Should a person approve what an AI system sends to customers?
Usually yes at the start. Have a person read each message before it reaches a customer until you have seen it handle real examples from your business well. After that, it can be reasonable to relax review for low-stakes, routine messages, while keeping a person on prices, promises, apologies, complaints and advice. Whatever you relax, keep a log so you can look back at what was sent.
When to keep a person reviewing
- The system is new, or you have just changed what it does.
- The message states a price, a date, a refund or a promise a customer could hold you to.
- The customer is upset, or the message is an apology.
- The topic is health, money, law or anything where wrong advice could hurt someone.
- The message goes to many people at once, or cannot be recalled once sent.
- The message would be the first words a new customer ever hears from you.
When it can be reasonable to relax
Two things usually need to be true. First, the system has a track record on real examples: run it on your own past emails, calls or jobs, including the awkward ones, then in review mode for a few weeks, and count how often a person has to change or reject what it wrote. Second, the message is low-stakes and routine, such as an appointment reminder or a confirmation that a form arrived.
Relax one message type at a time, not the whole system. A message built from a fixed template with a few fields filled in is easier to trust than free-written text. If a mistake gets through, put that message type back under review. Automatic texts or calls go out only with the person's consent, whatever the review setting.
What to check first
- Facts: names, dates, times, amounts and addresses match your records.
- Promises: it offers nothing you cannot or will not do.
- The question: it answers what the customer actually asked.
- Privacy: it mentions nothing about another customer.
- Honesty: it does not claim to be a person. The related answer on telling customers you use AI covers whether to say so.
Keep a log
Record what was sent, to whom, and when. Record what the system wrote, whether a person approved, edited or rejected it, and what changed. A log shows whether the edit rate is falling, lets you trace a complaint, and gives you evidence to decide whether to relax review. It can hold customer details, so limit who can read it.
Look at the log every month. If edits stay frequent, review stays. The related answer on what parts of lead generation you can automate draws the line on what usually stays human, and the related answer on what to ask before hiring someone to build AI covers who reviews output and how you switch it off.
Related answers
- Do you have to tell customers when you use AI to answer or write to them?
- How can I automate my lead generation process?
- What should you ask before hiring someone to build AI for your business?
AI systems and automation
A lot of an owner's week goes to the same tasks: answering the same questions, chasing the same follow-ups, retyping the same details. We build the systems that do that work, so your people can spend their time on customers.