Why your first AI project should be a thin slice

Most AI initiatives don't fail on the model. They fail on scope: a platform program that promises to transform everything, burns two quarters on integration meetings, and never puts working software in front of the people who do the work.

The alternative is unglamorous and it works: pick one workflow that hurts, build the thinnest slice of software that touches it end to end, and run it on real data within three weeks.

What a thin slice looks like

• One workflow, not a platform — quoting one product family, not "quoting".
• Real data from day one — your drawings, your spreadsheets, your inbox. Demos on synthetic data prove nothing.
• A human in the loop — the AI drafts, a person signs off. Trust is earned in weeks of side-by-side comparison, not claimed in a slide.
• A measurable before/after — minutes per task, quotes per week, stockouts per season. If you can't measure it, you can't defend the budget.

Why this beats the big program

A thin slice de-risks the three questions that actually kill projects: Is the data good enough? Will the team use it? Is the payoff real? You get answers in a month for the cost of a pilot, instead of discovering the answers in month nine of a platform build.

You see the result before you commit to it — that's the whole trick.

When the slice proves out, widening it is a series of small, defensible decisions. When it doesn't, you've spent weeks, not quarters — and you know exactly why.