Will AI replace customer service? What it can’t do

No, not all of it. AI already answers the repeat questions your docs cover, in seconds and at any hour, and at Klarna, AI handled 80% of customer service chats in 2025. People still handle what needs judgement, money, a look inside someone’s account or empathy. Even Klarna still offers every customer a person.

Deacon is an AI customer support agent built on that split, for founders who answer support themselves. It answers your customers in seconds from your own content. When the content runs out, it says it doesn’t know instead of guessing and asks for the visitor’s email. Once they leave it, you get an email, and the reply you type appears in their chat if they still have it open. It also shows you what customers are stuck on and what they want next. There’s a free plan, and it needs no card.

Every figure and quote below links to where it was first published, such as a research firm’s own release, a company’s own filing or a government page, and each was read there on 25 September 2026.

What the companies and the research say

Four sources answer most of the question between them.

What AI already answers well

AI is good at questions whose answer is already written down, such as how to export your data, whether a plan includes SSO, where to change a card or what an error message means. Those are the questions that repeat, they arrive at all hours, and answering one for the hundredth time teaches you nothing.

The biggest public example is Klarna, the payments company. It launched its AI assistant in February 2024, and a month later Klarna said it had held 2.3 million conversations, two-thirds of its customer service chats. In its annual report for 2025, filed with the US Securities and Exchange Commission on 26 February 2026, Klarna says the assistant handled 80% of its customer service chats that year, did the work of more than 850 full-time agents and saved about $59 million. It resolved chats in two minutes on average, against 12 minutes for Klarna’s human agents in 2024, and Klarna’s own satisfaction surveys rate it on par with them. It answers around the clock, in more than 35 languages. Chatbot examples follows its first month in more detail.

Then read the next part of that report. “Appreciating that certain consumers may nevertheless prefer to interact with human representatives, we continue to offer all of our customers that option,” Klarna writes, calling it a “dual-track approach” of AI at scale alongside “high-quality human support”. Klarna’s assistant is also wired into customers’ accounts, showing balances and handling refunds and returns, which takes engineering most founders can’t spare. The part that carries over to a small product is the repeat questions.

Customers are warming to it. In Gartner’s survey of 3,566 customers in February and March 2026, 50% said their interactions are easier when companies use generative AI. And the service leaders Gartner surveyed in autumn 2025 were doing more with the same people, as 55% reported stable staffing while handling more customers.

Gartner expects AI to take on more. It predicts that by 2029 agentic AI, meaning AI that acts for a customer, such as cancelling a membership, rather than only answering, will resolve 80% of common customer service issues without a person. AI agent vs chatbot explains the difference.

What these have in common is a routine question with a known answer. Gartner’s Kathy Ross put it plainly in September 2025. “While AI excels at handling routine and well-defined problems, it often struggles with exceptions and high-risk scenarios.”

What still needs a person

AI handles a question badly when there’s no written answer to find, or when the answer is a decision. Five kinds of conversation keep needing a person.

• Decisions about money. A refund, a discount or an exception is your call, not a lookup. Air Canada’s website chatbot told a customer whose grandmother had died that they could apply for a bereavement fare after booking, which the airline didn’t allow. On 14 February 2024 a British Columbia tribunal ordered Air Canada to pay C$650.88 in damages, plus interest and fees, and wrote that “it should be obvious to Air Canada that it is responsible for all the information on its website”. Why support bots make things up explains how that happens.
• Anything inside a customer’s own account. Gartner found that customers increasingly start with outside AI tools such as ChatGPT, but “still come to the company when they need to transact, access account-specific information or resolve more complex issues”. Unless an AI is connected to your billing and checks who’s asking, it can’t see why someone’s card failed.
• Advice. In a Gartner survey of 5,801 US customers in January and February 2025, 54% said they trust human agents more than AI for product or service recommendations, against 32% who trust AI more. A prospect asking whether your product fits their team wants that kind of judgement.
• Someone who is already upset. Only 27% of customers would try a chatbot again after a bad experience, Gartner found in its 2026 survey. A customer writing angry, or writing for the second time, should hear from you.
• Anyone who asks for a person. In the same survey, 87% said a way to reach a human agent is essential when a company uses generative AI for service. When Gartner asked 5,728 customers in December 2023, their top worry about AI in service was that it would get harder to reach a person.

