AI customer service is the use of language models to resolve support requests, draft agent replies, route incoming tickets, and read conversations for patterns. That definition is easy. The hard part is that almost every article about it cites the same handful of vendor statistics, and almost none of them check whether the companies named actually published a number.
So I went and checked. I threw out every claim I could only find on a third-party blog, kept only what the companies said in their own newsrooms, investor decks and price cards, and rebuilt this list around that. Four companies fell out on the way, including two that show up on nearly every roundup of this topic.
I have spent the last three-plus years at eesel putting AI agents on live support queues, which mostly means watching where they break. The lesson that survives every rollout is the one that shows up below in Klarna's numbers and then in Klarna's U-turn: a confident-sounding bot that quietly gives wrong answers costs more than no bot at all. That is why every eesel rollout now gets replayed against a customer's historical tickets before it answers a single real person.
What AI customer service actually does
Modern support AI is not one product, it is four distinct jobs, and conflating them is how buyers end up disappointed.
- Resolve. The AI answers the customer and closes the request without a human touching it. This is what automated resolution means, and it is the only job that moves your containment rate.
- Assist. The AI drafts a reply for a human agent to check and send. An AI copilot lowers average handle time without ever reducing ticket count.
- Triage. The AI reads intent, then tags and routes. Cheap, low-risk, and the usual place to start with ticket triage.
- Analyse. The AI clusters conversations to show you which missing help article is generating tickets, which is how self-service actually improves.
The real shift from the scripted bots of a decade ago is that a language model reads intent instead of matching keywords, so it degrades gracefully on phrasings nobody scripted. That difference is the whole gap between an AI agent and a rule-based chatbot.
What AI customer service costs
This is the most-asked question about the category and the most badly answered, because vendors bill in units that look similar and are not. Zendesk bills per resolution, Freshworks per session, eesel per ticket. A chatty customer generates several sessions inside one ticket.
Here are the three published rate cards run against the same monthly volume. Pick your volume.
Monthly cost at published rates
Same volume, three billing units. The unit is doing more work here than the price.
| Rate card | Monthly |
|---|---|
| Zendesk, pay-as-you-go $2.00 per automated resolution | $1,000 |
| Zendesk, committed volume $1.50 per automated resolution | $750 |
| Freshdesk Freddy AI Agent $0.49 per session, $49 per 100 | $245 |
| eesel $0.40 per ticket, no platform fee | $200 |
| Rate card | Monthly |
|---|---|
| Zendesk, pay-as-you-go $2.00 per automated resolution | $4,000 |
| Zendesk, committed volume $1.50 per automated resolution | $3,000 |
| Freshdesk Freddy AI Agent $0.49 per session, $49 per 100 | $980 |
| eesel $0.40 per ticket, no platform fee | $800 |
| Rate card | Monthly |
|---|---|
| Zendesk, pay-as-you-go $2.00 per automated resolution | $20,000 |
| Zendesk, committed volume $1.50 per automated resolution | $15,000 |
| Freshdesk Freddy AI Agent $0.49 per session, $49 per 100 | $4,900 |
| eesel $0.40 per ticket, no platform fee | $4,000 |
Rates as published on each vendor's own pricing page, August 2026. Seat licences, Copilot add-ons and platform fees are excluded, so these are AI line items only, not total cost of ownership. Freshworks also charges $29 per agent per month for Freddy AI Copilot.
Two things worth noticing. A session is not a resolution, so Freshdesk's cheaper-looking unit can bill several times inside one conversation that Zendesk would count once. And Zendesk's own comparison table is where those resolution rates live, not the plan cards, which is why so many Zendesk pricing write-ups miss them entirely.
The same normalising problem shows up in Freddy AI pricing. It shows up again in per-interaction models like Gorgias AI pricing, where the meter runs on messages rather than outcomes.
Whichever unit you land on, the number that decides whether it was worth it is cost per resolved contact, which means you need to compare it against your agent versus AI cost baseline before signing.
The containment numbers companies actually published
Most articles on this topic quote a single vendor's marketing figure. Four companies have put real numbers in public, and the spread between them is the most useful thing in this post.
Read that chart as a scope chart, not a quality chart. Air India's 97% is extraordinary, and it is measured on a query assistant answering baggage, booking and loyalty questions. Bell's 50% covers a broader technical-support surface where the alternative is a truck roll. Neither figure is audited, and neither one is a promise about your queue. What they do establish is a credible band, roughly half to two-thirds for general support, higher when the scope is narrow and the knowledge is clean.
