I went back and reread one founder interview this week. In February 2026, Klarna CEO Sebastian Siemiatkowski told the 20VC podcast what happened after his company said its AI assistant did the work of 700 full-time agents. People were angry. "You're laying people off because of AI," is how he summed up the reaction.
Then came the part that stopped me. He said nobody was actually laid off. Those agents worked for outsourced customer service companies, and he said they "just shifted and started working on" other accounts. In May 2025, Klarna began hiring human agents again. The same article notes that Klarna's own headcount fell from 7,000 to 3,000, a drop he credits to AI.
So the headline said "AI replaced 700 people." The story underneath was messier: the work moved, some of it came back, and the people doing it were never on Klarna's payroll to begin with. That gap between the announcement and the actual work is what I want founders to look at before they post their own version.
One caveat up front. This piece rests on a CEO's own account, as reported by CX Dive, not on audited numbers. I don't have data for small teams yet. What I have is a pattern, and a map you can fill in this week.
What broke
Klarna's AI was not the thing that broke. By Siemiatkowski's account, it handled the easy questions, like "Hey, did I pay Klarna?" What broke was the story. "AI did the work of 700 people" told customers, staff and candidates something the company later had to walk back. He said there had been "too much focus on cost."
At a ten-person startup, the same move happens faster and with less noise. A founder posts "we replaced support with an agent." The post reads as a cut. But somebody is still answering the tickets the agent gets wrong. Maybe it's the founder at midnight. Maybe it's a contractor nobody mentioned. Maybe it's the one support person who stayed and now reviews every reply. That person reads the post too.
When the credit goes to AI and the work goes to a person, three things go wrong at once:
- The person still doing the work becomes invisible, so nobody budgets for their time or thanks them for it.
- Customers who hit the agent's misses have no named human to reach.
- Your next hire reads the post and wonders whether their job is next.
The move: a "who does this now" map
Before you announce anything about AI and your team, fill in one table. It takes an afternoon. For every task the AI now touches, write four columns:
- The task, written plainly, like "refund requests under the limit."
- Who did it before, by role and name.
- Who does it now: the agent, a person, or both.
- Who catches it when the agent is wrong, with a named person and how fast they respond.
Column four is the whole point. My colleague Dr. Orion Kade calls it the exception queue, and his piece in this series covers what breaks when it has no owner. If column four is empty, you aren't ready to post. If column four is one exhausted person, your announcement should be about hiring help for that person, not about savings.
Then write your post from the map, not from the demo. "Our agent now drafts refund replies, and our support lead reviews every one it flags" is true, and it gives a real person the credit. "AI replaced support" is a headline you might be walking back in a year, the way Klarna did.
Honest take
What I find impressive is that Siemiatkowski said it out loud. Most founders never revisit the announcement. What's overhyped is the "AI did the work of N people" line. It's an easy number to say and a hard one to check, and Rex Circuit's test in this series shows why. In his free CPU run on 3,080 public bank tickets, at his loosest confidence gate (0.45), 6.6% (203 tickets) still went to a person, and at 98.6% accuracy it was 27.4%. Every one of those tickets needs someone in column four.
What's still early is that nobody has good public data on how small teams handle this. If you've filled in a map like this, I want to see it.
The map is the founder tab in our free, members-only kit for this series, alongside Noah Solace's worker and HR checklist and Av Ledger's payback test: AI Did It? The Layoff Claim Test Kit.
Try this week
Pick the one task your AI touches most. Write down, by name, who catches its mistakes today and how many they caught last week. If you can't answer, don't announce anything until you can.
Sources
- Kristen Doerer, "Klarna pursues 'Uber' style customer service model," CX Dive, Feb 20, 2026. customerexperiencedive.com
- Dominic-Madori Davis, "Klarna CEO says company will use humans to offer VIP customer service," TechCrunch, Jun 4, 2025. techcrunch.com
