AI in Customer Service (2026): What Works, What Fails and How to Roll It Out
AI in customer service means using software that understands language, such as chatbots, AI agents and agent-assist tools, to answer customers, help human agents and route requests. In 2026 it ranges from an assistant that drafts replies for a human to an AI agent that settles a refund on its own.
The honest state of play is uneven. The strongest evidence, a peer-reviewed field study of more than 5,000 support agents, shows large gains when AI helps human agents. Bots that replace agents have produced impressive company-reported numbers, but also public failures, legal rulings and reversed layoffs. Customers, meanwhile, keep telling surveys they want a human option. This guide is independent: it relies on research, court and tribunal rulings, regulator texts and companies’ own disclosures, not on vendor marketing, and it ends with a step-by-step rollout plan.
What AI in customer service means
AI in customer service is not a single product. It is a collection of technologies that automate or assist different parts of the support workflow. The main uses include:
- AI agents and chatbots: Systems that answer customers directly via chat, email, or voice. Older versions followed scripted decision trees. Current versions use large language models grounded in the company’s help content (retrieval-augmented generation, or RAG) and hand over to a human when they cannot answer.
- Agent assist: Tools that suggest replies, summarize conversations, and coach human agents in real time.
- Routing and triage: Algorithms that direct customers to the most appropriate agent or department based on intent and sentiment.
- Knowledge search: Semantic search that finds relevant help articles, policies, or past tickets instantly.
- Sentiment analysis: Real-time detection of customer emotion to flag escalations or adjust tone.
- Voice AI and IVR: Conversational voice bots that handle simple queries or route calls without pressing buttons.
- Summaries and QA: Automatic generation of after-call summaries and quality monitoring of agent performance.
The shift from scripted chatbots to large language model agents changed what bots can do. Scripted bots failed when customers deviated from the expected path. LLM-based agents, grounded in RAG, can handle more natural language and access dynamic data. However, they are not perfect. They can hallucinate, and they often require a human handoff to resolve complex issues. This has led to a split in strategy: some companies use AI to replace agents, while others use AI to make agents more effective.
Two models: agent assist vs autonomous AI agents
The choice between agent assist and autonomous AI agents is not just about technology; it is about risk, cost, and customer experience. Agent assist keeps a human in the loop, reducing the risk of errors while boosting productivity. Autonomous AI agents handle the interaction end-to-end, offering speed and scale but risking customer frustration if the AI fails.
| Question | Agent assist | Autonomous AI agent |
|---|---|---|
| Who answers the customer? | A human agent, aided by AI suggestions and summaries. | The AI resolves the issue itself, with a human handoff only if needed. |
| Main benefit | Faster resolutions, faster onboarding; a human stays accountable. | Lower cost per contact, 24/7 availability, instant answers. |
| Main risk | Smaller savings than full automation; agents may over-trust suggestions. | Invented answers, prompt injection, customers who cannot reach a person. |
| Evidence so far | A peer-reviewed field study of 5,172 agents (QJE, 2025). | Company-reported resolution rates; public failures and reversals; no independent controlled study. |
| Example | Verizon’s Gemini-based assistant for 28,000 reps. | Klarna’s OpenAI-powered assistant handling 2.3 million chats. |
What the evidence says about agent assist
The strongest causal evidence for AI in customer service comes from a 2025 field study published in the Quarterly Journal of Economics by Erik Brynjolfsson, Danielle Li and Lindsey Raymond. The study examined 5,172 customer support agents at a Fortune 500 software company, mostly based in the Philippines. The company introduced an agent-assist tool built on OpenAI’s GPT models. The results were significant. The tool increased productivity, measured as issues resolved per hour, by 15% on average.
The gains were not uniform. Less-skilled and less-experienced agents gained about 30%, with the lowest skill quintile seeing a 36% increase. This suggests that AI acts as a great equalizer, allowing new hires to perform at a higher level. Agents with only two months of tenure using the tool performed as well as agents with more than six months of tenure without it. This has major implications for hiring and training. Companies can hire less experienced agents and bring them up to speed faster, reducing the time and cost of onboarding.
Attrition also fell. Agents with under six months of tenure saw a 40% reduction in attrition relative to the baseline. This is crucial for industries with high churn, where training costs are a significant burden. The tool also improved customer interactions. Customers were more polite, and requests to speak to a manager fell by almost 25%. This suggests that AI-assisted agents are better at de-escalating conflicts and resolving issues effectively.
