Customer Support Went First

In February 2024 Klarna announced that its new AI assistant had handled two-thirds of its customer service chats in its first month, 2.3 million conversations. It said the assistant did the work of 700 full-time agents, that customers were as satisfied as with people, and that problems were solved in under 2 minutes instead of 11.

Fifteen months later its chief executive said cost had been "a too predominant evaluation factor", that quality had suffered, and that the company would invest in human support again. Customers should always be able to reach a person.

The AI was still handling about two-thirds of the inquiries. Nothing had been switched off. What changed was what the company decided to look at.

That sequence is the whole story of AI in customer support in miniature, and it's worth understanding well, because support is the department the rest of the company is about to repeat. Support went first because its work is unusually cheap to check and it already measured itself. And the two lessons it learned the hard way will apply to every department that follows: a conversation ending is not a problem ending, and the human work left behind gets harder, not easier.

Why support was ready

The tempting explanation is that support work is simple. It isn't. In the best study of AI in support, described below, the chats averaged about 40 minutes, much of it spent diagnosing technical problems.

The better explanation is that support work has five properties that make a new technology easy to deploy, easy to check and easy to measure.

Volume with repetition. The same questions arrive thousands of times, so a fix that works once pays back thousands of times. Contact centers also lose people fast, by industry estimates about 60 percent of agents a year, so they have a permanent reason to want something that teaches.

Text in, text out. A chat or email ticket has no physical step in the middle. Voice is harder, and customers dislike talking to AI by voice even more than by chat.

A record of past cases. Every closed ticket is a worked example: the question, the answer, how long it took, whether it worked. Add a knowledge base and a model has something authoritative to draw on.

Cheap, fast checking. The customer replies or doesn't. The ticket reopens within a week or doesn't. A reviewer can read twenty tickets in half an hour. Checking costs far less than producing, the same property that let code go first, with tests playing the role of the customer (Why Progress Arrives Unevenly).

A habit of measurement. Support teams have tracked handle time, resolution and satisfaction per agent per month for years. Most departments have no baseline to compare an AI result against. Support had one before anyone bought a model.

Cheap checking has an edge worth knowing. A wrong answer usually costs a recontact, but not always. In 2024 a Canadian tribunal held Air Canada responsible for its website chatbot's wrong advice about bereavement fares, and rejected the airline's argument that the chatbot was responsible for its own actions. The award was small, CA$812. The principle was not: whatever the bot says, the company said.

Scroll sideways to see the whole figure

The same five properties make a decent forecasting tool. Score any department from 0 to 3 on each, and the lowest score tells you where the work will be. I'll come back to which departments look like support at the end.

What the assistant actually did

The best evidence comes from a field study by the economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in 2025 in the Quarterly Journal of Economics. A Fortune 500 software firm gave its chat agents an AI assistant that suggested replies in real time and linked to the right internal documents. The agents stayed in charge and could ignore it; on average they followed 38 percent of the suggestions. This design is called agent assist: the AI drafts, the person decides. The tool reached agents in stages, so the researchers, following 5,172 agents, could compare those who had it with similar agents who didn't have it yet.

On average, agents resolved about 15 percent more issues per hour. That average hides the finding that matters.

The two-month agent with the assistant did as well or better. New agents with the tool reached in about two months a level that took eight to ten months without it. The least skilled fifth of agents gained 36 percent. Agents with more than a year on the job gained nothing, and the most skilled fifth saw small but real drops in quality.

The mechanism the authors propose is the most useful idea in the whole support story. The assistant had been trained on past conversations labeled by outcome and by whether the agent was a top performer. So its suggestions carried what the best agents tend to say and which documents they open. A new agent received, on the first day, patterns a good agent takes most of a year to absorb. An experienced agent already had those patterns, so the tool had little to give, and following it could pull them toward the average.

Two observations point that way. Agents who followed the suggestions more closely gained more. And during software outages, when the tool gave no suggestions at all, agents who had been using it still beat their own earlier numbers. That looks like learning, not copying.

The assistant didn't make agents faster typists. It handed the newest ones what the best ones knew.

The consequence for anyone rolling out AI in a team is direct. The gain flows to the people who lack the pattern. Measure your best people and you'll conclude the tool does nothing. Measure your newest and you'll be impressed. Both conclusions are correct, and neither describes the team.

The number that flatters: containment

Now the trap Klarna walked into, in a form every department will meet.

Support has two metrics that sound alike and mean different things. The containment rate is the share of conversations that ended without a person joining. The resolution rate is the share of problems actually solved, confirmed by the customer or by the customer not coming back about the same issue within a set window, often seven days.

A conversation can be contained because the bot fixed the problem, or because the customer gave up and closed the window. Containment can't tell those apart. It also rises when the human option gets harder to find.

Containment counts conversations that ended. Resolution counts problems that ended.

Most organizations can't yet tell which one they're getting. In an analysis of 432 customer service AI projects by the research firm Gartner, as reported in the trade press in 2026, about a quarter showed negative returns and 42 percent had unclear value. Unclear is the telling word: these were teams that couldn't say what their AI had produced.

A slower effect is coming. In another Gartner survey, customers were about three times as likely to take a service question to a general-purpose AI chatbot as to the company's own bot. A customer who has already tried one arrives with the easy questions answered. If that makes the tickets reaching you harder over time, a steady containment rate hides a decline.

Pricing has started to write the metric into contracts. Intercom charges $0.99 per "outcome" for its support agent, Fin, and an outcome counts when the customer confirms the problem is solved, or simply asks for no further help, or when Fin completes a workflow, handoffs to a person included. So a customer who gave up can be billable, and a hard ticket that ends with a person can be paid for twice: once to the vendor and once in human time. Paying for outcomes is a sound idea. Read the definition of the outcome before you sign.

