AI Customer Service Agents in 2026: What They Resolve, Where They Fail and How to Deploy Them

Most AI customer service projects are sold on one number: the share of conversations the agent resolves without a person. That number is defined, measured and billed by the vendor, and it rarely tells you whether the customer got what they needed or gave up and phoned. The rest of the case rests on less visible things: which contacts the agent can safely handle, what it is allowed to do with money and accounts, what the law now requires of it, and what happens to the contacts it cannot close. This guide covers what named deployments and independent studies show, how vendors count a resolution, where agents fail, which rules apply, what a contact costs and how to build an agent that holds up.

• 64%of more than 5 million conversations on Salesforce's own help site resolved autonomously, down from over 84% a year earlier²
• 87%of customers say a company using AI for service must offer a way to reach a person¹⁶
• 24%of service and support leaders demonstrated positive financial returns from their AI use cases¹⁵

What AI agents resolve in production

The best-known figures are large and self-reported. Klarna said its OpenAI-powered assistant had 2.3 million conversations in its first month, two-thirds of its service chats, doing "the equivalent work of 700 full-time agents."¹⁷ Its registration statement gives the more durable version: the assistant handled 69% of service chats in the twelve months to June 30, 2025, and delivered about $39 million of cost savings in 2024. The 700 figure was "estimated based on the average monthly reduction in chat and telephone conversations handled by full-time agents," so it is a modeled equivalent, not a count of people.¹⁴ Intercom says its Fin agent averages 76% resolution across more than 12,000 customers.¹⁸ Microsoft says Air India's agent answers 40,000 queries a day with a 97% success rate.¹⁹

Salesforce's series is the most useful in the public record because it tracks one deployment over time. The rate rose to over 84% in the first months, when the agent handled about 32,000 conversations a week, then fell to 70% of 4.3 million inquiries and 64% of more than 5 million as the volume, the range of questions and the stakes of counting grew.²¹¹²²² A stricter definition is the likely reason more than a fall in quality. Either way, a vendor's flagship figure moved twenty points under its own hand, and the first months of any rollout are when most announcements are written.

The cost side moved too. Klarna's customer service and operations expense fell $37 million, or 15%, in 2024.¹⁴ By the third quarter of 2025 it said the agent did the work of 853 employees and saved $60 million, while the same quarter's customer service and operations cost rose to $50 million from $42 million a year earlier.²³ In May 2025 its chief executive told Bloomberg that "cost unfortunately seems to have been a too predominant evaluation factor," and that the result was "lower quality," and Klarna began recruiting human agents again.²⁴

What vendors count as a resolution

The definitions explain why vendor figures cannot be compared. Intercom counts a resolution when the customer confirms the answer worked or "exits the conversation without requesting further assistance," which it calls an assumed resolution, and it deducts the charge only if the customer comes back to the same conversation.³ A customer who goes quiet for 24 hours after Fin's last answer counts as resolved.²⁵ Decagon concedes that the common test, no human ticket within 24 to 72 hours, "can overstate deflection when a customer gives up and contacts an untagged channel later," and that "an operation can post high deflection while resolution quietly declines."²⁶

Two vendors have put the gap in writing. Zendesk introduced resolution tiers on May 18, 2026: a conversation the agent finished without the customer asking for more is a "contained resolution" unless a model check confirms the request was actually resolved, which makes it a "verified resolution."²⁷ Ada's figures across more than 550 deployments put average containment at about 72% and automated resolution at about 52%.⁴ Decagon's own ranges are lower than most showcase numbers: 55% to 75% deflection in e-commerce and 40% to 60% in SaaS, with rollouts that "commonly start at 30% to 40%."²⁶

