The most useful design rule we have for AI employees fits on one line: anything customer-facing pauses for review before it sends.
An AI can be right most of the time and still cause damage with the occasional wrong reply. The risk is not just that it makes something up. It is that the answer reaches a customer before anyone checks it.
Build review into the process from the start. The AI prepares a draft, a person checks and approves it, and only then does it send. That review is part of the job.
Three things follow from this.
First, the AI prepares the work and the team approves the response. Researching the customer, reading the last thread, and finding the right tone take time. A useful draft brings that work together so the reviewer can check it without starting from scratch.
Second, sensitive categories always need human review: refunds, legal language, regulatory commitments, and cases that need a supervisor. A confident draft does not bypass that rule. A short delay is easier to recover from than a wrong promise.
Third, record the team’s corrections so later drafts need less editing. Keep clear rules for what to draft, flag, or archive. Faster review is useful; it is not a reason to stop checking the work.
A customer-service supervisor in financial services described an AI employee trained not to promise a guaranteed refund, even under pressure from a caller. A person still approves cases that need escalation. The useful safeguard is the review step, rather than a claim that AI will always follow the rule.
Build the review step first. Build the rest after.