Many companies approach AI through technology: "we want to do LLM" or "we want an agent." The right question is different: which problem, once solved, will have a visible impact within a few weeks? Here's the grid we use to select a first use case.
Criterion 1 · Frequency
A process run a hundred times a week offers far more opportunities for fast learning and measurable gains than a process run once a month, even if the latter seems more "strategic."
Criterion 2 · Data availability
The best use case in the world is worthless if the data it needs is scattered, incomplete, or inaccessible. Checking real data availability before picking a use case saves weeks of being stuck.
Criterion 3 · Tolerance for error
A first project should be able to get things wrong without serious consequences. Use cases with high regulatory or financial stakes are better candidates for a second or third iteration, once the method has proven itself.
The point of the first project isn't to impress: it's to prove, with numbers, that the approach works in your context.
Criterion 4 · An engaged business owner
An AI project with no business-side champion ready to test, correct, and defend the system to their team is statistically less likely to succeed, regardless of the technical quality of the solution.
How to use it
List 3 to 5 use-case ideas, score each one against these 4 criteria, and start with the one that gets the best overall score, not necessarily the one that seems the most ambitious.
