Summary
AI won’t replace customer service; it takes the repeat questions your docs already answer.
People keep what isn’t written down, anything that needs an action in your systems, and upset customers.
Forrester forecasts 49% of current customer service jobs to be gone by 2030; Gartner found 20% of leaders had cut agents due to AI.
No, AI won’t replace customer service on a team like yours. It replaces the part of the job where someone types the same answer for the tenth time.
Wonderchat’s study of 5,848,078 conversations measured where it stops, and its AI agent resolved 75.3% of scored conversations on its own in 2026.
In a random sample of 5,000 conversations it couldn’t resolve, 58.5% asked for an answer that had never been documented anywhere the agent could reach. Judgment calls, actions in your systems, and upset customers still need a person.
What do Gartner, Forrester, and the BLS predict for customer service jobs?
The forecasts about what AI will do to customer service come from serious firms, and each one counts something different, whether issues, jobs, or interactions. Surveys so far count leaders who cut staff and respondents who saw a decline, and the BLS projects US jobs, so read the unit before you compare.
What do the forecasts say AI will take over?
Gartner predicted in March 2025 that by 2029, AI agents will resolve 80% of common customer service issues “without human intervention.” The 80% counts common issues, so it says nothing yet about how many jobs go.
Forrester’s Kate Leggett and Laura Ramos go further on jobs. Their May 2026 forecast says that by 2030, AI will cause 49% of current customer service jobs to disappear, though lower-volume teams, typically B2B, will see AI resolve a smaller share of first contacts.
Forrester’s Max Ball explained the reasoning to CX Dive. “There are humans today doing jobs that don’t require the level of intelligence that a human has,” he said, adding, “That work is going to go away, so the number of agents is going to go down, and over time the automation is going to get more and more sophisticated.”
Early signs exist too. Stanford’s Digital Economy Lab finds that early-career customer service workers “show substantial employment declines.”
What do surveys and the BLS show about support jobs so far?
The records so far measure something else. In Gartner’s survey of 321 service leaders in October 2025, “only 20% of customer service leaders have actually reduced agent staffing due to AI.”
Three analyst forecasts, two survey results, and one official US projection, each measuring something different. Sources: Gartner, Forrester, McKinsey, and the US Bureau of Labor Statistics, each on its own page, checked October 21, 2026.
The same release says Gartner predicts that by 2027, half of the companies that cut customer service staff because of AI will rehire for similar work under different job titles. Gartner’s Emily Potosky put it plainly. “AI simply isn’t mature enough to fully replace the expertise, empathy, and judgment that human agents provide.”
Gartner’s April 2026 release adds that 85% of service leaders are expanding human agents’ responsibilities, while 31% have implemented, or are planning, layoffs in response to AI through the first quarter of 2027.
Across all functions, McKinsey’s 2026 survey found that 39% of respondents expect AI to decrease head count over the next year, so the pressure hasn’t gone away.
The US Bureau of Labor Statistics projects a 5% decline in customer service representative jobs from 2025 to 2035, with about 289,500 openings a year as people leave the job.
How much of customer service can AI resolve today?
On real docs, Wonderchat’s AI customer service agent resolves about three in four conversations on its own and stops most often where nothing is written down. In the study, the AI agent resolved 75.3% of scored conversations end to end in 2026, without a person.
The rate isn’t fixed. Among 687 businesses with enough volume to compare fairly, resolution ranged from under 46% in the weakest tenth to more than 88% in the strongest on the same software, with a median of 69%.
The reason the rest went unresolved is the useful part. In a random sample of 5,000 conversations the agent didn’t resolve, 58.5% asked for an answer that had never been documented anywhere it could reach.
The typical case is a question about a nonprofit discount that no page mentions. No setting fixes it, and one written page does.
Among 5,000 unresolved conversations sampled at random, answers never documented outnumber all other causes of a miss combined.
The rest of that sample split between questions outside what the team had set the agent to cover, questions too ambiguous to answer as asked, content that had gone out of date, and conversations with no clear cause.
Across 84 businesses that added missing documentation, resolution moved from 75.4% to 80.8% over three months, a tendency rather than a promise, since 57% improved and 23% declined.
The repeat questions an agent handles are the ones tier 1 support answers today, questions with a known, written answer.
The quarter it doesn’t resolve still needs people.
Which customer service work stays human, and why?
An AI agent’s limit is what’s written down and what it’s allowed to do. Everything past that line is people’s work, and a good setup sends it there on purpose.
Which questions should the AI agent hand to a person?
At Wonderchat, the agent hands a conversation to a person on three rules, a direct request for a person, a message count, or repeated failed answers, and the full conversation goes with it. The moment of escalation and handoff decides whether the customer has to repeat everything.
Any one of the three rules hands the conversation to a person, with the whole chat attached.
