The useful question was not whether AI could write outreach. It could. The useful question was where a human still changed the outcome.
The first map showed nine manual steps before a qualified account reached a seller. Five were data movement. Two were judgment calls with clear rules. One was a review. One was a relationship decision.
We did not replace the relationship decision. We rebuilt everything before it.
Agentic work should stop at the edge of judgment.
The flow now researches accounts, checks fit rules, drafts a reason to contact, logs the record, and prepares a review queue. A person still decides whether the account deserves human attention. The machine handles repetition. The team keeps judgment.
That distinction matters. Without it, AI becomes theatre with invoices.