On this page
When we build automations that write to customers, one rule never moves: the AI may draft the reply, and a person decides whether it goes out.
This post explains the thinking, what is allowed to be automatic, and the two habits that go with the rule: a plan for when the automation breaks, and a test with your worst messages before it goes live.
The rule, and why we hold it
An AI is good at a first draft and at remembering the details. It can also be confidently wrong. It can sound cold when the customer is upset. It can promise something you cannot do. And when it sends the wrong message, it is your name on it, not the AI's.
Two costs, side by side
The cost of checking is small. A draft arrives on your phone, you read it in seconds, and you tap approve.
The cost of a single bad message can be a lost customer, a public complaint, or a promise you now have to keep. Compare the two and the decision makes itself.
A small example
Say an order shows up days late and the customer writes in annoyed. A draft could start breezy and promise a refund that no one signed off on. Whoever reviews it spots both problems at a glance, deletes the promise, and replaces the chirpy opening with an honest apology.
A few seconds of editing separates a saved customer from an angrier one.
Does that mean nothing is ever automatic?
No. Fixed confirmations, like "your appointment is booked", can be plain templates that you wrote, with no AI improvising. Internal alerts can be automatic too.
The line we draw is simple: if an AI is writing words to a customer, a person approves them.
Over time the drafts improve, because you fix the instructions once and every later draft benefits. A good reply goes unnoticed and a bad one gets remembered, so review exists to catch the bad ones before anyone outside sees them.
Decide what happens when it breaks
Assume your automation will break at some point, and decide today what happens then.
If a step breaks, the customer should still get an answer. Send the message to a person and tell them what failed. Name one owner and one place for failures, so nothing vanishes unnoticed.
Test it with your worst messages first
Before your automation goes live, send it your worst messages:
- An angry customer.
- A message in another language.
- A blank one.
- A question it should not answer.
See what it does, fix what breaks, and decide who gets alerted when it does. Then switch it on.
How this shows up in the automations we build
You can see the rule in action across our other posts. The ten automations post has the approve-then-send pattern in most of its examples, and missed call text-back for trades shows the other side: a text you wrote yourself, sent exactly as written.
Frequently asked questions
Does an AI ever send a message to a customer alone?
No. The AI may draft the reply, but a person approves it before it goes out.
Is anything allowed to be fully automatic?
Yes. Fixed confirmations, like your appointment is booked, can be plain templates that you wrote with no AI improvising. Internal alerts can be automatic too.
What happens if a step in the automation breaks?
The customer should still get an answer. It is sent to a person who is told what failed, so no message vanishes unnoticed.
What should I test before going live?
Send it your worst messages: an angry customer, a message in another language, a blank one, and a question it should not answer. Fix what breaks and decide who gets alerted, then switch it on.