Automation Can Acknowledge, Never Promise: Where Agents Need a Person
Small business owners were asked whether an automated message had ever promised a customer something the business would not do. The stories came quickly. One business had a template replying "we'll ship it tomorrow" to every email, including customers asking to cancel. It took three days before anyone noticed. The owner pulled the whole trigger and handled those emails personally for a week.
The failure in that story is a message that sent perfectly. Every test the business ran said the automation worked, because the tests checked whether it fired. Nobody checked what it said to the customer who wanted out.
The rule worth copying
One owner put the rule in five words: "automation can acknowledge, never promise." A reply that confirms the message arrived and says when a person will answer is safe. A reply that says yes, tomorrow or no problem is a commitment, made by something that cannot see whether it is true.
Others drew the same line from different directions. Anything involving money, scope or a published price stays with a person. So does the first reply to an unhappy customer. One owner tried auto-acknowledging complaints with a short apology and found it made cancellations worse. A templated sorry, at the moment someone is angry, told them they were not worth a real person. It went back to a human within a day.
A third put the test in one line: "If a mistake can reach a customer before a person sees it, keep a human check."
Where wrong promises come from
Most wrong promises are created later, by a change nobody connected to the automation. Someone raises a price, drops a service or changes Saturday hours, and every template that mentions it keeps promising the old terms for months. One owner found a stale discount column feeding invoices for about two weeks before anyone caught it.
The fixes owners described are plain. Every automation gets a named owner and a review date. Any change to price, hours or scope triggers a check of every message that touches it, the same day. A short list of words forces a person into the loop: cancel, refund, chargeback, not happy. And testing covers the ugly cases, such as a booking on a public holiday or a request for a service you stopped offering.
Agents raise the stakes
A template says one wrong thing until someone notices. An agent can say many different wrong things, fluently, to many customers at once. The answer is to decide in writing what the agent completes on its own, what it queues for approval and what it hands to a person, before it goes live.
That split is the core of how we run digital staff. Adam, our operations agent, works inside explicit boundaries of exactly that kind, with an audit trail of every action. It is also guardrail five of our framework, human in control. Anything touching a customer, a payment or a consequential decision has a review point built in. A person can override the system at any time.
None of this means keeping agents away from customers. Order status updates, booking confirmations and reminders for things already decided are good work for an agent. Sending them early heads off most of the where-is-my-order messages. The skill is drawing the line between facts the system can check and promises it cannot. It is the same split we described in the 90/10 rule of AI agents.
Before your next automation goes live
List every message your automations send. Mark each one that states a price, a date, a yes or an exception. For each mark, either connect it to data the system checks before sending, or rewrite it to acknowledge and route. Then put a name and a review date beside the list.
If you would rather design the boundaries with us, a digital staff trial starts free on one scoped workflow, with the limits written down before it begins.