If you want to hold a vendor to a number, hold them to containment alongside escalation quality, not deflection rate alone. A ticket abandoned by a frustrated customer deflects beautifully and resolves nothing.
9 companies using AI for customer service
Every entry below follows the same shape: what they run, the number they published, and the part worth stealing. Where a company had no first-party number, it did not make the list.
1. Klarna
The most-cited AI support deployment in the world, and the most instructive because it has a second act.
What they run. An OpenAI-powered assistant in the Klarna app handling refunds, returns, payment issues, cancellations, disputes and invoice errors.
What they published. In its February 2024 announcement, Klarna reported 2.3 million conversations in the first month, two-thirds of its customer service chats, work equivalent to 700 full-time agents, a 25% drop in repeat inquiries, resolution in under 2 minutes against 11 minutes previously, coverage across 23 markets and 35 languages, and an estimated $40 million profit improvement. Its Q1 2025 results added that customer service cost per transaction had dropped 40% since Q1 2023 "whilst maintaining customer satisfaction levels."
The second act. In 2025 Klarna began recruiting human agents again after service-quality complaints, a reversal that readers picked apart at length. Both halves of the story are true, and the honest reading is that the efficiency was real and the coverage was pushed past what the model could carry.
What to steal. Publish a cost-per-transaction figure rather than a headcount figure. Headcount claims age badly; unit economics do not.
2. Air India
The highest containment rate anyone in this category has put in writing.
What they run. AI.g, a generative AI virtual assistant launched in May 2023, answering baggage, booking and loyalty questions in the app and on the site.
What they published. In an April 2026 update, Air India reported 17 million-plus customer queries handled, more than 18,000 daily sessions, and a 97% containment rate, "resolving most queries without human intervention." The same programme sends over one million customer notifications daily at 95.5% delivery.
What to steal. The containment number sits on top of a rebuilt knowledge and data core, part of a $200 million modernisation across 140-plus systems. The sequence matters more than the tooling: they fixed the knowledge, then pointed AI at it.
3. Bell Canada
The only entry here whose number was filed with securities regulators, which makes it the least marketing-shaped figure in the post.
What they run. A generative AI chat virtual assistant plus an AI-driven Virtual Repair tool for technical diagnosis.
What they published. Bell's parent BCE told investors in its Q2 2025 presentation that the Chat VA had passed 7 million-plus customer sessions since launch, "with over 50% of problems solved without going to live agent," and that Virtual Repair had eliminated 1.2 million technician support calls since 2022. A voice virtual agent was slated to follow.
What to steal. Counting eliminated field visits, not just closed chats. The most valuable ticket AI handles is often the one that would have dispatched a van, so if your support has an expensive physical tail, that is the metric to instrument first.
4. Vodafone
The clearest before-and-after in the set, because Vodafone reported the same metric on an old bot and a new one.
What they run. SuperTOBi, a generative assistant built on Microsoft Azure OpenAI, replacing the older TOBi chatbot. TOBi served 13 countries in 11 languages; SuperTOBi rolled out through Italy, Portugal, Germany and Turkey.
What they published. In Vodafone's own newsroom, first-time resolution in Portugal rose from 15% to 60% and online net promoter score improved by 14 points to 64. The company allocated €140 million to customer experience transformation that financial year.
What to steal. The 15% baseline. Vodafone is the only company here that told us how bad the old bot was, which is the only way to know the new one earned its budget. Measure your current first contact resolution before you buy anything.
5. Delta Air Lines
The freshest deployment in the set, and a good model for staged rollout.
What they run. Delta Concierge, an assistant inside the Delta app that uses your trip and account context to answer questions and complete transactions.
What they published. On 10 August 2026 Delta announced Concierge had reached 100% of SkyMiles Members, less than a year after its beta launch, having "supported thousands of customer interactions while maintaining strong customer satisfaction." It can now cancel an eligible flight and submit a refund or issue an eCredit in real time, rebook during disruptions, track bags and look up SkyMiles balances. Chief Digital Officer Eric Phillips framed it as putting "more control in customers' hands, while ensuring Delta people continue to deliver the care and expertise that set us apart."
What to steal. Two things. The rollout went small group, then phased, then everyone, over roughly a year. And it still carries a beta label at 100% coverage, which is an honest signal rather than a hedge. Note also that transactional actions arrived after informational ones.
6. Verizon
The counter-example to "AI means fewer humans on the phone."
