AI customer service: getting the balance right
The only test that matters is an angry customer. A system that handles one well is fit to deploy; one that answers with automated warmth doubles the anger.
Every discussion about automating customer service revolves around what the system can do. The right question is the reverse: what happens when it cannot?
Because that is what the customer remembers.
The angry customer test
This is the only test I insist on before any launch.
Write a genuinely angry message: the order is late, you have paid, and nobody has replied for two days. Then watch what the system does.
A bad system replies with automated warmth: "We're sorry for the inconvenience! We're happy to help." That sentence doubles the anger, because it proves nobody read it.
A good system does three things: acknowledges specifically ("your order is three days late"), gives real information or says it does not have it, and escalates immediately without being asked.
What automates safely
| Type | Why |
|---|---|
| Where is my order | Answered from data, not opinion |
| Hours and location | Fixed and definite |
| How to return something | A written procedure |
| Password reset | Already automatic |
| Confirming an appointment | A simple transaction |
The common factor: one correct answer, and no emotion in the question.
What never automates
Complaints. Any complaint is a person.
Money. Refunds, discounts, billing disputes. Errors here cost directly.
Exceptions. "My case is different" means, by definition, that the automation does not apply.
Cancellations. The last conversation with a departing customer should always be with a person. Sometimes they stay.
Designing escalation
Three rules are enough:
A visible exit, always. "Talk to someone" visible in every message, not hidden behind three failed attempts.
No repetition. When the customer reaches a person, that person must see the whole conversation. Asking a customer to explain their problem again is the worst thing automated systems do.
Say so. Declare it is an automated system from the start. People accept that and resent discovering it late.
The right metrics
First contact resolution. Regardless of channel.
Time to reach a human. When one is asked for.
Repeat contact rate. A customer coming back three times was not helped the first time, whatever the system recorded.
What it could not answer. This is the most valuable list you own, and it improves both the system and the product.
In closing
Design for failure rather than success. A system that knows its limits and escalates quickly builds more trust than one trying to answer everything.
Try this today: send an angry message to your current system and watch what it does. The result will decide your next plan.
Common questions
- What can safely be automated in customer service?
- Anything with one correct answer and no emotion in the question: order tracking, opening hours, returns procedures, password resets, appointment confirmations.
- What should never be automated?
- Complaints, anything involving money such as refunds and disputes, exceptional cases, and cancellations. The last conversation with a departing customer should be with a person.
- How do I test a customer service system before launch?
- The angry customer test. Send a genuine complaint and watch whether it acknowledges specifically, gives real information or admits it lacks it, and escalates immediately without being asked.
- Which metric should be avoided?
- Percentage of conversations resolved without a human. Once it is a target, the system gets designed to avoid escalation. Measure first contact resolution instead, regardless of who resolved it.
No comments yet
Leave a comment