Customer Service in the Age of AI: Read the Customer First

●  START HERE

Your AI is right most of the time. That is exactly why you are going to miss the times it is not.

The draft that lands in your queue is fluent, polite, and correct. It answers the question. So you skim it, you send it, and you move on to the next one. The trouble is not that AI writes a bad reply. The trouble is what a good reply does to your attention.

There is a fix, and it takes ten seconds. I call it the Cold Read.

●  NOBODY TAUGHT YOU THIS

There is a name for this, and it comes from aviation

It is called automation bias. When a system hands you a recommendation, you tend to go with it, even when the information in front of you says something different. Pilots have trained on this for decades. So have radiologists and anesthesiologists. It is one of the most studied failure modes in the entire field of human factors.

Raja Parasuraman and Dietrich Manzey published the definitive review of it in Human Factors in 2010, pulling together thirty years of research on what happens to a human being who monitors a machine that is usually right. Nobody in customer service has ever been trained on any of it. And we just handed everybody a draft button.

Here is the part that should stop you cold. The better the system gets, the worse you get at catching the times it is wrong.

●  THE PART THAT STOPS YOU COLD

A tool that fails constantly keeps you sharp

Think about a system that breaks all the time. You do not trust it. You check its work by reflex, because you have been burned. Your attention never fully switches off, and the errors stay visible to you because you are looking for them.

Now think about a system that is right nine times out of ten. The first week, you check everything. The second week, you check most things. By the sixth week you are scanning for tone and hitting send, because ninety percent of the time that has been the correct call. The system did not get worse. You did.

This is the finding that makes the whole thing hard to fix, and I want to be honest with you about it. Parasuraman and Manzey found that automation bias is not cured by knowing about it, and it does not go away with experience or ordinary practice. Reading this article will not protect you. Awareness is not a countermeasure.

●  BE HONEST ABOUT THIS

Your leaders think you were trained. You know you were not.

Zendesk found that 72 percent of customer experience leaders say they have provided adequate training on generative AI tools. Fifty-five percent of the people actually doing the work say they have received none at all. That is not a small perception gap. That is two organizations living in the same building.

And the work itself got harder while nobody was looking. Gartner found that only 14 percent of customer service issues are fully resolved in self-service. Everything else reaches a human. Which means the easy contacts are gone from your queue. The ones that land on you now are the ones the machine could not finish, and they arrive with a confident draft attached.

●  SAY IT WITH ME

The fix is ten seconds. It is just an order of operations.

Read what the customer wrote before you read what the AI wrote. That is the whole technique. The reason it works is not willpower. It is sequence.

Once you have read a good draft, you cannot un-read it. You will go back to the customer’s message looking for permission to send. You will not be reading them anymore. You will be checking a box. Reverse the order and the same ten seconds buys you an independent opinion, which is the only thing you were ever there to provide.

Here is how it runs, contact by contact.

●  ASK THIS EVERY TIME

The one question that finds what the draft missed

That is where the misses live. Every single time. AI is excellent at answering the question that got asked. It is blind to the one underneath it.

The sentence about the wedding. The line about how many times they have already called. The part where they said they do not want a refund, they want somebody to tell them it will not happen again. The draft will answer the ticket. It will not answer the person.

Now, a fair objection, because I would rather you hear it from me than from the skeptic in your next meeting. A capable model can often catch all four of these if you ask it to. That is true. The problem is not that the technology cannot see them. The problem is that your workflow never asks, and a fluent draft is remarkably good at convincing you there is nothing left to ask. The Cold Read does not fix your AI. It fixes the ten seconds in front of it.

●  FOR THE PERSON RUNNING THE HUDDLE

Run this with your team in twelve minutes

You do not need a training day for this. You need one huddle and a rule everybody agrees to. Here is the shape of it.

●  IF YOU LEAD A TEAM

Two numbers your dashboard is not giving you

Every metric on your wall was designed for a job that no longer exists. Handle time, first contact resolution, and quality scores all assume the person is composing the answer. They are not. They are reviewing one. So measure the review.

Track both for four weeks. You will find that the Cold Read costs you a few seconds per contact and takes back minutes per restart. That is the business case, and it is the one your finance partner will actually accept.

●  THE PART THAT MATTERS

Your job changed and nobody announced it

You are not writing the reply anymore. You are the last human who reads it. That is a different job with a different skill, and almost nobody has been given the ten seconds of technique that the new job requires.

●  PEOPLE ASK ME THIS

Questions I get about AI and customer service

What is automation bias in customer service?

Automation bias is the tendency to accept a system’s recommendation even when the information in front of you contradicts it. In customer service it shows up as approving an AI draft that answers the ticket correctly but never addresses what the customer actually said. It was first documented in aviation and medicine, and Parasuraman and Manzey reviewed three decades of the research in Human Factors in 2010.

Does using AI make a support team worse at their jobs?

Not on its own. What degrades is the checking behavior, not the skill. The research is consistent that a highly reliable system reduces how carefully people monitor it, so the small number of errors it does make are the ones most likely to get through. The fix is a procedure that does not depend on how alert somebody feels that day.

How do you train a team to review AI drafts properly?

Give them an order of operations rather than an instruction to be careful. Customer first, draft second, one question before send. Then measure whether replies are missing things the customer raised, so the standard is visible instead of assumed. If you want this built for your team, with your own tickets and your own language, that is the work I do with clients.

Is it really faster to read the customer first?

The Cold Read adds a few seconds per contact. It removes whole repeat contacts, because the reply lands the first time. Track your Restart Rate for a month and the arithmetic will make itself. Speed measured one ticket at a time is the wrong unit.

What should you never hand off to an AI draft?

Anything where the customer is telling you what this cost them. Apologies, service failures, and any message where somebody has already contacted you more than once. A draft can carry the facts in those moments. It cannot carry the acknowledgment, and the acknowledgment is the entire reason they wrote.

MYRA GOLDEN  ·  CONFIDENCE IN EVERY CONVERSATION