yet most teams just bolt AI onto a process they never changed, which is a big part of why so little of it lands. So selecting a use case isn't only choosing which work; it's committing to rebuild how that work flows around the tool. That's the point where use-case selection turns into AI adoption, and where the value stops being a slide and starts being a number.

A fair objection says you can't predict AI ROI in advance, so don't try to pick winners; hand the tools to everyone and let value emerge from broad experimentation, because rigid filters kill the ideas you'd never have guessed. It's a credible position, and the shadow-AI numbers back part of it: people already use AI on their own work far ahead of any official program, surfacing uses no committee would have scoped. The honest answer is that both are true, and they do different jobs. Use broad experimentation as the discovery mechanism, the way you generate candidates in the first step. But you still need selection to decide what gets operationalized and funded, because the bottom-up failure mode, isolated projects with no coordination, is exactly the scattered-pilot pattern that produces the 95% no-impact rate. Experiment widely to find use cases; select ruthlessly about what you scale. Predicting ROI to the decimal is a fool's errand. Choosing high-volume, measurable, owned work over a flashy demo is not, and it's the choice the data keeps rewarding, as part of a broader AI transformation.