A team of 40, traced

The second lesson only shows up when you follow the numbers. Take an illustrative support team of 40 agents handling 12,000 tickets a month. Every number here is an assumption, chosen to be plausible, not taken from any company.

About a third of the tickets are simple (order status, password resets) and take 10 minutes of a person's time. Two-fifths are standard (setup help, billing questions) at 22 minutes. A quarter are complex or sensitive (outages, exceptions, upset customers) at 46. Now route them three ways. AI resolves most of the simple tickets and a few standard ones, 30 percent of the total. The remaining simple and standard tickets go to a person with an assistant that saves 15 percent of their time, close to the study's average gain. Complex tickets stay with people. One in ten AI-resolved tickets comes back, and a person spends 20 minutes on each.

Scroll sideways to see the whole figure

Three things happen at once.

Human work falls by a fifth, and each remaining ticket gets harder. Human hours drop from about 4,760 to about 3,800 a month. The average time a person spends on a ticket rises from 23.8 minutes to 26. The easy tickets left, and what remains takes longer. A manager watching average handle time sees a team getting slower in the same quarter AI saved it a fifth of its hours. Judge that team by handle time and you punish it for the success.

Containment says 30 percent; resolution says 27. The gap is the 360 tickets that came back. Customers who gave up and never came back appear nowhere at all, which is why a weekly blind sample of AI-handled conversations, read against what a good agent would have said, isn't optional.

The saving exists only if hours leave. With staffing cut to match the new workload, cost per ticket falls from $16.00 to about $13. If nobody changes staffing, the team pays the AI fees on top of its old cost and cost per ticket rises slightly. The eight agents' worth of freed time becomes money only when it is no longer paid for, moves to work the company values, or absorbs growth (Where the Saved Time Goes). With the turnover support already has, attrition alone would get there in about four months.

Quality hides inside the hours, too. Each extra 10 points of reopen rate costs about one agent's worth of time. At a 30 percent reopen rate, staffing barely moves while three in ten AI answers fail the customer. That's how a company can look efficient on the dashboard and get worse in the customer's experience, which is roughly what Klarna said happened to it.

The beginners problem

Put the field study next to the worked example and a tension appears.

The assistant makes a new agent productive months sooner. Automation removes the simple tickets, which were exactly where beginners learned the job. A firm that needs the same volume handled can need fewer beginners, for both reasons at once.

Payroll data points the same way, with a caveat. Using payroll records through June 2026, Brynjolfsson and two coauthors find employment of 22-to-25-year-olds in AI-exposed occupations, customer service among them, about 19 percent below where it would have been had it kept pace with less exposed work, mainly because firms hired fewer young people. The authors call the result descriptive, not causal, and post-pandemic swings in hiring are a strong rival explanation. The question it raises stands anyway: where do the experienced agents of 2030 come from if the beginner tickets are gone? (Hiring When Producing Is Cheap)

The support job that remains is harder, more emotional and more about judgment. Whether pay and training change to match is a management choice, and most companies haven't made it yet.

Which department is next

Run the five properties down the rest of a company and the order is fairly predictable.

IT helpdesk is support's nearest relative: tickets, history, a knowledge base, service-level metrics. Checking is harder, because many tickets end in an action such as changing a permission, and a wrong action is a security problem. Start with assistance, and allow autonomy only for password-style requests with narrow permissions.

HR service desk has text, policy and repeated questions, at smaller volume. Answers can be checked against policy, but a wrong answer about leave or pay lands on a person.

Supplier inquiries in accounts payable ("where's my payment?") are checkable in the most literal way, because the answer sits in the ledger. The hard part is giving the AI properly permissioned access to the finance system.

Sales development has volume, text and a metrics habit, and fails on checking: the only test of an outreach message is a real person's slow, noisy reaction.

Contract review fails on checking in a different way. The work is text, but checking needs an expert, so drafting time saved comes back as review time. Field service mostly isn't text at all.

The lowest score is where the work is, and choosing the first process is a bigger decision than this score covers (Choosing the First Process).

What to do

My advice for support, and for whichever department looks like support next:

Route by the cost of checking and the cost of being wrong. AI first where the answer is a policy or a status lookup and a mistake is cheap. AI beside a person for standard tickets and anything with emotional or financial weight. A person first where a wrong move is expensive or resolving it needs authority, like a refund above a limit. Let the customer reach a person at any point.

Measure the end of the problem. Resolution with a recontact window, plus a blind quality sample, set up before you choose a vendor or a target (Measuring AI Impact Without Fooling Yourself). Containment is a cost metric dressed as a quality one.

Judge the team by cost per resolved problem, not handle time. Handle time will rise. That's a sign it's working.

Watch who gains, and plan how beginners learn. If the tool helps the newest people most, and automation removes the work they used to learn on, design the apprenticeship on purpose.

Invest in the record. The knowledge base, the outcome labels and the escalation logs are what the next model will use. Every escalation reason is a to-do: a missing document, a missing permission, or a kind of ticket that should stay human.

Support's record shows a technology that is productive, uneven and easy to oversell. The next department will get all three. The ones that do best will have decided what they're measuring before anyone tells them it works.

Check yourself

Why was the AI assistant more useful to new agents than to experienced ones?

It carried what the best agents tend to say and which documents they open. New agents got patterns that normally take most of a year to absorb, while experienced agents already had them.

Why can containment rise while customers get worse service?

A conversation can end because the customer gave up, or because reaching a person got harder. Containment can't tell that apart from a solved problem.

In the team of 40, why does average handle time rise even though the AI is working?

AI takes the easy tickets, so the ones left for people are harder and take longer. Judging the team by handle time punishes it for the success.