What to Build First When You Have Never Built AI

Four criteria

1. High volume — happens daily, so results appear in weeks
2. Low consequence — an error costs minutes, not money
3. Easily checked — a person can verify correctness at a glance
4. Genuinely wanted by whoever does the work now

A project meeting all four succeeds far more often than a better idea meeting two.

Five that reliably work

• Support triage — classify and route before anyone reads
• Sales research briefs — nothing customer-facing, immediate feedback
• Document extraction on one document type with review
• Quality checking of work people already produce
• Meeting notes to actions, into your task system

All five are checkable in seconds by the person receiving the output, which is the property that makes a first project safe.

What to avoid first

Anything customer-facing without review, anything financial without a rule layer, anything where the domain expert is your busiest person, and anything where success is defined vaguely.

Those are all buildable later, with more credibility and a team that has learned how these projects work.

Set the success measure now

One number, agreed before starting, that means it worked. Time per item, straight-through rate, response time — whichever fits.

Also agree what would mean stop. A first project with a defined stopping condition is much easier to get approved.

Then decide the second one

The value of a first project is partly the capability: an evaluation discipline, a monitoring pattern, people who know what these systems need.

The second project is faster and cheaper because of it, and it can be more ambitious for the same reason.

Frequently asked questions