When no page covers a question, the agent says it can’t find that, doesn’t guess, and offers a person. A clear fallback message keeps that miss honest, and the agent is just as plain about a request it can’t act on, like a request to reissue an invoice with a new VAT number.
Asked about a status page no page covers, the agent says it can’t find one and offers a live agent.
Actions stay with people too. The agent doesn’t process refunds, cancellations, or order and account changes.
In 5,000 scored conversations that Wonderchat’s study classified by stage, the agent resolved 73.6% for people still deciding to buy and 60.8% for existing customers with a problem to fix. Part of that gap is the routing working, because judgment calls, policy exceptions, and account access belong with a colleague.
What do customers expect when AI answers first?
Customers want an easier answer and a way out. In Gartner’s 2026 survey of 3,566 customers, half said GenAI makes their interactions easier, and 87% said an option to reach a human agent is essential.
Gartner’s Eric Keller is specific about the handover. “If the chatbot cannot resolve an issue with high confidence, it should provide a visible path to a human agent and pass along the information and context already collected.”
Businesses buying AI support worry about the same moment. Vera Sun and Luke Tan, Wonderchat’s co-founders, write in the study’s foreword that “the concern we hear most is trust,” because buyers “worry the agent will get something wrong in front of their customer.”
For the EU AI Act’s disclosure rule, the study’s checklist starts with “Disclose at the first message, in the language of the conversation, rather than in terms nobody opens.” Its next step is “Make the human path visible at any point, and name the conversation types that route automatically.”
Octopus Energy draws its own line on what AI handles, in a press release that says its AI “never handles vulnerable customers, sensitive cases or complex complaints, which always go straight to experienced members of the customer service team.”
What do AI agents and human agents each do best?
When you compare AI agents vs human agents, the old AI chatbot vs human agent framing of speed against empathy misses where the line falls. The real split is a written answer against an unwritten one, and an answer against an action.
A refund request shows the split. The agent can quote your refund policy with a link, and only a person can issue the refund.
The work | AI agent | Human agent |
|---|---|---|
Questions with a written answer | Answers from your docs and shows the source | Answers them as well |
Actions in your systems, like refunds or account changes | Answers from your policy page, if you have one, and hands over | Does the action |
When it doesn’t know | Says it can’t find the answer and offers a person | Asks a colleague |
A customer who asks for a person | Hands over with the full chat | Takes it from there |
How it gets better | Answers each new question from then on, once you add the answer | Training and experience |
What happened when companies handed support work to AI?
Shep Hyken, a customer service author and speaker, writes on his blog that “The goal shouldn’t be to remove humans from customer support.” Klarna, Commonwealth Bank, Air Canada, and Cursor show what happened when companies moved support work to AI, and that a wrong AI answer is still the company’s answer.
What did Klarna and Commonwealth Bank say after moving support work to AI?
In March 2024, Klarna’s CEO told CBS that its AI assistant was doing the equivalent work of about 700 full-time agents, and that Klarna’s human agents are outsourced.
What three companies and one tribunal said after AI took on support work, February 2024 to August 2025.
By June 2025, he told an audience at SXSW London that human customer service would stay, as a premium option, while AI takes the manual work.
In July 2025, Commonwealth Bank of Australia declared 45 customer service roles redundant after introducing a voice bot. On August 21, it reversed the decision and apologized, the one full reversal among these cases.
What happened when Air Canada’s and Cursor’s AI gave wrong answers?
Air Canada’s website chatbot gave a customer wrong information about bereavement fares. In February 2024, British Columbia’s Civil Resolution Tribunal held the airline responsible for what its website said and ordered it to pay $812.02 CAD in total.
In April 2025, an AI support reply at Cursor described a policy that didn’t exist. The Register reported that Cursor’s co-founder apologized, the developer who raised it was refunded, and AI email replies were labeled from then on.
On Wonderchat’s own support team, the failure seen most often is an agent pulling an archived policy and answering from it correctly.
Should you hire a support person or use an AI agent first?
If most of your questions already have a written answer, start with the agent and hire for the rest. Either way, decide first what never goes to the AI.
Decide the escalation list before launch. One step in Wonderchat’s study reads “Name the conversation types that should reach a colleague every time,” listing “refund disputes, policy exceptions, account changes, anything carrying a complaint.”
The same step goes on: “Configure them as routing rules before launch, so the decision sits with your team rather than emerging from how the agent behaves.”
Which questions should the AI agent take first?
Start the agent on the questions your docs already answer, and keep everything else with a person until you’ve written it down.
What the question needs | Who takes it first | Why |
|---|---|---|
An answer already in your docs | The AI agent | It answers from the page and shows the source |
An answer nobody has written yet | A person, who then writes it once | The agent answers it from then on |
An exception or a judgment call | A person | Gartner’s Kathy Ross says AI “often struggles with exceptions and high-risk scenarios” |
A refund in dispute or an upset customer | A person, once the customer asks or the agent can’t answer | They go on the escalation list first |
When is it time to hire a support person?
