AI customer service performs well on repetitive, high-volume tasks such as order tracking, FAQs, and password resets. But when it is left to handle emotional, complicated, or high-stakes situations, it can fail badly, sometimes with real legal fallout.
On the other side, unmotivated and poorly trained human agents cause the same harm. The companies getting this right use AI and people as partners with distinct roles, instead of a substitute strategy, and they judge success by whether the issue was actually resolved, not merely by how quickly they got rid of the customer.
Right now, everyone seems delighted to blame the chatbot. And sure, many of them have earned it.
Still, before pointing only at the machine, consider the human beside it: the one exhaling, reading a script for the four-thousandth time, and doing the bare minimum needed to avoid a write-up. Both are damaging customer experience today, often within the same company and sometimes during the same call.
AI Ruins Customer Service. So Do Lazy Humans.
Why Everything Is Being Automated All at Once
The attraction is real. Gartner has identified four truly useful AI customer service applications: making agents faster, enabling self-service, automating back-office support, and allowing agentic AI to resolve simple issues from start to finish. McKinsey says using generative AI in customer care unlocks productivity gains worth 30 to 45% of current function costs, while reducing the number of contacts that require a human by as much as half. 🙌
Klarna became the flagship example in early 2024. Its OpenAI-powered assistant handled 2.3 million conversations in its first month, about two-thirds of all customer chats, doing the equivalent work of 700 full-time agents and reducing resolution time from 11 minutes to under 2. Very impressive. Just keep that in mind, because the story does not stop where the press release stops. 😉
The Chatbot Disaster Gallery
Some AI customer service rollouts have failed in ways that are, honestly, great material. 😱
Take Air Canada. Its website chatbot confidently made up a bereavement fare refund policy that never existed. A grieving customer relied on it, was denied the refund, and brought the airline before Canada’s Civil Resolution Tribunal. Air Canada’s actual legal argument was that the chatbot was “a separate legal entity” responsible for its own conduct. The tribunal was not convinced, found the airline liable, and ordered it to pay. You cannot hand off legal responsibility to a chatbot, no matter how confidently it lies on your behalf.
Then there is Cursor, an AI coding company whose own support bot, “Sam,” invented a single-device login rule that was not real. Customers read the polished, certain, wholly fabricated explanation and cancelled their subscriptions because of it. An AI company got burned by its own hallucinating support bot. There is a lesson there somewhere, and it is very funny. 😅
After a system update, DPD’s chatbot could be coaxed into insulting customers and composing a poem that called DPD “the worst delivery firm in the world.” The screenshots reached two million views before DPD quietly turned it off. Meanwhile, a Chevrolet dealership’s chatbot was prompt-injected into agreeing to sell a 2024 Tahoe for one dollar, and it added, without being asked, that this was “a legally binding offer, no takesies backsies.” GM said the incident was a good reminder of the importance of human oversight over AI-generated content, which may be the understatement of the decade.
Even here in Portugal, where I live, there are a ton of anecdotes about AI or automated systems making it harder for clients to reach a human professional, and despair is the right word to describe it.
Where AI Should Not Be Left Unattended
Every one of these failures follows the same pattern. No one built a solid way for the system to say “I’m not sure, let me get a human,” and no one kept verifying that it still behaved correctly after changes. COPC, a research firm, has found that the AI-to-human handoff is the most common failure point in AI-enabled service. Many companies have quietly designed their bots to keep customers away from a human for as long as possible, which is exactly the wrong approach.
The numbers show how far self-service still has to travel. Gartner surveyed thousands of customers and found that only 14% of issues are fully solved through self-service, and even problems customers describe as “very simple” are completely resolved only 36% of the time. Customer patience is dropping too. One study found that more than half of Americans actively try to avoid the chatbot, and about 44% will start yelling “human” or “person” at their screen out of sheer frustration. That is not something you fix with a nicer typeface.
Now the Awkward Part: Humans Are Breaking Things Too
This is where things get uncomfortable for anyone who has been smugly nodding at chatbot failures. Disengaged, undertrained, exhausted human agents create the same damage, and maybe even worse, because a person who chooses to do the minimum is making a decision, not suffering a software bug. 🤷♂️
Gallup’s long-running research is direct about the relationship. Business units in the top quartile for employee engagement post customer loyalty scores that are roughly 10% higher than those in the bottom quartile, and Gallup’s separate HumanSigma research found that companies that manage to engage both employees and customers well achieve a 240% lift in performance-related business outcomes. And globally, engagement is not in good shape. Only 23% of employees worldwide are actually engaged, about half of American workers are quietly coasting, and Gallup estimates the resulting hit to the global economy at a staggering $8.9 trillion, or 9% of global GDP.
