The most useful design rule we have for AI employees fits on one line: anything customer-facing pauses for review before it sends.
Not because the employee is unreliable. Because the things the employee gets right 99% of the time will, on the 1%, be visible to a customer. That asymmetry — high accuracy, high blast radius on errors — is what kills agent products in production. It is not the agents that hallucinate. It is the agents that send.
The fix is review-before-send as a first-class concept. Drafts queue. A human approves. The send happens. This is not a workflow extension; it is the trust contract.
Three things follow from this.
First, the employee’s job is to prepare the work, not commit it. The expensive part of any customer email is the research, the tone, the timing, the recall of the last thread. The cheap part is hitting send. Done right, the employee does the expensive part and leaves the team with the cheap part. Most days they spend two seconds per draft.
Second, sensitive categories never bypass the queue. Refunds, legal language, regulatory commitments, anything escalatory. Even when the employee is sure, even when it is well within historical norms, the queue catches it. The cost of a slow send is one minute. The cost of a wrong send is a lost customer.
Third, the queue gets faster every week as the team teaches the employee what to draft, what to flag, what to silently archive. After a month the queue is small. After three months it is mostly noise. The team is doing twenty percent of the work and getting ninety percent of the credit.
A customer-service supervisor in financial services described how their AI employee never promises a refund guarantee, even under pressure from a caller. The team trained it on the rule. A human still approves anything escalatory. The point is not that the AI is more honest than a human rep. The point is that the queue is the harness.
Build the harness first. Build the rest after.
Related Where we wire approval gates →