Most failed AI projects didn't fail in the build, they failed in the choosing. Teams pick the flashiest idea, not the one with the clearest payoff, and three months later they have a demo nobody uses.
Here's the framework we use in every discovery sprint to avoid exactly that.
Score every idea on two axes
Value: how much time or money does this save, or how much revenue does it protect? Be concrete, hours per week, cost per transaction, tickets per day.
Feasibility: is the data available and clean enough? Is the task well-defined? What's the risk if the model is wrong? A high-value idea with terrible data readiness is not your first project.
Start where the work is repetitive and text-heavy
LLMs shine on repetitive knowledge work: reading, drafting, classifying, extracting, summarizing. If a task is high-volume, rule-ish and text-based, it's a strong candidate.
Avoid, for a first project, anything that needs perfect accuracy with no human in the loop, or where being wrong is expensive and hard to detect.
Pick one, prove it, then scale
Your first pilot should be small enough to ship in weeks and important enough that success is obvious. Momentum from one visible win unlocks the budget and trust for the rest of the roadmap.
That's the whole game: choose well, prove it, compound.