Two companies that leaned too hard on AI

Cursor, the AI code editor, used AI replies as the first filter for its email support. In April 2025 a thread titled “Cursor IDE support hallucinates lockout policy” reached Hacker News, and a Cursor co-founder replied there that “something very clearly went wrong here”. He said AI replies in email support were now clearly labelled as AI and that the customer had been refunded. Asked whether every Cursor inbox sat behind AI, he said its support system did. “We are standing up a much bigger team here but are behind where we should be.”

Commonwealth Bank of Australia said in July 2025 that 45 customer service jobs would go after it brought in an AI voice bot, which it said had cut call volumes. In August it reversed the cuts. The Finance Sector Union, which had taken the bank to the Fair Work Commission, said calls had in fact been rising, that management admitted it hadn’t properly considered that the increase would continue, and that the bank would apologise to the 45 staff.

Gartner doesn’t expect even the largest companies to go all the way. In September 2025 it predicted that by 2028 none of the Fortune 500 will have fully eliminated human customer service.

Where the handover sits

The design that works is AI first, with a person one step away. Gartner calls it “digital first, but not digital only”, and its analyst Eric Keller says chatbots “should serve as seamless connectors to human support, not containment traps”. Human in the loop covers the other places a person can sit.

Five rules make that work, whichever tool you use.

• Answer only from your own content, so every answer comes from something you wrote.
• Say “I don’t know” the moment the content runs out, instead of guessing.
• Ask for one way to reply, such as an email address, so the customer can leave and still hear back.
• Tell a person straight away and pass on the whole conversation, so nobody repeats themselves.
• Make it plain what’s AI and what’s a person, which is the lesson Cursor learned in public.

How Deacon hands a conversation to you

Deacon is built on those rules. It answers in seconds from your website, your files, text you paste in and the answers you write, replying in whatever language the visitor writes in. When that content doesn’t cover a question, it says it doesn’t know instead of guessing and asks for the visitor’s email. It always declines questions about the visitor’s own account, invoice or data, because it can’t see them, and it takes no actions in other systems, so it can’t issue a refund or change a plan.

The owners on your team get an email 3 minutes after a visitor leaves an address or asks for a person, unless someone on the team already has the dashboard open or has opened that conversation. The conversation shows every turn and the page the visitor was on. Type your reply in the conversation, and a visitor with the chat still open sees it under your first name and your company’s, while the box they type in changes from Ask AI to Reply to the team. Sending it takes the conversation from Deacon until you press Hand back to Deacon. A visitor who has left gets your reply by email, if they gave an address in that conversation.

From the first message it’s plain that visitors are talking to an AI. The default greeting introduces Deacon as your AI assistant, and the box visitors type in reads Ask AI on every plan.

What it means for a founder who answers support alone

You’re the person one step away, so the real question is how much of your day goes on questions your docs already answer. Let AI take those. Keep the ones that need you, such as refunds, billing disputes, bug reports, a prospect asking about a missing feature and anyone who’s upset or asks for you. Customer service automation goes through the rest of a customer conversation stage by stage, and the best AI agents for customer service compares the tools.

The questions an AI hands you are worth more than they look, because each one is a gap in your content. The share of conversations handed to you is your escalation rate, and it falls as you fill those gaps.

In Deacon, every question your content didn’t cover has an Answer this button. Write the answer once and press Save & check, and Deacon asks itself the question again, shows you the reply it now gives and tells you how many other open questions your answer covers. Every answer you write becomes part of what Deacon knows, so the more you answer, the less it needs you. Topics group similar questions and count the people who asked, so you see what customers are stuck on and what they want you to build next.

If you’re weighing up a person instead, customer support outsourcing covers what freelancers and agencies charge. If you hired someone yourself in the US, the Bureau of Labor Statistics puts the median wage for a customer service representative at $21.53 an hour in May 2025.

What it means for a job in customer support

Fewer people will answer the easy questions, and the job moves towards the hard ones. The official projections and the surveys of service leaders agree on that.

