Getting the dose of AI in customer service right: 5 reasons more automation is not the answer

You know this moment. You are stuck in a chatbot loop, typing your issue for the third time, and the bot replies politely past the point. At some stage you just type: “AGENT. HUMAN. PLEASE.”

At the same time, a lot of businesses are being told to put even more AI into customer service. More bot, more self-service, more automation. Almost everyone giving that advice sells the matching software. I build software architecture and always start from a different question: how can you use AI in a low-threshold, brand-safe way, if at all? Where a chat component then belongs, the brand voice and above all the handover to humans are anchored deep in the concept. I earn nothing from you automating more. Which is why I can tell you the sentence you will not find in the vendor guides: more AI in customer service is not the answer. The right dose is.

The numbers are clear. Nearly one in five people who have used AI in customer service says it brought them no benefit at all. That is a failure rate almost four times higher than for AI use in general (Source: Qualtrics, 2025). And here is the good news, right up front: this is not an argument against AI in customer service. It is an argument for the right dose. And getting the dose right is easier for a small business than for a corporation. That is what this article is about: five reasons why “more” points the wrong way, and what the right dose looks like.

Why does AI fail so often in customer service?

Because service issues are rarely standard. AI in customer service fails about four times as often as other AI applications, measured by customers who experienced no benefit.

That is reason one, and it comes from the Qualtrics 2026 Consumer Experience Trends Report: consumers rank AI applications in customer service among the worst of all AI use cases for convenience, time savings, and usefulness. Of all places, the technology delivers its weakest results exactly where your business builds trust.

Why? Think for a second about what actually reaches your service team. The standard questions are already handled by your website. Opening hours, prices, delivery status: that lives in the FAQ. What gets through is the rest. Edge cases. Emotions. Things that went wrong. Precisely the terrain where AI solutions are weakest.

And when something has a four times higher failure rate, “more of it” does not scale the benefit. It scales the bad experience.

Why do most bot conversations end up with a human anyway?

Because the bot cannot solve the hard cases. 82 percent of respondents say chatbot conversations usually escalate to a human in the end.

Reason two is what I call the escalation boomerang. The number comes from a survey the research firm Sago ran in September 2025 among 1,500 US adults, commissioned, of all companies, by an AI vendor. Which makes the result even more remarkable: 82 percent say the bot session escalates to a human anyway (Source: CMSWire, 2025).

That is like a detour that leads back onto the same road. You have not shortened anything. You have just driven longer. The bot did not replace the conversation, it just added a frustrating extra leg to the trip. And the human who then takes over often starts from zero, because the context from the bot conversation does not travel with the customer. This is exactly why I anchor the handover to humans in the concept from day one instead of leaving it to chance. What that looks like in practice comes further down.

The same survey shows something else: 60 percent say an accurate resolution matters most to them. Only 31 percent put speed first. So the bot optimizes exactly the metric people care about least. And your customer can immediately tell the difference between “they want to help me” and “they want to get rid of me”.

What other downsides does AI in customer service have?

The biggest one never shows up in a dashboard: customers see through the motive. They do not reject the technology. They reject saving money on the service they need.

That makes reason three the most uncomfortable one. 53 percent name misuse of personal data as their top concern when companies automate interactions with AI. That is eight points more than the year before. Half fear that AI will keep them from ever reaching a human being (Source: Qualtrics, 2025).

Isabelle Zdatny of the Qualtrics XM Institute puts it in one sentence: “Too many companies are deploying AI to cut costs, not solve problems, and customers can tell the difference.”

The same study also shows the reverse direction: trust can be repaired. 46 percent would share more data if it were transparent what is being collected. 45 percent each would do so with better control over usage and deletion. Transparency is not a nice-to-have. It is the lever.

Because trust is the real currency of customer service. And trust is not built where everything runs smoothly. It is built where something went wrong and someone made it right.

Will AI replace customer service?

No. Gartner expects that by 2028, not a single Fortune 500 company will have fully eliminated human customer service.

Reason four: the pioneers are already rowing back. The clearest story comes from Klarna. February 2024: Klarna launches its AI assistant. 2.3 million conversations in the first month, by the company's own account the work of roughly 700 full-time agents. The workforce shrinks from 5,500 to 3,400, mostly through a long hiring freeze. The company publicly celebrates this as the model of the future (Source: Fast Company, 2026). May 2025: CEO Sebastian Siemiatkowski publicly admits the company cut too deep. Klarna starts hiring humans again and has been rebuilding ever since: AI takes the routine volume, humans take escalations, complex cases, and the relationships that matter. The breaking point was not money. It was quality: edge cases, emotional situations, multi-step problems.

