Where to start when someone tells you to “do something with AI”

The mandate usually arrives with no detail attached. Someone above you has decided the company needs an AI story, and now it’s yours to produce. The instinct is to find the most impressive-sounding use case and staff it. Resist that. The flashy project is where quarters go to die.

Start with two questions instead. Both are about your own system, not about models.

First: can you reach the data? Pick the workflow you think AI might help with and ask, concretely, how you would get the data it needs out of the systems that hold it. If the answer involves a nightly export nobody owns, a database only one team can touch, or logs that aren’t searchable, you’ve found your real first project — and it isn’t an AI project.

Second: is there a task that is expensive, repetitive, and tolerant of being roughly right? Those three properties matter together. Expensive and repetitive is where value is. Tolerant of being roughly right is what makes it a good fit for a model instead of a rule. A task that has to be exactly right every time is a worse first candidate than one where a human already checks the output.

Once you have a candidate, the next step is a map, not a build. Spend a week or two understanding what the work actually requires, where the data lives, what “good enough” means, and what it would take to run it in production — not in a notebook. Write that down. A short, honest assessment that says “here is the one thing worth doing and here is what it costs” is worth more than three months of a project chosen on a whiteboard.

When you do start building, keep the first one small enough to finish. One workflow. One team that feels the pain. One outcome you can measure — minutes saved, tickets deflected, a manual step removed. A finished, modest thing that a real person uses beats an ambitious thing that demos well and ships never.

None of this is about AI, really. It’s the same discipline you’d apply to any system change: understand the constraints, find the highest-value low-regret move, ship it, measure it, then decide what’s next. The model is just the part that happens to be new.