The hard part isn’t building agents. It’s knowing which agents to build. Heard in a public company boardroom.Selection comes first. Choosing the right problem is higher-leverage than how you build it, and it’s where most of the value, or the wasted effort, is decided.
The screens
Mind the jagged frontier. AI is unevenly capable, strong at one task and surprisingly weak at an adjacent one. Don’t assume a use case is feasible because a neighbor is. Carry several candidates and test each. Favor problems that don’t demand multiple nines of accuracy. The best early candidates are high-volume and repeatable, have data already in the Company Brain, and carry an error cost a human gate can absorb. Qualify each candidate against three tests:
Know what not to touch. Irreducible human judgment, a regulatory floor on human review, or volume too low to pay back are signs to decline rather than force.
Apply an enterprise bar. Prioritize work that spans more than one team or touches a system of record and has a durable, named owner, rather than personal-productivity hacks.