No vendor publishes a resolution rate adjusted for customers who contact the company again on another channel. A customer who abandons a chat and calls the next morning can be counted as resolved, and the vendor that sets the definition is usually the vendor paid per resolution. Ask who measured a figure, over which window and which kinds of contact, whether silence counts, whether confirmed and assumed resolutions are reported apart, whether recontact is tracked on every channel, and whether it comes from the launch month or a steady state. Klarna's often-quoted 25% drop in repeat inquiries shows why the window matters: its filing dates the measurement to December 2023 and January 2024, before the global launch.¹⁴

What independent research shows

The strongest independent evidence is about a different product, AI that drafts and suggests for human agents. A study of 5,179 support agents found the assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, "with minimal impact on experienced and highly skilled workers."⁵ A 2026 field experiment in Alibaba's service operation found that AI assistance raised customer ratings but had "no significant effect on objective service quality, measured by customer retrials," and that top-performing agents got worse on both.²⁸ No peer-reviewed study yet compares an autonomous agent's resolution or recontact rate with a human control group in production.

Customers set conditions. In Gartner's survey of 3,566 customers in early 2026, 50% said GenAI makes service easier and 87% said a company using it must offer a way to reach a person.¹⁶ Only 7% used a chatbot in their most recent service interaction, and only 27% would try one again after a bad experience.²⁹ Use of company chatbots "has remained statistically unchanged since 2022," while use of third-party AI tools nearly doubled in a year.¹⁵ Qualtrics found nearly one in five consumers who used AI for service saw no benefit, a failure rate almost four times that of AI used for other tasks.³⁰

The workforce effect so far is mostly redeployment. In an October 2025 survey of 321 service leaders, only 20% had reduced headcount because of AI, and 55% held staffing flat while handling more volume.³¹ The largest public cut is Salesforce, where Marc Benioff said on a podcast that support went "from 9,000 heads to about 5,000."³² Gartner predicts that by 2027, half the companies that attributed headcount cuts to AI will rehire for similar work under different titles.³³

Where support agents break

Public benchmarks are the only independent, repeatable view of autonomous support agents, and they agree on the weak spots. On Sierra's tau-bench, GPT-4o completed 61.2% of retail tasks on a single try, but succeeded on all eight repeated tries of the same task less than 25% of the time.⁶ That consistency measure is the one that matters, because a deployed agent meets the same request thousands of times. When the customer also has to act, as in telecom troubleshooting, GPT-4.1's score fell from 74% in retail to 34%.³⁴ When answers sit across about 700 interlinked banking documents, frontier models reached about 25.5%.³⁵ Voice agents reached 31% to 51% in clean audio and 26% to 38% with noise and varied accents, against 85% for a text model on the same tasks.⁷

Policy written in a prompt does not hold against a persistent customer. Salesforce's CRMArena-Pro found leading agents at about 58% success in single-turn tasks and about 35% in multi-turn ones, with "near-zero inherent confidentiality awareness" that prompting improved only at a cost to task success.³⁶ IBM researchers' CRAFT method talked an airline agent out of its policy 70% of the time, against 35% for a generic jailbreak, and within four attempts its success rate passed 80% even against the strongest prompt defense.³⁷ Agents also find routes the policy's author never pictured: Anthropic reports Claude Opus 4.5 getting around a no-changes rule on basic economy by upgrading the cabin first and then changing the flight.³⁸ A rule that matters belongs in the system the agent calls, where it cannot be argued with.

The public failures

Two lessons sit under the table. The first is legal: the tribunal in the Air Canada case held the airline responsible for information on its own website "whether the information comes from a static page or a chatbot," and New York's disclaimers did not make wrong answers right.⁸⁴² The second is a shift from tone to permissions. An agent with access to records and tools is a new entry point to your systems, and the damage in the 2025 cases came from what the agent could reach, so the security review matters as much as the script.

