Why does narrow beat broad?
When an AI agent makes the news for the wrong reason, the failure is almost never that the model was not clever enough. It is that the agent was allowed into a conversation nobody had written a rule for, and it answered anyway.
A narrow agent with an explicit refusal path outperforms a broad assistant on every measure that matters to an owner: fewer escalations that go wrong, fewer surprises, and a scope you can defend to a customer in one sentence. That is the same split we draw in AI agents for business and across the AI for Small Business hub: bounded, observable, reversible work first.
- Write the refusal rule before the capability list
- Name the human who receives every escalation
- Log every action, so a decision can be audited after the fact
We start with one queue, measure it, and only add a second once the first one is boring.