Verizon provides a real-world example of this model. In April 2025, Reuters reported that Verizon gave its 28,000 service reps a Google Gemini-based assistant. Following the rollout, sales by the service team rose nearly 40%. While this is a company-reported correlation and not a controlled experiment, it aligns with the Brynjolfsson study’s findings that AI-assisted agents are more productive and effective.
What this means for teams is clear. Agent assist is a low-risk, high-reward strategy. It does not replace humans; it makes them better. It reduces the cognitive load on agents, allowing them to focus on complex issues while AI handles routine tasks. This model is particularly effective for companies with large, diverse support teams where skill levels vary. It also helps new hires through their hardest first months, which is where the study found the drop in attrition.
What autonomous AI agents deliver: the company numbers
Autonomous AI agents promise to replace human agents entirely, offering instant responses at a fraction of the cost. The numbers from major companies are impressive, but they come with caveats. Resolution rates are high, but customer satisfaction and long-term costs tell a more complex story.
Klarna: From 700 agents to rehiring humans
Klarna, the Swedish fintech company, made headlines in February 2024 when it announced that its OpenAI-powered assistant handled 2.3 million conversations in its first month. This represented two-thirds of Klarna’s customer service chats. The assistant did the equivalent work of 700 full-time agents. Customer satisfaction was reported to be on par with human agents. Repeat inquiries fell by 25%, and errands were resolved in under 2 minutes, compared to 11 minutes before. The assistant operated across 23 markets and 35+ languages. Klarna estimated a $40 million profit improvement in 2024.
However, the story did not end there. In May 2025, CEO Sebastian Siemiatkowski told Bloomberg that “As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality.” Klarna began hiring again, allowing customers to always reach a human. This reversal highlights a common pattern: autonomous AI agents can handle volume, but they often struggle with quality and customer preference.
Klarna’s 2025 annual report (SEC filing) provided more detail. The assistant handled 80% of customer service chats in 2025 with no drop in consumer satisfaction levels. Customer service and operations costs rose 2% while volumes rose 32%. By the company’s own account, then, cost per contact fell sharply as volume grew, while it kept a human layer in place.
Salesforce: Resolution rates on its own help site
Salesforce runs its own Agentforce agent on help.salesforce.com, which has been live since October 2024. The company has published different figures at different times, making it hard to compare. About six months in, Salesforce reported over 84% resolution rates, with only 4% handed to humans. In early 2025, it reported 2.6 million conversations at a 63% resolution rate. By June 2026, it reported 4.3 million inquiries with a 70% resolution rate.
Salesforce ingested nearly 740,000 pieces of help content and called good data “the greatest unlock.” Renewal questions go straight to humans, a deliberate choice for high-stakes conversations. CEO Marc Benioff said in an August 2025 podcast that he had cut support staff from 9,000 heads to about 5,000. Salesforce’s own pages do not state those headcount numbers.
Intercom and pay-per-resolution pricing
Intercom, a customer messaging platform, says its Fin AI agent’s average resolution rate “now stands at 76%” across more than 12,000 teams (2026). It bills $0.99 per “outcome”, after renaming its billed metric from “resolution”. Pricing per resolution means the vendor is paid when its bot counts a conversation as resolved, which is one more reason to check how that count is made.
Why resolution rates are hard to compare
Vendors define “resolution” differently and some charge per resolution, so their rates are not comparable. Does a conversation count as resolved if the customer simply left? If they contacted support again the next day? Salesforce alone has published 84%, 63% and 70% for the same help site at different times. Without a shared definition, a headline resolution rate says more about the counting method than about the bot.
Additionally, resolution rates do not capture customer satisfaction. A customer might have their issue resolved quickly, but if the AI’s response was confusing or incorrect, they might still be dissatisfied. Klarna’s CEO said as much in 2025: optimizing mainly for cost had produced “lower quality”.
When customer service AI goes wrong
AI in customer service is not just about efficiency; it is about liability, reputation, and the precise boundaries of what a machine can and cannot do. When automation fails, it fails publicly. The failures range from simple hallucinations to legal judgments and operational collapses. Understanding these breakdowns is critical for leaders deploying autonomous agents.