What they run. A pairing of Verizon Assistant in the My Verizon app with a human role called Customer Champion, both drawing on Google Cloud AI including Gemini models.
What they published. Verizon's June 2025 launch describes a dedicated expert for complex issues so that "customers only need to call once, and we take it from there," alongside expanded live call hours and new round-the-clock chat. Chairman and CEO Hans Vestberg framed it as "focusing on both people and technology."
What to steal. They spent AI savings on more human availability rather than less, then marketed that. If your escalation management is the weak link, this is the shape of fix that customers actually notice.
7. Shopify
The clearest example of AI support that sells rather than just deflects.
What they run. An AI sales associate inside Shopify Inbox, living on the merchant's storefront.
What they published. Shopify's Spring 2026 edition notes say it can "answer buyer questions, suggest products, and handle order inquiries using the data already in your admin, like catalog, inventory, and policies," personalise recommendations when a buyer signs in with Shop, and that "merchants have control over the tone and when to loop a human into the conversation."
What to steal. Grounding the assistant in commerce data rather than help articles. Most ecommerce support volume is order tracking and returns, and neither is answerable from a knowledge base alone. Both need live order state.
8. DoorDash
The entry that is not about ticket volume at all.
What they run. SafeChat+, which reviews in-app conversations between customers and delivery couriers for abuse.
What they published. DoorDash's own announcement says that "with SafeChat+, we're utilizing AI to review in-app conversations to detect and prevent verbal abuse or harassment," giving the courier a one-click exit and flagging the conversation to a human safety agent. The same technology now underpins its merchant chat.
What to steal. Using intent and sentiment detection as a routing trigger, not a reporting metric. The same capability that spots harassment spots a customer about to churn, which is the practical basis for a good handoff.
9. eesel
Full disclosure: this is our product, and it is on the list because it is the counterweight to the other eight. Klarna, Air India, Bell and Vodafone all built. Most teams cannot.
What we run. An AI layer that connects to the helpdesk you already have, learns from your past tickets and existing docs, and can be replayed over thousands of historical conversations before it goes live. It handles resolution as an AI agent, drafting as an AI copilot, and triage on the same connection.
What we publish. $0.40 per ticket or helpdesk conversation handled, billed per ticket rather than per reply, with no platform fee, no per-seat charge and no monthly minimum. $50 of free usage to start with no card. Enterprise adds a $1,000 per month platform fee for SSO, HIPAA and a BAA. An annual commit saves up to 25%.
What customers report. Kim Simpson at Gridwise put it plainly: "In the first month, eesel is resolving 73% of our tier 1 requests." That figure sits right in the band the big deployments report, which is the point, and it came from a 7-day trial rather than a multi-year programme.
Want AI customer service without a rebuild? eesel plugs into the helpdesk you already run, replays against your own resolved tickets so you see the automation rate before anyone else does, and bills $0.40 a ticket with no seats. Try eesel free, or book a demo.
Build it or buy it
Four of the nine built custom systems. That is a real option, and the numbers above show it works. It is also the most expensive way to find out your knowledge base was the problem.
The pattern in every build story here is that the AI was the last step. Air India rebuilt 140-plus systems first. Bell rebuilt self-install and diagnosis first. Klarna had 96% daily internal AI adoption before the assistant shipped. If you skip that groundwork, a custom model gives you fluent wrong answers faster.
Karel at GENERAL BYTES described the calculation the way most teams eventually do: "We could try to write our own LLM application but we didn't want to invest our time into that. We wanted something that we would not have to maintain." Maintenance, not construction, is where build costs actually land. Our build versus buy guide has the fuller comparison.
How to roll it out without Klarna's U-turn
Four steps, in this order, drawn from what the successful rollouts above have in common.
- Measure your baseline first. Vodafone knew its old bot resolved 15% of contacts first time. Without that number you cannot tell improvement from noise. Capture your current resolution rate alongside satisfaction and handle time before you shortlist anything.
- Fix the knowledge before the model. Every high-containment deployment here rebuilt its content and data layer first. Coverage tracks how complete your sources are, so connect past tickets as well as help articles.
- Test on your own history, not on scenarios you invented. Replay the AI over resolved tickets and compare what it would have said with what your team did say. This is the single step that separates the deployments that held from the one that reversed.
- Go live on one ticket type, with a confidence threshold and a clean escalation path. Delta started with information, then added transactions. Start with your highest-volume, lowest-risk intent and widen from there.
