Two failure modes bracket AI client communication. Full auto: the model answers everything, and one hallucinated commitment to an angry client costs the account. Full manual: every AI draft waits for a human click, which is the old workload wearing a new interface. The design question isn't whether humans stay in the loop — it's where the loop is.
Three lanes, drawn in advance
Green — send freely. Provable from the board, reversible, routine: status answers, deadline confirmations, "received, it's on the board" acknowledgments. If it's wrong, the board was wrong, and that's auditably fixable.
Yellow — draft and hold. Anything touching money, dates or feelings: the reply exists in seconds, waits visibly for approval, and a person sends it with their name on it. Out-of-scope answers live here — the agent classifies and drafts; the human owns the quote.
Red — humans only. Complaints, negotiations, bad news. The agent's only job is routing it fast to the right person with full context, and staying silent — knowing when not to answer is half of what makes an agent trustworthy.
The two details that make it real
The lanes must be rules, not vibes — a model deciding per-message how confident it feels is full-auto with extra steps. And every held draft needs a timeout: yellow items that sit unapproved escalate like any overdue task, so "waiting for approval" never becomes the new silence.
Done this way, the human touches perhaps a tenth of the traffic — but it's precisely the tenth where judgment, accountability and a name matter. That's not a compromise on automation. That's the design.
← All posts