Rules that apply to customer-facing agents

For a company serving customers in the EU and California, the practical floor for 2027 is to label the agent as AI at first contact, keep a staffed route to a person with bounded waits, and keep a person on refusals that matter, such as refunds, credit and account closures. Gartner's forecast that regulation will push assisted volume up 30% by 2028 means the human tier is a permanent cost to plan for.⁵⁵

What a contact costs

The model itself is a small part of these prices. OpenAI's realtime audio is billed per token, at one token per 100 milliseconds of caller audio and one per 50 milliseconds of agent audio, and a text conversation on a small model costs a fraction of a cent.⁶⁰ The rest of a list price pays for retrieval, orchestration, evaluation, integration and margin.

On the human side, the median US customer service representative earned $21.53 an hour, across 2.67 million jobs that BLS projects to decline 5% over the next decade.¹³ At Klarna's 12-minute average for a human chat, that is about $4.30 in wages before benefits, management, tools and idle time.¹⁴ Per contact, AI wins today. The catch is what the meter leaves out: knowledge upkeep, integrations, testing, and the human queue for the contacts the agent cannot close. Service leaders put a median 12% of their 2025 budgets into AI, and only 24% demonstrated positive financial returns.¹⁵ Gartner expects the cost per AI resolution to exceed $3 by 2030, "higher than many B2C offshore human agents."⁵⁵

How to build a support agent that holds up

Identity and data scope come first. Authenticate the customer, then let the agent read only that customer's records, because models do not reliably refuse confidential requests on their own.³⁶ Money and irreversible actions belong in code. OpenAI's guidance is to rate each tool low, medium or high risk by write access, reversibility, permissions and "financial impact," and to use the rating to trigger checks or a person's approval.⁶¹ The retail policy in tau2-bench shows the pattern: authenticate before acting, list the details of any change and get an explicit yes, and refund only to the original payment method.⁶² Every customer message, form field and document is possible injected instruction, and every model output is untrusted until it has been encoded and checked.⁴⁵⁴⁴

Knowledge is the biggest lever. Intercom's published work found that fine-tuning its retrieval model on its own support queries lifted English retrieval precision from 44.79% to 74.33% on its test set, a change in the search layer with the language model left alone.⁶³ Writing procedures as structure helps too: a 2026 benchmark found that driving the agent from a graph of standard operating steps let GPT-4o-mini beat GPT-4o on policy adherence.⁶⁴ Voice needs its own decision. People expect a reply in about 500 milliseconds, while a typical cascaded voice pipeline takes around 1.3 seconds.⁶⁵ OpenAI notes that a chained pipeline lets you "inspect or transform intermediate text," which is where policy checks before a refund or an identity step can sit.⁶⁶

1. 1Pick contacts by volume and risk Group six months of contacts by reason and start with high-volume, low-risk requests such as order status, returns within policy and password resets.
2. 2Fix the knowledge first Rewrite the articles and policies the agent will use, remove contradictions, and write down the answers your best people give that exist nowhere else.
3. 3Put money and identity rules in the tools Authenticate before any account action, scope data to that customer, and enforce refund limits, eligibility and confirmation inside each tool.
4. 4Test repeatedly on your own cases Build a suite of real and adversarial conversations, run each many times, and release only versions that pass every run on the risky requests.
5. 5Label the agent and keep a person close Tell customers at first contact that they are talking to AI, and give them a one-step route to a person with a bounded wait.
6. 6Measure resolution the hard way Report confirmed and assumed resolutions separately and track recontact on every channel within seven days.
7. 7Re-test after every change Rerun the suite after any model, prompt, knowledge or vendor update before it reaches customers.

AI agents already handle a real share of customer contacts, and the gains are largest on simple, repeated requests and for newer staff. The resolution rate on a dashboard is a commercial term set by the party that bills on it. The companies that will look good in two years are the ones that can show their recontact rate, their policy violations and their time to a person, because those are what customers, regulators and tribunals now check.

This is how we approach AI agent development for customer service: start with the requests that are safe to automate, put identity, refund and data rules in the systems the agent calls, and measure every release on your own conversations before customers meet it. The policies, test suite and logs stay with you, so you can change models or vendors without starting again.