The US Bureau of Labor Statistics counted 2,666,000 customer service representative jobs in 2025 and projects a 5% fall by 2035, about 141,800 fewer. It gives automation as the reason, since self-service, social media and apps let customers do simple tasks without a representative. It still expects about 289,500 openings a year, all from people retiring or moving to other work, and says some companies will keep their own service centres, “particularly for complex inquiries such as refunding accounts or confirming insurance coverage”. The page was last updated on 27 August 2026.

Employers are cutting more slowly than the headlines suggest. In Gartner’s survey of 321 service leaders in September and October 2025, only 20% had reduced agent staffing because of AI, and 42% were hiring for new roles such as conversational AI designers and automation analysts. The same survey found 85% adding tasks to frontline roles and 75% moving agents into new roles. Just 31% had made or were planning layoffs because of AI through the first quarter of 2027, and 63% were letting frontline headcount fall gradually as people leave.

Gartner also predicts that by 2027 half the companies that cut customer service staff because of AI will rehire people for similar work under different job titles. Its researcher Emily Potosky said AI “simply isn’t mature enough to fully replace the expertise, empathy, and judgment that human agents provide”. Nor will AI stay cheap. Gartner expects an AI resolution to cost more than $3 by 2030, more than many offshore agents at consumer companies. It also expects rules giving customers the right to talk to a human to raise the number of conversations people handle by 30% by 2028.

For you, that means the part of the job that repeats an answer is going. What stays and grows is judgement, the conversations that need a decision or some care, and running the AI itself, which means writing the content it answers from, reading its conversations and filling the gaps it finds.

Questions about AI replacing customer service

Will AI replace customer service jobs?

Some of them. The US Bureau of Labor Statistics projects customer service representative jobs to fall 5% between 2025 and 2035, while still expecting about 289,500 openings a year. Only 20% of the service leaders Gartner surveyed in late 2025 had cut staff because of AI, and Gartner predicts none of the Fortune 500 will have fully eliminated human customer service by 2028. The jobs that remain lean towards harder conversations and towards running the AI.

How is AI used in customer service today?

In four main ways, according to Gartner’s survey of 265 service leaders in April and May 2025. It helps agents with summaries and quick answers, answers customers directly through self-service chat and search, automates back-office work such as quality assurance and writing help content, and, newest, acts on requests through agentic AI. For a small team, the one that matters first is answering customers from your own docs. Conversational AI for customer service follows one support chat turn by turn.

Do customers mind talking to AI?

They mind being stuck with it. Half the customers Gartner surveyed in early 2026 said generative AI makes service easier, but 87% said a way to reach a person is essential, and only 27% would try a chatbot again after a bad experience. Answer fast from your own content, and keep the way to you easy to find.

Is it safe to let AI answer my customers?

It is when the AI answers only from your content and says when it doesn’t know. The risk is a bot that guesses, since the tribunal held Air Canada to what its chatbot said. Try it on the questions your customers really ask before it goes live, and read its conversations afterwards. With Deacon, the Playground lets you try it before a customer does, and every conversation is saved and searchable by what visitors asked.

Can Deacon handle refunds or account changes?

No. Deacon takes no actions in other systems, so it won’t issue a refund, look up an order or change a plan. It answers general questions about them from your docs, such as how to update a card. It always declines questions about the visitor’s own account, invoice or data and asks for their email so you can reply. Your site can pass details such as the visitor’s plan, and Deacon takes them into account when it answers.

Does Deacon answer phone calls or support email?

No. It answers in a chat on your website. It has no phone or voice agent, no WhatsApp, Instagram or SMS channel, no inbox for support email and no ticketing. It also never starts a conversation, because the chat opens only when a visitor clicks it or your site calls Deacon.open().

Does Deacon tell visitors they’re talking to AI?

Yes, by default. Its greeting introduces it as your AI assistant, and the box visitors type in reads Ask AI on every plan. You can write your own greeting, so keep it saying so. When you reply, your words appear under your first name and your company’s, without the Deacon mark.

What does Deacon cost?

Every feature is on every plan, the free one included. Free gives you 50 answers a month with no card and no end date. Starter is $50 a month, or $40 a month billed yearly, for 1,000 answers a month. Growth is $150 a month, or $120 a month billed yearly, for 4,000 answers a month. Nothing is billed past a plan’s answers, and no plan upgrades itself. Pricing has the full plan card.