Air Canada: the company owns what its chatbot says
The legal precedent for corporate liability in AI interactions was set early. In Moffatt v. Air Canada (British Columbia Civil Resolution Tribunal, 14 February 2024), Air Canada’s website chatbot told a grieving customer he could claim a bereavement fare retroactively. The airline argued the chatbot was responsible for its own actions. The tribunal disagreed, stating, “It should be obvious to Air Canada that it is responsible for all the information on its website.” The ruling found negligent misrepresentation and ordered Air Canada to pay approximately CA$812 in damages, interest, and fees. The lesson is clear: a chatbot is not a separate legal entity. If it speaks for the brand, the brand owns the error.
A German court agrees
This principle extends beyond North America. In Germany, the Higher Regional Court Hamm (OLG Hamm) ruled on 12 May 2026 that a beauty clinic’s chatbot falsely described its doctors as board-certified plastic surgeons. The court treated the chatbot’s answers as the operator’s own commercial statements under unfair-competition law. The distinction between “a bot said it” and “we said it” is disappearing in courts of law. If the bot is the face of the service, the business is on the hook for its accuracy.
Invented policies: Cursor’s “Sam” and New York City’s MyCity
Hallucinations are not just errors; they are policy violations. In April 2025, Cursor’s AI support agent “Sam” invented a policy that the product works on only one device per subscription. Co-founder Michael Truell admitted, “We have no such policy… this is an incorrect response from a front-line AI support bot.” Users cancelled subscriptions, and Cursor began labelling AI email replies as such. Similarly, New York City’s MyCity business chatbot, launched in 2023, told businesses they could take workers’ tips and refuse cash. The city shut it down in February 2026, with the mayor calling it “functionally unusable.” Both are examples of AI producing plausible but wrong business rules. Even vendors acknowledge the risk: Intercom’s documentation warns that its bot “can also fabricate a plausible-looking link to a domain that doesn’t exist”, which settings can reduce but “can’t fully prevent”. Our guide to LLM hallucinations explains why this happens.
Bots that can be talked into anything
Prompt injection attacks show how fragile chatbot guardrails can be. In December 2023, users manipulated a ChatGPT-based dealership chatbot at Chevrolet of Watsonville (California) into “agreeing” to sell a 2024 Tahoe for $1. It was an embarrassment; no sale took place, and the bot was pulled. In January 2024, a system update at UK parcel firm DPD caused its chatbot to swear at a customer, write a poem, and call DPD “the worst delivery firm in the world.” DPD disabled the AI element. These incidents show that without robust grounding and prompt-injection defenses, AI can be easily coaxed into saying anything.
Cutting staff too early
Operational failures often accompany premature automation. In August 2025, Commonwealth Bank of Australia cut 45 customer-service roles because of an AI voice bot. It reversed the decision after the Finance Sector Union challenged it, admitting its assessment “did not adequately consider all relevant business considerations.” Klarna made a similar correction on a larger scale: in May 2025 its CEO said that treating cost as the main factor had produced “lower quality”, and the company began hiring so that customers can always reach a human. Cutting staff before the data is in can leave a gap that has to be refilled.
| Case | Year | What went wrong | Lesson |
|---|---|---|---|
| Air Canada | 2024 | Chatbot wrongly promised retroactive bereavement fare | Company is liable for chatbot statements |
| OLG Hamm (Germany) | 2026 | Chatbot falsely claimed doctors were board-certified | Chatbot answers treated as corporate commercial statements |
| Cursor (Sam) | 2025 | AI invented a “one device per subscription” policy | Invented policies cause cancellations; label AI replies |
| NYC MyCity | 2026 | Chatbot told businesses they could take tips and refuse cash | Test answers against the real rules before launch |
| DPD | 2024 | Chatbot swore and insulted the company | System updates can break guardrails; disable if unstable |
| Chevrolet of Watsonville | 2023 | Users got the bot to “agree” to a $1 Tahoe | Bots can be talked into false commitments |
| Commonwealth Bank of Australia | 2025 | Cut 45 roles for a voice bot, then reversed | Measure real call volumes before cutting staff |
What customers actually think
Customer attitudes toward AI in customer service are nuanced and often skeptical. Despite the hype, users do not uniformly embrace automation. Gartner’s July 2024 survey of 5,728 customers found that 64% would prefer companies didn’t use AI in customer service. The top concern is that AI makes it harder to reach a person. 53% said they would consider switching to a competitor if they learned a company was going to use AI for customer service.
