Why customer service is hardest to make AI-native
Every other function hands AI a clean on-ramp. Finance has a stable, low-stakes close to automate. Sales has research and sequencing to hand off. Customer service has one queue, and that queue carries a password reset and a grieving customer's refund dispute through the exact same door. That is the whole problem, and no model solves it.
One service queue carries a routine question and a legally binding promise through the same door. The redesign that works starts by fixing triage.
Short answer: Customer service is the hardest function to make AI-native because a single queue carries routine lookups alongside emotional, high-stakes, and legally binding moments. The fix is a redesigned triage layer that decides which interactions an agent ever sees, sends the routine to AI, and routes everything else to a person.
We rebuild service workflows for operators, and the ones that fail almost always automated the queue before anyone redesigned it.
Why is customer service harder to make AI-native than sales or operations?
Because in service, the routine work and the judgment work arrive through the same channel, so an agent cannot tell which is which until it has already answered.
In finance or operations, the routine and the exceptions run on separate tracks. A reconciliation is a reconciliation; the M&A call goes somewhere else entirely. Customer service has no such split. "Where is my order" and "your bot told my mother the wrong bereavement fare and now I want a refund" land in the same inbox, in the same tone, often in the same sentence. Gartner projects that agentic AI will autonomously resolve 80% of common customer service issues by 2029 and cut operating costs by 30%. The load-bearing word is "common." The uncommon interactions are the ones your brand actually lives or dies on, and they show up unlabeled.
The cost of getting the sort wrong is brutal and immediate. Zendesk's CX Trends 2026, which surveyed more than 11,000 people across 22 countries, found that 85% of CX leaders say a customer will leave a brand over a single unresolved issue, even on the first contact. Misroute one high-stakes moment into an automated dead end and you do not get a second attempt. Sales can afford a bad email. Service cannot afford a bad answer to the wrong person.
Why did Klarna bring back human agents after going AI-first?
Klarna ran the most aggressive AI-first service experiment in the market, then reversed part of it in public, and the reversal is the more useful data point.
The early numbers were real. Klarna's OpenAI-built assistant handled two-thirds of its customer service chats in the first month, 2.3 million of them, and cut average resolution time from 11 minutes to under 2. Then came the part nobody puts in a launch post. By May 2025, Klarna was hiring human agents again, and CEO Sebastian Siemiatkowski said the company's AI-first push had gone too far, with quality the casualty. The technology worked on the two-thirds it was built for. It broke on the third that needed a person, and that third was expensive enough to walk the strategy back.
The market read this and adjusted. Three months after its 80%-by-2029 forecast, Gartner released a second prediction: 50% of organizations that planned to shrink their service workforce because of AI will abandon those plans. Both forecasts are from the same firm, months apart. Neither is wrong. They describe the two halves of the queue.
Can a company be held liable for what its AI chatbot says?
Yes. A tribunal has already ruled that the company owns what its bot tells a customer, and the argument that the bot is a separate entity did not survive contact with a judge.
The case is Moffatt v. Air Canada, decided by British Columbia's Civil Resolution Tribunal in February 2024. Air Canada's chatbot gave a grieving passenger inaccurate advice about bereavement fares, the airline refused to honor it, and the tribunal held the airline liable and rejected the argument that the chatbot was a separate legal entity responsible for its own words. If your bot says it, your company said it.
The failure modes are not always so tidy. A DPD chatbot, prodded by a frustrated customer, swore and called its own employer "the worst delivery firm in the world"; DPD disabled it within a day. A Chevrolet dealership's bot was prompt-injected into agreeing to sell a $76,000 Tahoe for a dollar, and told to call the deal "legally binding". Both were shut down. This is the exposure that makes service different: the same open text box that resolves a shipping query is also an unmonitored channel for anyone who wants to make your brand say something, promise something, or defame itself. We wrote about where that accountability lands in who owns the mistake when your AI agent gets it wrong.
Agent-assist or fully autonomous - which one actually works?
Give the human the last call and both speed and satisfaction go up. Take the human out of the loop and speed still rises, but customers rate the experience lower. The evidence on this is now field-tested at scale.
Two large experiments at Alibaba's Taobao make the split clear. An agent-assist model, where human agents kept discretion over the AI's suggestions, improved both resolution speed and customer satisfaction. A fully agentic model, where AI resolved cases autonomously and humans only supervised and handled escalations, ran across 647 workers and 680,676 chats in August 2024; it cut chat duration but lowered customer ratings. Same company, same customers, two designs. The one that kept a person in the decision won on the metric that matters.
None of this argues against scale. Salesforce's State of Service report, its seventh edition and built on more than 6,500 service professionals, found AI resolved 30% of service cases in 2025 and projects 50% by 2027, with AI leaping from the tenth priority for service leaders to the second in a single year and 79% now calling AI agents essential. The volume is going to AI. The question is which volume, and where the handoff sits. The handoff itself is fragile: Zendesk found that 74% of consumers find it frustrating to repeat themselves across agents and channels, so a clumsy escalation from bot to human can undo the resolution it was meant to save.
Redesigning a service function around agents? Sign up for the newsletter to get the next piece.
What does the redesign that works actually look like?
Redesign the triage layer first. The service functions that make AI work rebuild the thing that decides which interaction an agent ever sees, then build everything else around that decision. Here is the sequence we run.
