Core question
Should our first AI project be large or small?
Short answer
Small, narrow, and measurable. Pick one workflow. Fix the data issues in that workflow. Learn from it. Then expand. Small focused wins beat big vague moonshots, because a narrow project produces evidence you can act on and a broad one produces a status update.
Start with focused pilots aimed at narrow, well-defined problems. Gather information. Evaluate. Adjust. Expand. Micro-stack applications, meaning small contained builds that touch one workflow and one set of data, are the best place for most companies to begin.
The distinction that matters is not small versus ambitious. It is narrow and measurable versus broad and performative. A company that mandates "implement AI" and lets teams pick tools first ends up with press releases instead of results. One year later there is nothing to show, and the skeptics inside the building have won by default.
There is a second reason to start narrow. A contained project is where your team learns the unglamorous disciplines that determine whether anything larger will work: data quality, cleaning, validation, security, and ownership. Learning those on a small internal project is far cheaper than learning them on a customer-facing one.