The AI triage: what to build, buy, or ignore

Founders don't have a shortage of AI ideas. They have a shortage of a filter. Every roadmap review surfaces a dozen plausible ones, and the honest answer is most of them shouldn't get built. This is the filter I run with founders — four questions, in order, each one a gate the idea has to clear before the next one is even worth asking.

What metric does this move, and by how much. Not "it'll help with X" — a number, an owner, a before and after. No metric, no funding. This one question kills more ideas than the other three combined.

What happens when it's wrong. The cost of error sets the guardrails, not the other way around. An AI that drafts marketing copy can be wrong cheaply. An AI touching pricing or credit at scale needs a completely different standard of containment before it ships.

Is this your edge, or is it a commodity. Buy the commodity layers. The fifth company to build its own embeddings pipeline isn't building a moat, it's building a maintenance burden. Build only where the capability is the thing customers are actually paying for.

Can you verify it in production. Not does it work in the demo — can you tell, on real traffic, whether it's working. No, and it doesn't ship. Doesn't matter how promising the sandbox looked.

Run enough of these and a pattern shows up fast:

No metric → not funded. Commodity → buy. Unverifiable → doesn't ship.

The value here is subtraction, not addition. The ideas this stops a founder from funding save more time and credibility than the ones it greenlights ever earn back. A good triage isn't a list of what to build. It's a list of what you were right to walk away from.