The pitch for AI-run teams is that they finally get rid of the tension that comes with people: no egos, no meetings, no pushback. That pitch has the order backwards. The tension people bring to a team was never the problem. It was doing real work, and a growing number of companies that cut it are now paying to put it back.
The Rehiring Boomerang
In 2023, Klarna's CEO announced that an AI chatbot was handling the workload of 700 customer service agents, and the company became the public face of running support without a human team. By mid-2025 the story had reversed. Their customer satisfaction scores dropped and complex complaints kept getting generic answers. By early 2026 Klarna was quietly rebuilding its human support team in a hybrid model.
Klarna isn't an outlier. A February 2026 survey of 600 HR professionals found that two in three employers who cut jobs because of AI are already rehiring, and more than a third of that group brought back over half the roles they'd eliminated. Forrester's 2026 Future of Work report asked the question directly and got a blunt answer: 55% of employers said they regretted their AI-driven layoffs. Duolingo cut its contractor writers and translators in 2023 and 2024, shipped 148 new courses in under a year, then spent 2025 fielding complaints that lessons had gone formulaic, and started quietly rehiring for the roles it had cut.
None of these companies lost because AI failed the task. Klarna's bot could answer questions. Duolingo's models could generate lessons. What disappeared when the humans left was harder to spot on a spreadsheet. What disappeared was the willingness to say no to a bad instruction on the way to executing it.
Structured Work Holds, Judgment Work Doesn't
Not every AI-for-headcount swap collapsed. Pilot's AI bookkeeper has run the full accounting cycle for thousands of startups without a public reversal. Wendy's FreshAI now handles most drive-through orders across more than 160 locations. Salesforce cut support headcount from roughly 9,000 to 5,000 using its own AI agents, without the backlash that hit Klarna.
The difference isn't the industry. It's what the job actually asks for. Bookkeeping, order-taking, and support routing have clear rules and a measurable right answer, and removing a person from that kind of process mostly removes redundant labor. Customer escalations, translation nuance, and sales development ask someone to weigh context that doesn't reduce to a rule, and removing a person from that kind of process removes the judgment along with the labor.
Solo Operators Are Running the Same Experiment, But Faster
Business Insider's "AI-Powered Solopreneur" series profiled three solo founders who each independently found their AI "employees" agreed with them too easily to be useful.
Yesim Saydan, a branding consultant, built 17 custom GPTs modeled on mentors she admires. Asking any of them "what do you think of this idea" produced reliable flattery, so she changed the question: now she asks for a score from one to ten, then asks what would make it a ten. The rating format forces a critique the open question never did.
Aaron Sneed, a solo defense-tech founder, built what he calls "the Council": fifteen AI agents covering legal, HR, finance, compliance, and engineering, standing in for a leadership team he can't afford to hire. His first version agreed with him by default on everything, so he went back and retrained each agent specifically to push back, then built a priority order so legal and compliance override the rest when they conflict. Tim Desoto, a non-technical founder, runs an "AI conveyor belt". The same idea is fed to multiple models on purpose so the models can disagree with each other before he commits to a direction.
None of these three hired a person to get this. All three deliberately built friction back into a system that defaults to agreement.
"An AI system that only executes is a tool. An AI system that can tell you the instruction is wrong is closer to a team."
Klarna's reversal, Duolingo's rehiring, and Saydan's rating-scale fix all involve models that were already capable of the underlying task before anyone fixed anything. What changed wasn't the AI. It was whether the system had a mechanism built in for it to disagree. For anyone building a system out of their own expertise instead of hiring a team, that's the design decision worth making on purpose, not which model to use, not how fast it ships, but whether there's anywhere in it that's allowed to say no.