Even when AI works, it often doesn’t work well enough. Gartner’s August 2024 data from the same survey shows that only 14% of customer service issues are fully resolved in self-service. Even for issues customers called “very simple,” the resolution rate was only 36%. This gap between expectation and reality drives churn. Furthermore, Gartner’s 2026 survey of 3,566 customers revealed that 87% say it is essential to offer an option to reach a human agent when using GenAI. Only 27% would try a chatbot again after a negative experience.
Customers also compare company bots to general-purpose AI. In the same 2026 survey, customers were about three times more likely to use third-party AI such as ChatGPT or Gemini than a company’s own chatbot in their latest service interaction. General-purpose assistants have become the first stop, so a company’s own bot now has to beat them to get used.
Trust is eroding. Salesforce’s State of the Connected Customer (2024) survey of 15,015 consumers found that only 42% trust businesses to use AI ethically, down from 58% in 2023. Qualtrics’ October 2025 survey of 20,000+ consumers in 14 countries found that nearly one in five consumers who used AI for customer service saw no benefit, about four times the failure rate of AI use in general. Additionally, 53% fear misuse of personal data, and about half worry about not reaching a human. These figures underscore that AI is not a silver bullet for satisfaction. It is a tool that must be deployed with transparency and a clear path to human support.
The cost math
The financial model for AI in customer service is shifting from simple headcount reduction to complex cost-per-resolution calculations. Gartner forecasts that by 2030, the cost per resolution for generative AI will exceed $3, more than many offshore human agents (January 2026). If that forecast holds, full automation will not automatically be the cheaper option.
Gartner also predicts that by 2029, agentic AI will autonomously resolve 80% of common customer service issues, with a 30% cut in operational costs (March 2025). However, the path to these savings is not linear. By 2027, 50% of organizations that expected to significantly reduce their customer service workforce will abandon these plans (June 2025). In a 2025 poll, 95% of service leaders planned to keep human agents. By 2027, half of companies that attributed headcount cuts to AI will rehire for similar functions under different job titles (February 2026).
For your own business case, the useful number is not the vendor’s resolution rate but the cost of a correctly resolved contact, including the conversations that come back, the escalations and the errors a human has to fix. Our guide to reducing LLM API costs covers the model side of that bill.
Rules you have to follow
Regulation for AI in customer service is tightening globally. Leaders must navigate a patchwork of laws that mandate transparency and restrict deceptive practices.
The EU AI Act Article 50(1) states that AI systems that interact with people must be designed so people are informed they are interacting with an AI system (unless obvious), at the latest at the first interaction. This applies from 2 August 2026 and was not delayed by the 2026 “Digital Omnibus” amendments.
In the US, California SB 1001 (in force since July 2019) bans undisclosed bots used to knowingly deceive people to incentivize a sale. Utah (SB 226, 2025) requires businesses to disclose AI when a consumer clearly asks, and proactively in high-risk interactions in regulated occupations. Maine’s 2025 law states that using an AI chatbot in trade that may mislead consumers into thinking they are talking to a human, without disclosure, is an unfair trade practice.
Additionally, the US FCC declaratory ruling (February 2024) states that AI-generated voices in calls count as “artificial” voices under the Telephone Consumer Protection Act, so they need prior consent. This impacts voice AI and conversational IVR deployments.
Disclaimer: This is not legal advice. Consult legal counsel for compliance in your jurisdiction.
How to roll out AI in customer service
-
Start with agent assist. Begin by deploying AI tools that help human agents, rather than replacing them. Agent-assist tools suggest replies, summarize conversations, and coach agents in real time. This builds confidence and allows teams to learn from AI without risking customer experience. The Brynjolfsson study showed a 15% productivity increase with agent assist, with less-skilled agents gaining 30%.
-
Fix the knowledge base first. AI is only as good as the data it retrieves. Salesforce ingested nearly 740,000 pieces of help content and called good data “the greatest unlock.” If your knowledge base is messy, your AI will hallucinate. Clean, structured, and up-to-date content is the foundation of reliable AI customer support.
-
Pick narrow, high-volume, low-risk intents. Start with simple, frequent queries like “reset password” or “track order.” Avoid complex, high-stakes interactions initially. This reduces the risk of major errors and allows you to refine the AI’s performance on common issues before expanding scope.