Hire when the rows that need a person fill someone’s week. Until then, whoever answers customers today covers the rows the table gives to people, plus the gaps the agent reports.
Try the agent on your own help center before the job ad goes up, and see plans when you need more than the free plan.
How does the first support hire’s job change with an AI agent?
With an agent in place, the first hire stops typing the same answers and owns what the agent can’t answer. Forrester’s analysts and two people quoted in Wonderchat’s study describe what that takes.
What work moves to the first support hire?
Forrester’s Leggett and Ramos see AI already replacing coaching and scheduling jobs in contact centers, and write in the same forecast that when this happens, “the human workforce’s core mandate shifts from reactively interacting with customers directly to managing the AI that interacts with them.”
Shmuel Saklad, a senior manager for enterprise accounts at B&H Photo, says in the study that hiring changes too. “You need people who can use judgment and not just follow a script.”
He added that the job also needs “people with higher emotional intelligence, who are better equipped at connecting with others and understanding their frustrations.”
How does the hire keep the AI agent’s answers current?
Every question the agent couldn’t answer lands in one review inbox, sorted into user reported, knowledge gaps, quick fixes, conflicts, and model issues. The hire pastes a URL, picks a file, or writes the answer, and the agent answers it from then on.
One pasted page answers a missed question, with its source, from then on.
For a question that needs one approved answer, a correction is an approved question and answer the agent uses next time.
“A correction (a small one-off FAQ) is usually good for extremely specific questions or edge cases,” says Luke Tan, CTO and co-founder of Wonderchat.
Ian Golding, a CCXP-certified customer experience specialist, makes the case for an owner in the same study. A customer can ask the same unanswered question hundreds of times, he says, but “unless there is a mechanism that turns that signal into an action, with an owner and a date, it stays an insight rather than an improvement.”
Service leaders see the shift coming. In a Gartner survey of service leaders released in February 2026, 58% said they aim to upskill agents into knowledge management specialists.
To set it up, start with the help center pages behind your ten most-asked questions.
When is an AI agent that acts in your systems the better choice?
Ada, Decagon, or Sierra is the better choice when most of your conversations are account changes, refunds, or other tasks done inside your own systems.
The study’s industry table shows where that work sits. Among the 49% of businesses whose industry could be inferred, Education & E-learning resolves 70.8% of its conversations and E-commerce & Retail 60.1%.
The report traces the gap to live data. A question about an order or a refund needs the record itself, which no help center page holds.
Ada describes “AI customer service agents that resolve, act, and continuously improve,” Decagon lets teams “define agent workflows in natural language,” and Sierra invites teams to “break complex outcomes into specific steps.”
They sell to a different buyer, often with a deployment team that helps set them up. Wonderchat answers from your docs, hands over, and reports, and it doesn’t take actions in your systems, so a refund still ends with a person on your team.
How Wonderchat splits the work between AI and your team
AI won’t replace your customer service. Wonderchat, an AI customer service agent and AI helpdesk in one, takes the repeat questions and leaves your first hire the rest.
A customer asks how to change their invoice email, and the agent answers from your billing article with a link. The next asks about a nonprofit discount no page covers, so the agent says so and hands the whole chat to a person.
The handover lands in Wonderchat’s help desk or yours, and the question waits in the review inbox until your team writes the answer once.
The handover reaches Wonderchat’s help desk or the one your team already runs, with the whole chat attached.
The free plan refills 50 AI replies every month, with no credit card. Start for free, paste your help center URL, and ask your hardest question, then one your docs can’t answer, to see the source and the honest miss, or book a demo.
Frequently asked questions
Is AI taking over customer service jobs?
AI is taking over part of the work faster than the jobs. In Gartner’s October 2025 survey, 20% of service leaders had reduced agent staffing because of AI. The BLS projects a 5% decline in US customer service jobs from 2025 to 2035, with about 289,500 openings a year.
Do customers prefer AI or human customer service?
Customers want an easier answer and a way to reach a person. In Gartner’s 2026 survey of 3,566 customers, half said GenAI makes service easier, and 87% said a way to reach a human agent is essential when a company uses GenAI.
How can you tell if customer service is AI?
You can tell when a company labels its AI. Cursor labeled AI email replies after its 2025 support mix-up, and Octopus Energy labels every AI-written email. In the EU, the AI Act has required since August 2, 2026, that people be told they’re talking to AI, unless it’s obvious.
What are the disadvantages of AI in customer service?
An AI agent answers from what it has been given, so gaps in your docs become gaps in its answers. It can also be wrong with confidence, and the company still owns that answer, as Air Canada’s 2024 tribunal ruling showed.
Can an AI agent handle customer support without hiring more people?
For questions with a written answer in your docs, yes, with the source on each answer. Actions in your systems, judgment calls, and upset customers still need a person, so hire once those fill someone’s week. That person then works the handovers and the gaps the agent reports.
