Most teams do not struggle to find AI ideas. They struggle to pick the one worth doing first. The list is long, the demos are convincing, and everything sounds promising in a slide. The hard part is deciding where to spend the first dollar.
A useful starting question is not "where could we use AI" but "where is the work slow, costly, or error-prone today". AI earns its place when it removes real friction, not when it adds a feature nobody asked for.
Start with the problem, not the model
Pick a process you already run. Watch how the work actually flows, including the manual steps and the handoffs between systems. The places where people copy data, re-key information, or wait on someone else are usually the best candidates.
This is an engineering habit more than an AI one. You are looking for the bottleneck, then asking whether AI or automation is the right tool to remove it. Sometimes it is. Often the answer is a simpler integration or a workflow change.
Weigh return against feasibility
Once you have a few candidates, rank them on two axes:
- Return: how much time, cost, or error does it remove, and how often does the work happen.
- Feasibility: do you have the data it needs, can your systems support it, and how clear are the rules.
The best first project is high on both. A high-return idea that depends on data you do not have is not a first project; it is a later one, after you fix the data.
Be honest about what to skip
A good plan says no as often as it says yes. If a use case is exciting but low-return, or feasible but not worth much, it goes to the bottom of the list. Saying so early saves budget and credibility.
What you should walk away with
The output of this work is not a deck. It is a short, ranked list of use cases, with a clear note of what each one needs and what it returns. That is enough to start building the right thing first, instead of the loudest idea in the room.