One customer, a CX lead at a DTC supplements brand, framed the whole thing better than any vendor deck:
"The AI will never be able to answer 100% of the questions. I need an AI who is only handling the tickets that it's confident to handle and all the other ones, leave them alone."
That is the design goal, and it is the opposite of the 700-agent headline. Restraint is what makes the numbers hold.
The honest summary
AI customer service works, at roughly half to two-thirds containment on general support queues and higher when the scope is narrow and the knowledge is clean. It costs between $0.40 and $2.00 per unit depending on whose unit you buy. And the failure mode is not the AI being obviously bad, it is the AI being confidently wrong on the tickets it should have left alone.
If you already run a helpdesk, you do not need Air India's $200 million or Klarna's platform team. You need clean knowledge, a way to test against your own tickets, and the discipline to start narrow. That is the actual finding from all nine.
For a wider shortlist of tools that do this, my roundup of the best customer service AI compares them side by side, and tier-1 deflection covers where to point the AI first. If you support customers in several languages, multilingual support is the other place the economics change fastest.
Frequently asked questions
How much does AI customer service cost in 2026?
It depends entirely on the billing unit, and the units are not comparable. Zendesk publishes $1.50 per automated resolution on a committed volume and $2.00 pay-as-you-go on its pricing page. Freshworks sells Freddy AI Agent sessions at $49 per 100, so $0.49 a session, plus $29 per agent per month for Copilot. eesel bills $0.40 per ticket or helpdesk conversation with no platform fee and no per-seat charge. A session is not a resolution and a resolution is not a ticket, so always normalise to your own volume before comparing. Our breakdown of AI customer service cost walks through the arithmetic.
What is AI customer service, exactly?
AI customer service is the use of language models to do four separate jobs on a support queue: resolve tickets end to end, draft replies for human agents to review, tag and route incoming requests on intent, and analyse conversations to find gaps in your help content. Vendors usually bill for the first two. The distinction matters because a tool that only drafts replies will never move your containment rate.
What containment rate do companies using AI for customer service actually hit?
The published figures span a wide range. Air India reports a 97% containment rate for its AI.g assistant across 17 million queries. Bell Canada told investors its Chat VA solves over 50% of problems without a live agent. Vodafone lifted first-time resolution in Portugal from 15% to 60% with SuperTOBi. Klarna's assistant handled two-thirds of chats in month one. All of these are self-reported, and the spread tracks how narrow the scope is more than how good the model is. See measuring containment quality for how to read them.
Do companies using AI for customer service replace their human agents?
The most-cited attempt walked itself back. Klarna announced its assistant was doing the work of 700 full-time agents in February 2024, then began recruiting human agents again in 2025 after service quality complaints, a reversal widely discussed by readers. Every current deployment we could source keeps humans on the complex tail. Delta is explicit that Concierge exists while "keeping Delta people at the center of the care." The useful question is not replacement but where to hand off to a human.
How much technical work does AI customer service take to set up?
It splits cleanly by approach. Klarna, Air India and Bell built custom systems on multi-year programmes, and Air India's wider digital rebuild ran to $200 million. Buying an AI layer that sits on your existing helpdesk is a different scale of job: an AI agent that connects to Zendesk or Freshdesk is a matter of minutes rather than quarters. Our build versus buy guide covers where each one wins.
What data do companies using AI for customer service train the AI on?
Shopify's storefront assistant answers from the data already in the merchant's admin, meaning catalog, inventory and policies. Air India's assistant runs on a rebuilt knowledge and data core. In practice the best signal is your own resolved ticket history, because it captures how your team actually phrases answers rather than how your help centre wishes they did. Public help articles, internal wikis and past tickets together are what make RAG answers specific instead of generic.
How do companies using AI for customer service measure whether it works?
Four numbers, watched together. Resolution rate tells you volume handled, CSAT tells you whether that was welcome, average handle time catches the tickets the AI made worse by half-answering them, and escalation quality tells you whether handoffs land with context. Klarna's Q1 2025 report is a good template: it paired a 40% drop in customer service cost per transaction with the claim that satisfaction levels held. Here is how to measure support ROI the same way.
Is it risky to let AI talk to customers?
It is, and the risk is concentrated in confident wrong answers rather than obvious failures. The mitigation that matters is testing on your own history before launch: replay the AI over thousands of resolved tickets, compare what it would have said to what your team did say, and only then route live traffic, starting with one ticket type. That plus a confidence threshold and a clean escalation path is most of the work. See preventing hallucinations in support.