Klarna is not an outlier. It is the start of a wave. Gartner predicts that by 2027, half of the companies that attributed headcount reductions to AI will rehire people for similar roles (Source: Gartner, 2026).

And maybe the most surprising number from that same survey of 321 service leaders: only 20 percent have actually reduced staff because of AI at all. The headlines tell a different story than the data.

By 2028, Gartner adds, not a single Fortune 500 company will have fully eliminated human service (Source: Gartner, 2025).

Gartner analyst Emily Potosky explains why: AI simply is not mature enough to replace the expertise, empathy, and judgment of human agents. Relying on AI alone right now is premature.

Whoever buys “more AI in customer service” as a strategy today is buying the model the pioneers are currently unwinding.

Why the same logic applies to your team, and what the Upwork data on human-AI collaboration shows, is covered in “Why AI fails without humans”. Short version: pure AI agents fail regularly. The combination of human and AI lifts project completion rates by up to 70 percent.

Why do more tools not fix the architecture problem?

The fifth reason has nothing to do with AI at all. It is about the foundation underneath. In practice, “more AI” usually means: one more tool on a stack that is already fragmented. The Puzzel State of Contact Centres 2026 report shows that only 3 percent of contact centers run on a single, unified platform. The average organization juggles 3.9 different technologies (Source: CMSWire / Puzzel, 2026).

But an AI is only as good as the context it can see. If order history, ticket system, and email threads live in three separate systems, even the best model can only guess. That is not an AI problem. That is an architecture problem. And another tool makes it bigger, not smaller.

This is where a small business has a paradoxical advantage: fewer systems, shorter paths, a manageable data landscape. The preconditions for connecting an AI solution cleanly are often better with 10 people than with 10,000.

How can AI be used well in customer service?

By automating more precisely instead of more broadly: AI answers the factual instantly, humans resolve the emotional properly, and escalation is built as a feature.

What if the question was never “how much AI” but “where AI”? The companies that do this well automate more precisely. The dose makes the difference, like with medicine: the right amount heals. Too much harms. You can try out where the sweet spot sits right here:

Four steps help you hit that dose.

Step 1: Draw the line before you look at any tool

Factual requests with one clear answer are perfect for AI: “Where is my order?”, “When are you open?”, moving an appointment. Around the clock, instant, no hold music. Emotional, trust-critical moments belong to a human, no detour: the complaint, the hesitation before a purchase, the case where something went wrong. I've described how to draw this line systematically, with a three-zone map for green, red, and gray tasks, in “Where AI creates room and humans stay human”.

Step 2: Build escalation as a feature, not an emergency exit

The “I want a human” path is not a defeat for the AI. It is part of the design: always visible, one click, and the human receives the full conversation history. Customers never repeat themselves. Remember the 82 percent: escalation happens anyway. The only question is whether it feels like service or like a penalty lap.

Step 3: Measure resolution quality, not deflection

Many systems measure the “deflection rate”: how many requests the bot kept away from humans. That measures brushing off, not helping. Better questions: Was the issue actually resolved? Did the customer come back? 60 percent of people want the right answer, not the fastest one.

Step 4: Be transparent, generously transparent

Say clearly when AI is answering. Say what happens with the data. The Qualtrics numbers show that transparency is exactly the lever that rebuilds trust. Whoever hides it confirms the suspicion that the service is being cut.

And here is the punchline for your business: while corporations are rowing back, you can draw the line correctly from the start. Your advantage is not having more technology. Your advantage is that a human still picks up when it matters. A well-dosed AI solution protects exactly that advantage: it clears away the factual and keeps the human time free for the moments customers never forget.

The right dose is your opportunity

More AI in customer service is not the answer. The right dose is. AI for the factual. Humans for what builds trust. And an architecture that connects both cleanly.

What I bring from my work is a conviction: the goal is never to remove humans from the conversation. The goal is to give them back the conversations that are worth having.

If you want to know where you stand: take 15 minutes this week. Look at your last 20 service requests. Mark them: Which were purely factual? Which needed a human? The ratio you see there is your dose. Not the one a tool vendor sells you.

The next two years will show who treated customer service as a cost center and who treated it as a relationship. And if you would rather not determine the dose alone, we can find it together in coaching.