I've watched this pattern play out at more than one company now. Leadership decides it's time to "do something with AI." Someone puts together a shortlist of use cases based on what they think employees are doing all day. A team starts building. A few months later the tool works fine, technically, and almost nobody touches it, because it was never solving the problem people actually had.
That gap has nothing to do with the technology. Nobody checked what was actually happening before deciding what to build.
Here's the framework I use to find out where AI genuinely creates value before any money gets spent building anything. Five steps, in order, no shortcuts.
Step 1: Talk to the people doing the work, not just the people describing it
Most AI initiatives run on leadership's mental model of the business, and that model is almost always wrong in specific, fixable ways, because leadership sits several steps removed from the actual day-to-day work.
The fix is simple, if unglamorous: interview people at every level, not just department heads. Talk to whoever is actually doing the process, every day, right now. This is where most misguided AI initiatives go wrong from the start, and it's also the cheapest mistake to catch, if you catch it early enough.
Step 2: Map the process by watching it, not just asking about it
Interviews get you most of the way there, but self-reported process is often a little flattering. People describe their job the way it's supposed to work, not the way it actually works once they're behind on a deadline.
So map the real workflow. Where does the friction actually live? What are people quietly working around? Session recording tools help here, since they show what people are literally clicking through in their browser instead of what they remember doing later. You want the real map, not the org chart's polished version of it.
Step 3: Build the use case list, and prioritize by impact
Once you actually understand what's happening, you'll usually end up with more use cases than you can reasonably act on. The mistake most companies make at this point is prioritizing whatever's most visible or most annoying, instead of whatever actually moves the business.
Here's a concrete version of that trap. Say an employee making $50,000 a year spends nearly their whole day on one repetitive task. It's tedious, it's visible, everyone agrees it should be automated. But if there's a separate opportunity somewhere else in the business worth $500,000 in new revenue, or $100,000 in cost savings, that one needs to go first, even though nobody's complaining about it out loud.
Plot everything on a real matrix: revenue impact against time savings. Not gut feel about what looks the most automatable.
Step 4: Sort everything into three buckets
Not every use case is the same kind of problem, and treating them all the same is where a lot of AI budgets get wasted. I sort into three:
Quick AI wins. Cases where AI is genuinely the right tool and it's fast to stand up. Sometimes this is nothing more than a well-built AI skill or workflow, not a whole platform.
Process fixes. This one surprises people every time. A good chunk of what looks like an AI problem isn't one at all. Here's a real example, generalized: a team was struggling to get pricing and historical spend data from suppliers for a set of municipal accounts. The instinct was to build something smarter to work around the low response rate. The actual fix had nothing to do with AI. Outreach was going out from a consulting firm's email address instead of the client's own municipal one, and suppliers respond very differently to someone they expect an ongoing relationship with than to a consultant who'll be gone in six months. No model required, just the right read on what was actually going on. It's a small thing, but it's the kind of small thing that gets missed constantly.
Big bets. The stuff with the biggest revenue or bottom-line impact, usually needing more investment, more change management, more time. You don't build these first. But they're the whole reason the assessment matters in the first place, because without it, big bets get missed in favor of whatever's loudest in the room.
Step 5: Turn it into a roadmap
The deliverable here shouldn't be a slide deck that gets presented once and forgotten. It should be a sequenced roadmap: what gets built first, second, later, based on the real impact-versus-effort read from steps three and four. That's the point where a company can actually make decisions about where to spend money, instead of guessing based on whoever's loudest internally.
Why this matters more when there's no room to get it wrong twice
A large enterprise can usually afford to miss on an AI project and try again. Most companies can't. When the budget for a project is a real chunk of what's available, and there's no second attempt built into the plan, getting this right the first time stops being a nice idea and becomes the whole game. It's the difference between a project that gets funded again next year and one that quietly disappears.
I'm working on a simple scorecard version of this that a leader could run themselves against a single business unit or opportunity. I'll add it here once it's ready.