12 lessons from the boss who let AI do the work of 700 agents, then brought humans back

The future sounded like a two-minute customer-service chat. In February 2024, Klarna said its AI helper had handled 2.3 million chats in one month. It covered two-thirds of the service talks. Its work equaled 700 agents.

Reuters later reported that staff fell from about 5,000 to 3,800, mainly through attrition. Then the story bent. By September 2025, Klarna’s chief executive said the firm had pushed cost cuts too far. It was hiring again.

This wasn’t a tale of 700 named workers being fired, then called back. It was something more useful: a bright promise meeting the hard grain of real work.

AI changed the work, but it didn’t make the work vanish

Klarna’s bot could handle two-thirds of service chats. Reuters said it cut the stated average resolution time from 11 minutes to 2. Yet fast answers did not erase angry customers. Or odd cases. Or fraud fears. Or the need for judgment

An OECD review covered nearly 100 workplace cases in finance and manufacturing across 8 nations. Job redesign was more common than job loss. Tasks moved. Roles changed. People still held the loose threads. That matters.

The first lesson for any boss with a gleaming new tool is simple. A job is a bundle of acts, not one button. AI may lift the plain boxes. A person still needs to read the torn label and hear the worry in a voice. They must decide what the rules missed.

A smooth demo can hide a rough day at work

A test room is quiet. Real customers arrive with typos, grief, rushed questions, and rare account problems. Their stories don’t fit a menu.

Verint’s 2024 survey found that 68% of customers had suffered a bad chatbot experience. Almost half had met a bot that gave them no path to a person. That costs trust. That gap explains why a system can look grand on a chart. It can still feel cold at the kitchen table.

Klarna’s first-month figures were strong, with a 25% drop in repeat questions. Yet one month cannot reveal every edge case. A careful boss tests the rainy days, not just the clear ones. They watch failed handoffs, wrong replies, repeated contacts, and complaints. Some cases take three tries to fix.

Human oversight is part of the design

AI needs a brake pedal, a rearview mirror, and someone awake at the wheel. McKinsey’s 2025 survey found that only 27% of groups had staff review every output before use. A similar share checked 20% or less. It also found that 47% had experienced at least one adverse outcome from the technology.

Klarna’s own turn fits that warning. Reuters quoted CEO Sebastian Siemiatkowski saying, “We probably over indexed a little bit on that, and then in the last six months we have been trying to course correct.”

Reviewers, appeal paths, audit logs, and clear owners matter. They may feel less thrilling than a launch. They are still the beams that keep the roof from sagging. They also keep trust in the room. That is the job.

The people who remain may face a harder job

Automation often eats the easy cases first. What stays on a worker’s screen can be tense, rare, or messy. OECD case studies across 8 countries found reports of higher work pace and stress after AI arrived. Some staff also lost dull tasks.

Pew Research Center found a similar split in 2025. Among workers who had used chatbots on the job, 40% said the tools helped them work much faster. Just 29% said they improved work quality as much.

A smaller team may move at a brisk clip while carrying heavier thoughts. The work shrinks in count, not weight. Leaders should track wait times, errors, breaks, and burnout together. Saving 9 minutes on a routine chat means little. Every human agent may inherit a line of fire.

Skills shift before whole occupations disappear

The World Economic Forum’s 2025 survey covered more than 1,000 employers. It found that close to 40% of job skills may change by 2030.

It also found that 77% planned to train workers for AI. Yet 41% expected to cut staff as some tasks became automated. The smarter path starts with the people who know the work.

A published study covered 5,172 support agents. It found a 15% average rise in issues solved per hour after workers got an AI helper. MIT professor Danielle Li put the learning effect plainly: “newer workers move down the experience curve faster.”

A seasoned agent knows where the process creaks. Train that person to guide the tool. Old know-how can become tomorrow’s map instead of a relic left behind.

AI can make a job better if people stay in the plan

There is a hopeful version of this story. The same study of 5,172 agents found that the lowest-skill group increased output by 36%. The most skilled group saw no clear gain. Customers also used less hostile language. New workers became less likely to quit.

OECD’s nearly 100 cases found less dull work and more interest. It also found better physical safety. People gain. Those benefits bloom when AI serves as a quiet helper. It should not be a trapdoor under someone’s chair.

Let software fetch a policy or draft a reply. Let a person read the room, spot harm, and choose the final words. A good tool can clear stones from the path. It need not pretend it knows the whole road ahead. That matters too.

An AI rollout is a people project

Buying software is quick. Rebuilding a workflow takes time, trust, training, and many small fixes. In 2025, McKinsey found that 78% of surveyed groups used AI in at least one business function. Yet only 21% had rebuilt even some workflows around it.

