The question we get asked is usually some version of where should we start with AI. The honest answer is that the starting point matters less than the filter you apply, because the difference between a pilot that reaches production and one that quietly stops is rarely the model.
Three ways a use case dies
- The economics never worked. The task was cheap to do manually and the system costs more per run than the person it replaces.
- The data was not what everyone assumed. It exists, but it is inconsistent, undocumented, or locked in a system nobody wants to touch.
- The workflow was never going to change. The output is good and no one uses it, because using it means someone doing their job differently and nobody agreed to that.
Note that none of these are about whether the technology can do the task. That question is usually the easiest to answer and gets the most attention.
A filter that works
Score each candidate on three axes before writing any code. Value: how much time or money does this actually free, in numbers you would defend to a CFO. Data readiness: could you assemble a decent evaluation set this month, without a project to do it. Risk: what happens when it is wrong, and who notices.
Anything that scores badly on data readiness is not an AI project yet. It is a data project with an AI project behind it, and sequencing those correctly saves a great deal of money.
Start where being wrong is survivable
The best first use cases tend to have a human already in the loop, a tolerance for imperfect output, and a clear measure of whether it helped. Drafting rather than deciding. Summarising rather than acting. Suggesting rather than sending.
This is not timidity. It is that you learn the shape of the problem, your data and your team's appetite far faster from something in production than from a pilot in a sandbox.
Decide what you will measure first
Before the build, agree the number. Minutes saved per case, share of drafts accepted without edit, reduction in time to first response. Write it down and instrument it on day one. A system without that number cannot be defended when someone asks what it is costing, and that question always comes.