When a company asks us to assess where AI fits in their business, they usually expect a technology evaluation. Which model, which vendor, which platform. That's the last ten percent of the work. The first ninety is figuring out whether the problem they want to solve is real, frequent, and actually solvable with the data they have.
Here's what we look for, in the order we look for it.
question one: where does repetition live?
We interview the people doing the work — not their managers. We're listening for the tasks people describe with a sigh: the weekly report assembled by hand, the same six questions answered forty times, the copy-paste ritual between two systems that don't talk. Repetition with judgment attached is the AI sweet spot. Repetition without judgment is just automation — cheaper to solve with a script.
question two: what does the data actually look like?
Every opportunity dies or thrives here. The org chart says "we have all our support tickets since 2019." Reality says they're split across three tools, half are tagged wrong, and the export is a PDF. We audit what exists, what's accessible, and what's trustworthy — because a model grounded in bad data doesn't fail loudly. It fails plausibly, which is worse.
"Don't build it" is a deliverable. It's the one that saves the most money, and the one nobody else seems willing to write.
question three: what happens when it's wrong?
Every AI system is sometimes wrong. The question is whether the workflow can absorb that. Drafting an email a human reviews? Wrong is cheap. Quoting a price to a customer? Wrong is expensive and possibly contractual. We map each opportunity onto a simple grid — frequency of use against cost of error — and the upper-left corner (used constantly, mistakes cheap) is where the first project lives. Almost without exception.
question four: who owns it in six months?
An AI tool without an internal owner is abandonware with an API bill. Before recommending any build, we name the person who'll own prompts, monitor quality, and handle the "it said something weird" Slack messages. If that person doesn't exist and can't be hired, the recommendation changes — usually toward a smaller, more self-service tool, sometimes toward nothing at all.
what comes out the other end.
The deliverable is a ranked opportunity map, an honest "won't help" list, and a 90-day roadmap for the top one or two items — scoped tightly enough that a team can execute it with or without us. Some clients build with us. Some take the report and run. Both outcomes are the assessment working as designed.
The pattern across every assessment we've run: companies overestimate what AI can do for their hardest problem, and underestimate what it can do for their most boring one. The boring one is where to start.