Short answer: It depends on who is affected, how bad the worst case is and whether a mistake can be undone. Rate it before you build. A reply that confirms delivery hours is a very different risk from a reply that promises a refund, quotes a price or answers a complaint.
I'm Vinay Pasricha, founder of GoodSpace AI and author of AI for Business Leaders. At GoodSpace, AI now talks to many people on our behalf, so I've thought about this question in practice, not just on paper.
The question isn't "AI or no AI"
Owners often ask whether AI should talk to customers at all. That's the wrong frame. Some replies are almost risk-free. Others can cost you a customer, money or your name in the market. The useful question is: for this kind of reply, what controls do we need?
Rate the use before you build it
In the book I use a ladder with four rungs, from Tier 1 (low) to Tier 4 (critical), so that the controls match the stakes. To place a customer reply on it, ask:
- Who is affected? One customer, many customers, or the public?
- How bad is the worst case? A slightly awkward message, a wrong promise, or something that harms someone?
- Can it be undone? A wrong delivery time can be corrected. A public reply, a wrong price that was accepted, or a leaked detail often can't.
- Does it involve money, safety, legal terms or personal data? If yes, it climbs the ladder fast.
What this looks like in practice
- Low: order status, opening hours, where to find a document. AI can reply on its own, with a clear way to reach a person.
- Moderate: product questions and first replies to enquiries. AI drafts, and either sends within tight limits or a person checks a sample every day.
- High: prices, discounts, delivery commitments and complaints. AI prepares the reply; a named person approves it before it goes.
- Critical: anything legal, anything about safety, any dispute. People handle it. AI can help them find the facts.
The rule behind all four rungs: AI recommends, humans decide. The higher the rung, the more literally that applies.
Controls that matter at every rung
- Tell customers when they are talking to AI, and how to reach a person.
- Give the AI only the information it needs for that kind of reply.
- Review a sample of real conversations every week, especially the ones that went badly.
- Decide in advance which words or topics hand the conversation to a person.
One thing to do this week
List the five kinds of message customers send you most often. Rate each one with the free Risk Ladder tool, which places a use on its tier and suggests the safeguards to match. Start with the lowest-rated kind, and leave the high ones until you have seen how the low ones behave.
Where this comes from
This answer comes from The Risk Ladder, one of the frameworks in AI for Business Leaders. The book is in hardcover and on Kindle.
To check a specific use, describe it on Ask Vinay's books.
Vinay Pasricha is the founder of GoodSpace AI and the author of six books, including AI for Business Leaders.