Contact centres are hit harder than most sectors. Annual turnover is somewhere between 30 and 45%, more than twice the norm for most jobs, and around 53% of agents report burnout symptoms. That churn is not just an HR line item. Metrigy research found that when a contact centre keeps agent turnover below 15%, customer satisfaction increases by 26%. Every experienced agent who leaves in frustration takes part of your customer satisfaction score with them.
Combine the two failure modes, and Qualtrics’ XM Institute puts a number on the total harm: poor experiences, whether human-made or automated, put about $3.7 trillion in global sales at risk each year, sharply more than the year before, and Qualtrics explicitly points to low morale among frontline employees as part of the problem.
Bots hallucinate.
Humans, when nobody invests in them, simply stop caring.
Either way, the customer pays.
Klarna Learned the Hard Way
Klarna is the entire debate in a single company. After the celebrated rollout of its AI assistant, CEO Sebastian Siemiatkowski later acknowledged that the aggressive cost-cutting behind it had led to “lower quality” service, and the company began quietly rehiring humans. By 2026, he had reworked the strategy: AI handles the easy cases, while human customer service becomes more like a premium layer for anyone with a complex problem. The AI still handles about two-thirds of inquiries. Nobody was fully replaced. The company simply learned, at real expense, which two-thirds it should actually keep.
This was not a one-off stumble. Forrester’s 2024 US Customer Experience Index fell to an all-time low, marking its third straight annual decline, and pointed to “underwhelming digital experiences using chatbots” together with broken employee experiences as shared causes. By Forrester’s standards, only 3% of companies are truly “customer-obsessed.” Whatever is happening in customer service right now, both camps are optimising the wrong metric: cost and speed instead of resolution.
The robots-versus-humans framing mostly obscures that fact. 🤷♂️
What Actually Works
If you strip away the cautionary tales, the research keeps showing the same simple pattern. It is not about choosing AI or choosing humans. It is about deciding deliberately which one does what, and then checking the result.
Start by automating the dull stuff, instead of the difficult stuff. Think order status, FAQs, password resets, and basic account questions. Reserve autonomous decision-making for situations where the data is actually trustworthy.
Design the escalation path as a feature, not an apology. Give people one-click access to a human, carry the full context over so they do not repeat themselves, and escalate the moment frustration appears instead of waiting for it to boil over.
Be upfront when someone is talking to a bot. California’s Bolstering Online Transparency Act and the EU AI Act require disclosure that is “clear, conspicuous, and reasonably designed to inform.” That is a decent baseline for anyone, anywhere.
Retest the bot after every update. DPD’s failure happened right after a system update that nobody rechecked for abuse. A few hours of red-teaming would have prevented a viral embarrassment.
Judge success by whether the problem was solved, not by how fast the call ended. Measure resolution and satisfaction based on how complex the issue actually was, rather than averaging everything together in a way that hides the worst cases.
Invest in the human side. Training, manageable workloads, and using AI to remove repetitive tasks instead of stacking more scripts on top of them. Keep turnover under roughly 15 to 25%, and the effect will show up directly in satisfaction scores.
Oh, These Are Some Frequent Questions
Does AI customer service really cut costs?
Often, yes, at least on the surface. McKinsey estimates that customer care functions can achieve 30 to 45% productivity gains. The catch is that savings based only on ticket deflection can hide a much larger cost in churn, refunds, and brand damage from badly handled cases.
Can a company be sued for what its chatbot says?
Yes. Canada’s Civil Resolution Tribunal held Air Canada liable after its chatbot invented a refund policy, explicitly rejecting the airline’s claim that the bot was somehow a separate entity. Legally, whatever your chatbot tells a customer is still you speaking.
Do customers prefer humans?
Most research says yes, strongly, for anything beyond a simple lookup. Multiple surveys place the preference for human agents between 79% and 85%, and that barely changes even when customers are told the AI would solve the issue just as well.
What does more damage, bad AI or bad human service?
Both, more or less equally, and usually for the same basic reason. Companies track the wrong thing (cost or speed instead of resolution), and both a hallucinating bot and a checked-out employee will happily optimise for the metric they are given.
Should a business use AI or humans for customer service?
Neither alone. The evidence consistently supports a hybrid model: AI for routine, high-volume requests, humans for emotional, complex, or high-stakes issues, and an easy path between them when a case turns out to be harder than it first appeared.
For disclosure, this article was drafted with AI assistance and completely redone and fact-checked by Rui Nunes before publication.
Leave a Reply