Should you let AI reply to customers without a human?

Treating this as a single yes or no is what produces both of the bad outcomes: businesses that let a system answer anything, and businesses that will not let it send a message at all. The question separates cleanly into three, and the three have genuinely different answers.

Acknowledging is the first and it should be autonomous without much argument. Confirming that an enquiry arrived, restating what was asked so the person can see it was received, and committing to a specific next step and time. The failure mode is trivial — an acknowledgement that was slightly generic — and the alternative is the enquiry sitting unanswered, which is the loss that actually costs money. There is no serious case for a human in this loop.

Answering is the second and it depends on what is being answered. Factual questions with a stable, checkable answer — opening hours, whether you cover an area, what a service includes, whether a part is in stock — are well suited, because a wrong answer is discovered immediately and corrected cheaply. Questions whose answer depends on the specific customer's situation, or where being wrong sends them away with a false belief, are not. The dividing line is not the difficulty of the question but the cost and visibility of getting it wrong.

Acting is the third and it is where the boundary should be firm. Booking, rescheduling, quoting a price, promising a date, issuing anything. Not because a system cannot perform the action but because the action cannot be withdrawn, and the business will spend more managing a wrong one than it saved by automating a hundred right ones. Where speed genuinely requires autonomous action — an appointment slot that will be gone by morning — the mitigation is a reversible commitment: a held slot with confirmation, rather than a booking.

Underneath all three sits the property that matters more than the category boundaries: what the system does when it does not know. A system that stops, says so and escalates is safe in a much wider range of tasks than one that continues plausibly, because fluency is exactly what makes a wrong answer expensive. A customer told something confidently and incorrectly acts on it, and the business finds out when the consequence arrives. That behaviour is testable before purchase by asking the thing questions outside its scope and watching whether it declines or improvises.

The second property worth designing deliberately is disclosure. There is no good reason to hide that a first response is automated, and a real cost to being caught: it converts a minor annoyance into a question about the business's honesty. Saying plainly that a reply is automatic and a person will follow up costs nothing, sets accurate expectations, and removes the failure where a customer replies to a machine believing it was a person and receives nothing back.