Klarna, Duolingo and the AI-first memo: what happened after companies replaced people with AI

Between early 2024 and the middle of 2025, "AI-first" became a genre of CEO memo. The idea was always roughly the same: before you hire a person, show that software can't do the job. Some of those memos aged well. A few came back to bite the companies that wrote them, and the way they came back is more useful to a ten-person business than the original announcements ever were.

Klarna: 700 agents' worth of work, then a U-turn

In February 2024 Klarna said its AI assistant had handled 2.3 million conversations in its first month, two thirds of all customer service chats. It put the workload at the equivalent of 700 full-time agents, said the average time to resolve an issue had dropped from 11 minutes to under 2, and estimated a $40 million profit improvement for the year (Klarna). The company was also shrinking by not replacing people who left, and its CEO, Sebastian Siemiatkowski, talked openly about AI doing more of the work.

Fifteen months later he told Bloomberg something different: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." Klarna started recruiting human agents again, on a flexible remote model, so customers would always have the option of speaking to a person (CX Dive).

Note what didn't happen. Klarna didn't switch the bot off. It still answers the simple questions, the refund-status and where's-my-payment messages that make up most of the volume. What changed is that a person is back on the other side of the hard ones.

Cost was the measure. Quality was the thing that got worse. Nobody was counting it until customers did.

Duolingo and Shopify: the memo was the problem

In April 2025 Duolingo's CEO, Luis von Ahn, sent an all-hands memo declaring the company "AI-first". It said Duolingo would gradually stop using contractors for work AI could handle, and that new headcount would only be approved where a team couldn't automate more. Users and staff read it as a layoff notice, and the backlash on social media was loud enough that he followed up a few weeks later: "I do not see AI as replacing what our employees do (we are in fact continuing to hire at the same speed as before)" (Entrepreneur). By August he was saying the memo hadn't given enough context and that Duolingo had never laid off full-time employees (Fortune). The contractors were the part that did change.

Shopify's Tobi Lütke had sent a similar memo a few weeks earlier: using AI was now a "fundamental expectation", and teams asking for more people had to show first why AI couldn't do the work (CNBC). It drew less anger, mostly because it was about future hiring, not current jobs.

The lesson from both is about words, not models. Saying "AI-first" to your staff and customers tells them what you value. If what you mean is "we'll use AI to do more with the same team", say that instead.

Commonwealth Bank: the numbers didn't hold

In July 2025 the Commonwealth Bank of Australia cut 45 customer service roles after introducing a voice bot on its inbound line, saying the bot had reduced calls by about 2,000 a week. The Finance Sector Union took the case to the workplace tribunal, arguing that call volumes were actually rising and that the bank was offering overtime and pulling team leaders onto the phones. In August the bank reversed the cuts, apologised to the staff and said its assessment "did not adequately consider all relevant business considerations" (ACS Information Age; Bloomberg).

This one is the most instructive, because it's so ordinary. The roles were cut on a forecast of what the bot would do, before anyone had checked what it was doing.

The pattern in the surveys

These aren't isolated stories. Orgvue surveyed more than 1,000 business leaders in 2025: about four in ten had laid people off because of AI, and 55% of those said they'd made the wrong call (HR Dive). In February 2026 Gartner predicted that by 2027 half of the companies that attributed customer service cuts to AI will rehire for similar work, often under different job titles. Its own survey of 321 service leaders found that only 20% had actually reduced agent numbers because of AI; most recent cuts had more to do with the economy than with automation (Gartner).

So the story isn't "AI failed". The bots are still running at Klarna and at the bank. The story is that companies took a tool that handles the routine part of a job and treated it as if it handled the whole job.

Where agents fit in a small business

A small company has an advantage here: you can see every conversation, and you know your customers by name. Use that. Agents earn their keep on work that is:

High volume and repetitive. Order status, opening hours, password resets, "did my payment go through". If you answer the same question twenty times a week, an agent should answer it.

Checkable. Drafting a reply, a quote or a first version of a document that a person reads before it goes out. The agent saves the typing; the person keeps the judgement.

Easy to undo. Tagging tickets, sorting email, filling a spreadsheet. If it gets one wrong, nobody loses money or trust.

Where people stay

Money and disputes. Refunds outside policy, billing errors, chargebacks. These are exactly the moments Klarna moved back to people.

Exceptions. Anything the rules didn't anticipate. A bot follows the policy; a person knows when the policy is wrong for this customer.

Anyone who's already upset. A customer who has tried the bot twice and is now angry needs a human, fast. Make that path visible: a "talk to a person" option that actually works is cheaper than the review you'll get without it.

Four rules before you change a role

Measure before you cut. Run the agent alongside the people for a month. Count what it resolves without a follow-up, not just what it answers. Commonwealth Bank skipped this step.

Track quality next to cost. Repeat contacts, complaints, refunds and customers who leave. If cost is the only number on the dashboard, it's the only thing that will improve.

Shrink by not hiring, not by firing. Most companies that did this well let AI absorb growth: the team stays the same size while the business gets bigger. It's slower, and it's reversible.

Count the AI in your margin. Model and API costs grow with every customer conversation. A team that looks lean on revenue per head can still be running at a thin gross margin, which is the trap we covered in small teams, big revenue per head. It's the same gap between use and returns we wrote about in how businesses are coping with AI in 2026.

We built a small calculator for that last point. Put in your revenue, your headcount and what you spend on models and hosting, and it shows revenue per employee and gross margin side by side, against the published benchmarks.

Sources

Klarna, AI assistant handles two-thirds of customer service chats in its first month (February 2024); CX Dive, Klarna changes its AI tune and again recruits humans for customer service (May 2025); Entrepreneur, Duolingo CEO clarifies AI stance after backlash (May 2025); Fortune, Duolingo CEO admits memo did not give enough context (August 2025); CNBC, Shopify CEO memo (April 2025); ACS Information Age, CBA reverses AI-driven job cuts (August 2025); Bloomberg, Australia's biggest bank reverses plan to replace jobs with AI (August 2025); HR Dive on the Orgvue survey (2025); Gartner, half of companies that cut customer service staff due to AI will rehire by 2027 (February 2026). Hero photo from Pexels (Pexels License).