Automating Bad Customer Service Just Gives You Faster Bad Customer Service

AI is transforming customer service, but without fixing policies and processes, automation risks scaling poor experiences instead of improving them.

Every week, I speak to a brand that has just signed an AI contract for its customer service operation. They’re excited, their CFO is excited, and somewhere in the business, a board slide is being updated with projected deflection rates. 

What isn’t being openly discussed is whether any of this will improve the customer experience.

In most cases, it won’t. Not because the AI doesn’t work, but because the problem it solves isn’t the problem customers actually have.

Here’s a version of events I keep seeing: a subscription brand instructs its AI not to answer questions about how to cancel. The bot is technically functioning – handling hundreds of queries a day, routing, responding – but the moment a customer types “how do I cancel,” it quietly goes around the question. 

The customer tries again, gets nowhere, eventually finds a phone number, waits in a queue, cancels out of frustration, and leaves a one-star review on the way out. What that brand spent significant budget on automating was the experience of not being helped. The AI didn’t create that outcome. 

A decision made somewhere in commercial or finance, that friction was good for retention, created it. The AI just runs it faster and at a greater scale.

AI Is Only as Good as the Decisions Underneath It

I sell AI for customer service, which means I have an obvious incentive to overstate what it can do. 

So let me say something people in my position rarely say publicly: 

• AI cannot fix a bad refund policy.
• It cannot fix hidden fees or a checkout designed to obscure the true cost. 
• It cannot undo the decision to route customers through five automated steps before they can reach a human.

These are the choices brands have made about what they are willing to do for their customers, and technology can execute them very efficiently. It cannot make them less bad.

When I look at brands with Trustpilot scores in the twos and threes, the diagnosis is almost never outdated software. Its policies are written by the finance department and optimised for cost avoidance, handed to a support team with an impossible brief: make people feel good about decisions that weren’t made with them in mind. 

There is only so much that good people, or good AI, can do with that starting point.

Many Brands Aren’t Ready for AI, and the Solution is Simpler Than They Think

Before any AI conversation, there’s one question I think every brand should ask itself: Do you have a documented knowledge base where your policies are written down, aligned, and consistent across the business? 

The honest answer, more often than not, is No. 

What most brands have instead is a group of experienced agents with their own Word documents and copy-pasted answer archives — people who have been around long enough to know the unofficial rules that never made it into any system. 

Any AI layered on top of that will simply inherit the gaps and contradictions underneath.

And below that is a more basic problem. At least half the brands I speak to are still handling queries manually that have nothing to do with human judgment or empathy: order tracking, address changes, return initiations, and basic status updates. 

These should have been automated years ago, not with large language models, but with simple logic. They weren’t, because no one prioritised it. Now they’re being bundled into AI projects that are far more complicated and expensive than the underlying issue requires.

The right sequence is to fix the basics first, document what you actually do, make sure your policies reflect what you want customers to experience — and then automate.

Successful Brands Started with a Different Conversation

I bought engraved shirts from a UK retailer with my initials on them, and ordered the wrong size. I assumed I was stuck; they had my name on them. Their response was to tell me to send them back, no problem. That decision cost them a shirt and bought them a customer who hasn’t gone anywhere else since.

That wasn’t an AI decision. It came from a brand that had already worked out what kind of experience it wanted to deliver, and built everything else around that answer. 

When that same brand faces peak-season volume, its automated flows handle high-frequency, low-complexity queries quickly and effectively — not because the technology is particularly exceptional, but because what sits beneath it is clear and fair. 

Contrast that with brands where the C-suite conversation is about how to make a cancellation button harder to find, or why customers in different countries should be shown different prices from the same warehouse. The support team, human or AI, will bear the consequences of those decisions. They cannot fix them.

Until Support Can Prove Its Value, Nothing Changes

There’s a deeper problem running through all of this. Marketing has spent 20 years building rigorous ways to measure its own contribution through attribution models, return on spend, and cost per acquisition, and the result is budget, headcount, and genuine influence over strategy. Customer service hasn’t done that work. 

A Trustpilot score moves from three stars to four, the team is pleased, and then structurally nothing changes, because nobody has calculated what that improvement was worth in reduced churn, in customers who didn’t need to be reacquired, in referrals that came for free. 

So support stays a cost line, and when support is a cost line, AI gets deployed to shrink it rather than improve what it produces. 

That loop is hard to break, but it starts in the same place: with brands being honest about what they actually want their customer service to do before they decide how to automate it. 

They shouldn’t spend too much time picking which AI vendor or which AI model. They should ask themselves whether they are on their way to automating something they are proud of. If the honest answer is no, the technology will make that faster. It won’t make it better.