Most AI initiatives don't fail because the technology fell short. They fail because the first project was chosen for excitement instead of payback — too broad, too deep in the critical path, or impossible to measure. Here's the filter we apply with every client before a single model is touched.
Start where judgment is repetitive
Look for decisions your team makes many times a day using roughly the same reasoning: routing an inquiry, checking a document, flagging an anomaly. Repetitive judgment is where AI pays back fastest — it's well-bounded, it happens at volume, and a human can verify the output. Creative, one-off, high-stakes decisions are exactly the wrong place to start.
Demand a number before you build
A good first use case has a metric attached before the build begins: hours of data entry per week, response time to leads, percentage of no-shows. If you can't say what number should move, you won't know if it worked — and neither will the sceptics in your team whose support you'll need for use case number two.
Stay off the critical path
Your first AI system should assist, not gatekeep. Let it draft the reply a human approves, pre-fill the form a clerk verifies, flag the exceptions a manager reviews. Once it has months of accuracy behind it, you can promote it. Starting in the critical path means one early mistake buys the whole programme a bad name.
Use the data you already have
The best first project runs on data that already exists — your chats, invoices, sales history, registers. If step one of the plan is 'first, collect data for six months', that's a second project wearing a first project's clothes. There is almost always a use case sitting on data you already own.
Pick a repetitive, measurable, assist-mode use case that runs on existing data. Prove it in weeks, publish the number, and let that result buy permission for the next one. That's the entire playbook — the rest is engineering.