AI for customer support: help your agents, don't wall off your customers
Klarna's bot did the work of 700 agents, then the humans came back. Where AI actually earns its place in a support team.
In February 2024, Klarna announced that its new AI assistant had handled 2.3 million conversations in its first month: two-thirds of all support chats, the equivalent work of 700 full-time agents, resolution times down from 11 minutes to under 2, an estimated $40 million profit improvement for the year. It was the most quoted number in customer service. Fifteen months later came the less quoted one: CEO Sebastian Siemiatkowski admitted the cost focus had gone too far, quality had suffered, and Klarna was recruiting human agents again.
The wrong reading of that arc is “AI support failed.” It didn’t: Klarna’s assistant still fronts the bulk of its chats today. The right reading is about placement. AI in customer support works brilliantly beside your agents: drafting replies, retrieving answers from your real docs, triaging queues, summarizing threads. It fails expensively when it’s placed between your customers and your humans as an unsupervised wall. This post is a practical toolkit for support leads on the first kind, with the receipts on what the second kind costs.
Klarna and Air Canada learned the same lesson
The second cautionary tale is smaller than Klarna’s and more instructive. Jake Moffatt booked a full-price Air Canada flight after his grandmother died, because the airline’s website chatbot told him he could apply for the discounted bereavement fare within 90 days of buying the ticket. The real policy allowed no refunds after travel. The chatbot had made its version up. When Air Canada refused to honor it, a British Columbia tribunal ordered the airline to pay $812. Air Canada had argued the chatbot was “a separate legal entity that is responsible for its own actions.” The tribunal called that submission remarkable and rejected it: everything on your website is you, including the bot.
Two different failures, one root cause. Klarna put AI in front of customers faster than quality could follow; Air Canada let a bot state policy with no connection to the actual policy documents. In both cases the model stood between the customer and the company, with nobody checking its work in the moment it mattered. Customers already suspected this would happen: in a Gartner survey of 5,728 customers, 64% said they would prefer companies didn’t use AI in customer service at all, and 53% would consider switching to a competitor over it. Their top stated concern was not accuracy. It was that AI would make it harder to reach a person.
Read that carefully and the survey stops being an argument against AI. Customers aren’t anti-model; they’re anti-wall. Every pattern in the rest of this post either keeps the human reachable or keeps the human in the loop entirely.
The unglamorous wins: drafts, summaries, triage
The highest-return uses of AI in support never appear in a press release, because the customer never sees them. The first is drafting. A model that has read your help center and a few hundred of your best past replies can draft an answer in your brand voice for nearly every incoming ticket. The agent edits and sends. Response quality stays under human control, the accountability chain stays intact, and the time per ticket drops because agents stop writing the same three paragraphs from scratch forty times a day. New hires sound like veterans in week one, which shortens onboarding before it shortens queues.
The second is summarization. Every escalation and shift change used to begin with the most hated sentence in support: “Could you repeat your issue?” A model that condenses a 40-message thread into five lines of who, what, and what’s been tried lets the next person start where the last one stopped. The third is triage: classifying intent, urgency, and sentiment on arrival, routing the billing dispute to billing and the furious enterprise customer to a senior agent. And the fourth is analytics: models reading the whole queue and surfacing what no individual agent can see, like a spike in tickets mentioning checkout errors that started an hour after Tuesday’s deploy. The same pass upgrades QA from sampling to coverage: score every conversation against your rubric instead of auditing 2% of them, and hand coaches a short list of the calls actually worth replaying.
Notice what these four have in common: none requires a customer-facing bot. A team can capture all of them while remaining, from the customer’s side of the glass, 100% human. If your customers match Gartner’s 64%, this is the mode that improves your numbers without spending any of their goodwill.
Ground the bot in your docs or it invents policy
When you do put AI in front of customers, the Air Canada failure is the one to engineer against. Language models answer from patterns, not databases; asked about a refund policy they haven’t seen, they will produce a fluent, plausible, wrong one, for reasons we’ve covered in why AI makes things up. The fix is retrieval-augmented generation: before the model answers, the system searches your actual help center, policy pages, and product docs, hands the model the relevant passages, and instructs it to answer only from them.
Grounding is what separates a support agent from an improviser, and it’s worth doing strictly. Require the bot to cite the help article it drew from, so customers can verify and so your team can trace wrong answers to stale docs. Give it an explicit path for “I don’t know”: a question with no matching passage should trigger a handoff, not a guess. Scope the index deliberately too: the public bot reads your published help center, not your internal wiki, so nothing can travel from a private page to a customer’s chat window. And treat your knowledge base as the bot’s source code, because it now is. A stale refund page used to mislead the occasional reader; wired into a bot, it misleads every customer who asks, in confident prose, with your name on it. The tribunal’s message in the Moffatt case was blunt: the company remains liable for what its chatbot says. Draft every bot policy sentence as if a rep said it on a recorded line.