Fewer than 1 in 5 tracked clear performance measures for those tools. That is a wide gap between owning a machine and knowing how to work with it. University of Toronto researcher Kristina McElheran offered a crisp warning: “AI isn’t plug-and-play.”

Workers need a voice before launch and a safe way to flag faults. They also need paid time to learn. A rollout done with people can feel like a new door. A rollout done to them feels like a lock on the old one.

Law and risk keep humans close to the controls

Some work carries stakes that a speed chart cannot hold. Europe’s AI Act, Regulation 2024/1689, uses 4 broad risk levels. It requires human oversight for high-risk systems.

The European Commission also says staff need enough training to deal with such tools. In the United States, NIST’s AI risk guide rests on 4 linked steps. They are govern, map, measure, and manage. These rules matter in banking, hiring, health care, and insurance.

One wrong output can close a door on a life. A company may automate a first pass. It still needs named people who can check the result. They must be able to explain it, stop it, and answer for it. That duty stays. Accountability cannot be sent to an empty desk.

The strongest results come from assistance, not erasure

The best field evidence looks less like a robot takeover. It looks more like a good coworker at your elbow. In the 5,172-agent study, AI raised average output by 15%.

New and lower-skill workers gained far more than top performers. The tool spread patterns learned from skilled staff. Yet workers used only about 38% of their tips, according to an earlier account of the research. They kept their judgment.

The World Economic Forum found that almost half of employers planned to move staff from AI-exposed roles by 2030. Those workers would shift to other parts of the business. That is the wiser shape.

Machines sort, search, and suggest. People decide, soothe, and solve. The lamp works best because someone still chooses where to point it.

Trust is slower to rebuild than a software stack

People remember being told that a machine can replace them. Pew’s 2025 survey of 5,273 U.S. workers found that 52% felt worried about AI at work. Some 33% felt overwhelmed, while 36% felt hopeful. Only 6% thought it would create more job chances for them.

A later Pew survey found that 21% used AI for at least some work. That was up from 16% a year before. The worry now sits beside daily use. A boss asking former staff to return must offer more than a warm email. Pay, role safety, honest goals, and a voice in the next rollout matter.

The offer must repair both income and pride. A real plan is part of the apology. Trust is a small garden. One blunt cut can take a season to mend. Even good tools cannot rush that repair.

Speed and savings are not enough

Klarna’s early scorecard shone. Reuters reported that revenue per worker rose 73%. The service bot also reduced resolution time from 11 minutes to 2 minutes. Yet customer care is not a footrace.

Verint found that 70% of people might switch brands after a terrible service experience. Some 85% might buy again after an excellent one. McKinsey found that more than 80% of surveyed groups saw no clear company-wide profit impact from generative AI. Only 17% tied at least 5% of operating profit to it.

Leaders need a fuller page. It should track error rates, solved cases, repeat contacts, staff strain, customer loss, and appeals. A sound scorecard catches hidden bills. Cheap can become dear if a fast wrong answer sends someone out the door.

Bringing people back does not mean AI failed

Klarna did not turn off the future. Reuters reported in September 2025 that the firm still backed AI. It had shifted its focus from blunt cost cuts toward better products and services.

At that point, it had more than 24 open jobs. Its chief executive said the company was correcting course. That is not a neat win for humans or machines. It is a hybrid lesson written in pencil.

AI handled work equal to 700 agents. Yet people returned because quality, care, and hard cases still had weight. Course correction shows judgment, not shame. That takes nerve.

The mature move is not to defend a bold call forever. Study the miss. Keep what works. Restore what was lost. Let the next plan carry both speed and grace.

A quieter lesson

Pew found that 65% of U.S. workers still used little or no AI at work in September 2025. For most people, the grand office shift is arriving one task at a time.

That gives leaders room to choose well. A bot can shorten a queue. A human can hear the tremble behind a question. Work has always changed with its tools.

Dignity should not be treated like old software. The firms that remember may move a little more slowly at first. They may also build something people want to stay and care for.

Key Takeaways

Klarna’s case shows why the 700 figure needs care. It described the bot’s workload. It did not describe a verified group of 700 people who were fired and rehired.

The wider evidence points to a steady middle path. OECD found job redesign more often than job loss in nearly 100 cases. The 5,172-agent study found a 15% output gain with AI help, led by newer workers.

WEF found that 77% of employers planned AI training. At the same time, 41% expected some staff cuts. Good leaders set narrow goals and keep people in the loop. They test quality as hard as speed. They change course in public when the facts change.

Disclaimer – This list is solely the author’s opinion based on research and publicly available information. It is not intended to be professional advice.

Like our content? Be sure to follow us