Designing an AI agent that actually ships work

Drafting is not shipping

The demo version of an AI agent writes a convincing first draft and hands it to a human. That is genuinely useful, and it is also where most implementations stop. The agent produces output; a person still does the filing, the routing, the follow-up, and the closing. The hours saved are real but modest.

The version that changes the economics is the one that completes the transaction. It writes the reply, sends it, logs it against the right record, sets the follow-up date, and escalates when the reply does not arrive. Each of those steps is unglamorous. Together they are the difference between a tool and a teammate.

Scope the first version around one closed loop

Pick a single workflow with a clear beginning and a clear end, and automate all of it rather than most of several. A narrow agent that fully owns inbound lead qualification will outperform a broad one that half-owns six processes, because half-owned processes still require a human to hold the context.

Give it a hard boundary and an explicit escalation path. The agent should know precisely which cases it is not allowed to decide, and handing those off cleanly is a feature, not a limitation.