Customers forgive a bot that hands off fast
The rage-inducing chatbot is not an AI problem; it’s a design choice. It happens when a company treats the bot as a gate customers must defeat before earning a human. The alternative is a small set of rules, none of them technically hard. Say it’s AI, plainly, in the first message. Keep a visible path to a person in every exchange, not buried behind three failed answers. Escalate automatically on frustration signals: repeated rephrasing, all-caps, an explicit “let me talk to someone.” And when the handoff happens, pass the transcript and a summary with it, so the human opens the conversation already knowing everything the customer just typed. Write these rules down where agents can see them, so everyone knows what the bot owns and what it must surrender, and read a sample of looping transcripts weekly to catch the gates forming.
Klarna’s own correction is the template. It didn’t remove the AI; it re-tiered the system, keeping the assistant on the high-volume routine chats it resolves well and reinvesting in humans for the complex and high-stakes cases, with Siemiatkowski framing quality human support as something the company would treat as a premium offering. Measure the split honestly and it maintains itself: track resolution and satisfaction, not deflection. A bot that “deflected” a ticket by exhausting the customer into giving up looks like savings in one dashboard and churn in another.
Per-resolution pricing makes the math honest
The vendor market has converged on a pricing model that quietly confirms everything above: you pay for outcomes, not conversations attempted. Fin charges $0.99 per resolution, counted only when the customer confirms the answer worked or leaves without asking for more; unsuccessful attempts cost nothing. Salesforce prices Agentforce at $2 per conversation, or by the action through Flex Credits. When the vendors themselves only get paid on resolution, “how many tickets did the bot touch” is officially the wrong metric even by the sellers’ accounting.
The comparison that matters is against your own numbers: take your loaded support cost from payroll, divide by monthly resolved tickets, and you have the human cost per resolution that $0.99 is competing with. At list price, 2,000 bot-resolved routine tickets a month meter out to about $1,980; the same volume handled by people is a payroll line. For most teams that gap is wide enough that the bot pays for itself on the routine tier even at modest resolution rates. Just keep two asterisks attached. A billed “resolution” that didn’t actually fix the problem costs you a customer, not 99 cents. And when Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029, the load-bearing word is common. The uncommon 20% is where loyalty is won and lost, and it’s exactly the tier Klarna re-hired for.
Customer data is an obligation, not fuel
Support tickets are among the most personal text a company holds: names, emails, addresses, order histories, and, depending on your product, health details and financial records. Routing that text through an AI system is a data-processing decision with legal weight under privacy regimes like GDPR, not just a tooling choice. The operational floor: send the minimum (the ticket, not the customer’s full record), redact identifiers where the model doesn’t need them, and use business API tiers under a data processing agreement with retention limits, never consumer chatbot accounts, which may train on what your agents paste.
For the sensitive end of the spectrum there’s a stronger option: don’t send the text anywhere. Open-weight models running on your own hardware now handle drafting, summarization, and classification well, and the ticket never leaves your infrastructure. We keep a current shortlist in the best open models you can run locally. A sensible split many teams land on: local models for the PII-heavy internal work, a cloud agent under a DPA for the public-facing docs bot, which by design only reads your published help center anyway.
Start with drafts, earn your way to deflection
The sequence that works runs opposite to the sequence that makes headlines. Start inward: turn on ticket summaries and reply drafting for your agents this week; it’s low-risk, customer-invisible, and it produces a number (the share of drafts agents accept) that tells you whether your AI actually understands your product. While that runs, clean the knowledge base and wire up retrieval, because grounding quality caps everything downstream. Only then put a bot on the front door, scoped to your top three routine intents, with disclosure, citations, and a one-click human handoff. Expand intent by intent as resolution and satisfaction hold.
Klarna ran this ladder roughly backward, in public, with a strong model and real engineering, and still spent 2025 climbing back down. The companies that get AI support right won’t be the ones that automated the front door first. They’ll be the ones that automated the back office first, kept a human within one click, and let the bot earn the counter one intent at a time.
Disclaimer: This article is general information, not legal advice, and reading it creates no attorney-client relationship. Laws, regulations, and court rulings summarized here reflect sources available as of July 2026 and may have changed. Consult counsel licensed in your jurisdiction before acting on any of it.