- Sort by stakes. The old routing rules sent tickets by subject line: billing here, shipping there. Rebuild the front of the queue to classify each interaction by consequence first. Is this reversible? Is it emotional? Does it commit the company to money or a legal position? The answer to that decides who handles it, and the topic barely matters.
- Give AI the routine, in full. The order-status checks, the password resets, the return labels, the policy lookups. This is Klarna's genuine two-thirds and Salesforce's 30% heading to 50%. Let the agent own it end to end, with no human tax on work a person adds nothing to.
- Route the consequential to a person, by design. Anything emotional, high-value, or legally binding gets a human before the company commits to anything. Air Canada is the price of skipping this step.
- Keep a human on the escalation seam. The Taobao result is that autonomous resolution without human discretion costs you satisfaction. So design the handoff from bot to person as one moment that carries full context. A cold restart forces the customer to repeat the story that Zendesk says 74% already resent repeating.
- Define what the bot may never say or promise. A hard boundary on commitments, quotes, and legal language, enforced before the text ever reaches the customer. Chevrolet's dollar Tahoe is what an undefined boundary costs.
- Measure resolution and satisfaction together. Deflection alone rewards exactly the failure mode Klarna reversed: a bot that closes tickets and loses customers. Track two things: was the interaction actually resolved, and would the customer come back.
The through-line is that the intelligence belongs in the triage as much as in the answers. Most programs pour the budget into a smarter bot and leave a dumb queue in front of it, which is a version of the mistake we covered in why adding AI to an existing workflow makes it worse.
What does the exec actually decide?
The AI-native service model comes down to four calls, and they belong to the CEO or the function head.
- Where does the stakes line sit? Which interactions are safe for full autonomy, and which touch money, law, or emotion. Draw it too wide and you inherit an Air Canada. Draw it too narrow and you automate nothing worth automating.
- What happens at the handoff? The seam between bot and human is where satisfaction leaks. Fund the context transfer, or the escalation undoes the resolution.
- What can the bot never commit to? The hard boundaries on promises and legal language, defined before launch. Chevrolet drew them after the incident, once the bot had already agreed to the dollar Tahoe.
- What do you measure? Track resolution and retention together, and stop reporting deflection on its own. The metric you pick is the behavior you get.
Answer those four and you have a service function that scales AI without betting the brand on it. Skip them and you have a fast bot in front of a queue that was the problem all along. For the wider version of this split across every function, see the AI-native operating model, function by function.
Frequently asked questions
Will AI replace customer service jobs? Partially. AI is taking the routine volume, and Salesforce reports it resolved 30% of cases in 2025, projected to reach 50% by 2027. But Gartner predicts 50% of organizations will abandon plans to shrink their service workforce because of AI. The high-stakes and emotional interactions still need a person, so the role narrows.
Why did Klarna reverse its AI-first customer service strategy? Because quality suffered on the interactions that needed a human. Klarna's assistant handled two-thirds of chats and cut resolution time from 11 minutes to under 2, but by May 2025 the company was rehiring human agents and its CEO said the push had gone too far. The AI worked on the routine third it was built for and struggled on the rest.
Can a company be held legally liable for what its AI chatbot says? Yes. In Moffatt v. Air Canada, a British Columbia tribunal held the airline liable for its chatbot's inaccurate bereavement-fare advice and rejected the argument that the bot was a separate legal entity. If your bot makes a representation to a customer, the company is treated as having made it.
What is the difference between AI agent-assist and a fully autonomous AI agent? Agent-assist keeps a human in control, with the AI suggesting and the person deciding. A fully autonomous agent resolves cases on its own, with humans only supervising. At Taobao, the agent-assist model raised both speed and satisfaction, while the fully autonomous model cut chat duration but lowered customer ratings.
Why is customer service harder to automate than sales or operations? Because one queue mixes routine lookups with emotional, high-value, and legally binding moments, and they arrive unlabeled. Other functions separate routine work from judgment work onto different tracks. Service routes both through the same channel, so an agent cannot tell which kind it is handling until it has already answered.
We write about what it takes to run an AI-native organization. No tool reviews, no hype. The operating decisions a CEO, CRO, or COO actually has to make. Join the newsletter to get the next piece.
Sources
- Gartner, "Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues by 2029" (March 2025). gartner.com
- Gartner, "50% of Organizations Will Abandon Plans to Reduce Customer Service Workforce Due to AI" (June 2025). gartner.com
- Klarna, "Klarna AI assistant handles two-thirds of customer service chats in its first month" (2024). klarna.com
- Fortune, "Klarna is rehiring human agents after its AI-first customer service push" (May 2025). fortune.com
- Salesforce, "State of Service Report," 7th edition (6,500+ service professionals). salesforce.com
- Zendesk, "CX Trends 2026" (11,000+ respondents, 22 countries). cxtrends.zendesk.com
- Moffatt v. Air Canada, 2024 BCCRT 149 (February 2024), reported by CBC and analyzed by McCarthy Tétrault. cbc.ca / mccarthy.ca
- Field experiment on agent-assist customer service, arXiv (2026). arxiv.org
- Field experiment on fully autonomous customer service (647 workers, 680,676 chats, August 2024), arXiv (2026). arxiv.org
- Time, "A DPD chatbot swore and criticized the company" (January 2024). time.com
- Envive, "Case study: the Chevy dealership AI chatbot" (December 2023). envive.ai
By Christopher Kliebenstein. We build and run AI-native workflows for operators who want real results.