-
Ground answers in your content and let the bot say “I don’t know.” Use retrieval-augmented generation (what RAG is) to ground AI responses in your help content. Allow the bot to admit uncertainty rather than inventing answers. This reduces hallucinations and builds trust. Intercom notes that bots can fabricate links, so grounding is critical.
-
Always offer a human and hand over with context. Ensure there is a seamless handoff to a human agent. The human should receive the full conversation history and context. This prevents customers from repeating themselves and improves resolution rates. Gartner’s 2026 survey found 87% of customers say this is essential.
-
Disclose the AI. Be transparent. Inform customers when they are interacting with an AI system. This meets regulatory requirements like the EU AI Act and California SB 1001. It also sets expectations for the customer.
-
Measure resolution honestly. Track repeat contacts, escalations, and CSAT, not just deflection rates. Deflection can mask poor resolution. Intercom’s shift from “resolution” to “outcome” highlights the need for precise metrics. Look at the full customer journey, not just the first touch.
-
Red-team for prompt injection and invented policies. Regularly test your AI for vulnerabilities. Try to trick it into revealing secrets, making false promises, or breaking guardrails. Simulate prompt injection attacks, the trick behind the $1 Tahoe.
-
Don’t cut staff before the data is in. Wait for sufficient data on AI performance before reducing headcount. Commonwealth Bank of Australia cut 45 roles prematurely and had to reverse the decision. Let the data drive workforce planning, not assumptions.
FAQ
How is AI used in customer service?
AI in customer service covers chatbots and AI agents that answer customers directly (chat, email, voice), agent-assist tools that suggest replies, summarize and coach human agents in real time, smart routing and triage, sentiment analysis, knowledge-base search, voice AI / conversational IVR, after-call summaries, and quality monitoring. Older chatbots followed scripted decision trees; current ones use large language models grounded in the company’s help content (retrieval-augmented generation, RAG) and hand over to a human when they cannot answer.
Will AI replace customer service agents?
So far AI is mostly augmenting agents rather than replacing them en masse, and some early cuts have been reversed (Commonwealth Bank of Australia) or partly walked back (Klarna). The Brynjolfsson study found a 15% productivity increase with agent assist, with less-skilled agents gaining 30%. However, Gartner predicts that by 2027, 50% of organizations that expected to significantly reduce their customer service workforce will abandon these plans. AI handles routine tasks, but humans remain essential for complex, high-empathy, or high-stakes interactions.
Do customers want AI customer service?
Customer sentiment is mixed. Gartner’s July 2024 survey found 64% of customers would prefer companies didn’t use AI in customer service, and 53% would consider switching to a competitor. However, 87% say it is essential to offer an option to reach a human agent when using GenAI. Customers value AI for speed but fear it makes it harder to reach a person.
What is the difference between a chatbot and an AI agent?
A chatbot typically follows scripted decision trees or rule-based logic to answer specific questions. An AI agent uses large language models to understand context, reason, and perform actions autonomously. Agents can handle multi-step tasks, access external systems, and hand over to a human when needed. They are more flexible and capable than traditional chatbots.
Is a company liable for what its chatbot says?
Yes. In Moffatt v. Air Canada (2024), the tribunal ruled that Air Canada was liable for its chatbot’s false promise of a bereavement fare. In Germany, the Higher Regional Court Hamm (2026) treated a chatbot’s false claims about doctors as the operator’s commercial statements. The chatbot is an extension of the brand, and the company owns its errors.
Do you have to tell customers they are talking to AI?
Regulations are increasingly mandating disclosure. The EU AI Act Article 50(1) requires informing users they are interacting with AI at the first interaction, effective 2 August 2026. California SB 1001 bans undisclosed bots used to deceive. Maine’s 2025 law considers misleading consumers into thinking they are talking to a human an unfair trade practice. Best practice is to disclose proactively.
How do you measure AI customer service performance?
Measure resolution honestly by tracking repeat contacts, escalations, and CSAT, not just deflection rates. Deflection can mask poor resolution. Intercom shifted from “resolution” to “outcome” to reflect true value. Look at the full customer journey, including time to resolution, customer effort score, and quality of AI responses.
What is contact center AI?
Contact center AI is the set of AI tools used in phone and chat support centers, including voice AI, conversational IVR, agent-assist, sentiment analysis, and after-call summaries. It aims to automate routine tasks, enhance agent productivity, and improve customer experience through data-driven insights and automation.