A memo got leaked in April 2025. Or, well, it was about to get leaked, so the CEO just posted it on X himself to get ahead of it. That CEO was Tobi Lütke of Shopify, and the memo basically said: before you ask for a new hire, you have to first prove that AI can’t do the job. Every team. No exceptions. Including the executive team.
People read it, shared it, argued about it on LinkedIn for two weeks straight. And then slowly, quietly, other companies started doing the same thing. Not always with a memo. Sometimes just in performance review criteria that suddenly included an AI usage section. Sometimes in job descriptions that started asking for “AI fluency” as a skill. And sometimes in contractors who stopped getting their contracts renewed, with no explanation except that the team had “found efficiencies.”
That’s where we are right now, in May 2026. The AI-first workplace mandate is not a future thing anymore. It’s already here. And it’s messier than the headlines suggest.
What Shopify and Duolingo Actually Did
Shopify’s memo was pretty direct about it. Lütke wrote that “reflexive AI usage is now a baseline expectation” and that AI questions would be added to performance reviews and hiring decisions. He also gave everyone access to a bunch of tools — GitHub Copilot, Claude, Cursor — and said the product design team is now expected to use AI for all feature prototypes.
Duolingo’s situation was a bit more chaotic. CEO Luis von Ahn announced in April 2025 that the company would go “AI-first” and that they’d “gradually stop using contractors to do work that AI can handle.” He also said new headcount would only happen when automation couldn’t solve the need. Users found out, got angry, started deleting the app. von Ahn then went on a clarification tour — he told podcasts, journalists, and anyone who’d listen that full-time employees weren’t getting fired. He even quietly dropped the part about AI usage being in performance reviews, saying “I’m not going to force you” in an interview about a year later, in April 2026.
So Duolingo basically did the announcment, got the backlash, and then walked back the harshest parts. Shopify has been quieter about the results. But both of them signaled something that a lot of other companies have been thinking.
The thing is — Klarna had already been doing this for a while. Their CEO Sebastian Siemiatkowski said their AI chatbot does the work of 700 customer service agents. By early 2026, they were down to around 4,000 employees from a peak of much higher, and Siemiatkowski had talked about eventually going to 2,000. That’s a 50% reduction in their workforce, if it happens. Nobody threw as big a fit about Klarna’s approach because they were doing it more slowly, without a dramatic memo.
The Part Nobody Talks About Much
Here’s what I found interesting when I looked at this more closely. The actual job losses that are happening right now are mostly contractor losses, not full-time employee losses. Duolingo cut 10% of its contractor workforce back in January 2024, months before the famous memo even went out. Amazon announced plans to cut 14,000 corporate jobs in late 2025 and was pretty open that AI efficiency was part of the reason.
But the way companies talk about it is always the same — they say something like “one person will be able to accomplish more” instead of “we’re going to have fewer people.” The math of that can go either way. If one person can do three people’s work, you might just have three times as much output with the same team. Or you might just… keep one person. Both outcomes are technically true to the statement.
I spent some time trying to find clear examples of full-time employees who were replaced, not contractors, and it’s much harder to find. Most of the documented layoffs include AI as one reason among several — UPS cut 30,000 jobs partly due to automation but also because they were winding down a big Amazon partnership. Salesforce cut 700 in 2024, didn’t call it AI, but also didn’t hire back. The data is genuinely messy and I don’t think anyone has a clean picture of it yet.
A November 2024 MIT study found that about 11.7% of jobs could be automated using existing AI — not future AI, existing AI right now. That’s a lot. But whether companies actually do that automation is a different question, because it costs money upfront and requires a lot of workflow redesign that companies are not great at.
The Mustafa Suleiman Prediction and Why It Made People Nervous
In early 2026, Microsoft’s AI chief Mustafa Suleiman said something that got a lot of attention. He said that “most tasks” in white-collar jobs — lawyers, accountants, project managers, marketing people — will be “fully automated by AI within the next 12 to 18 months.” That would put us somewhere around mid-to-late 2027.
That’s an aggressive timeline. I think he probably means “tasks,” not “jobs” — but even so. The thing that makes white-collar workers nervous is that these AI tools have gotten very good at exactly the kind of work that was supposed to be “safe.” Writing first drafts. Analyzing data. Summarizing documents. Scheduling. A lot of what a junior lawyer or a junior analyst does is exactly this stuff.
The World Economic Forum’s 2025 Jobs Report said 41% of employers expect to downsize by 2030 due to technology. So the companies themselves are expecting to reduce headcount. They’re just not saying it out loud, at least not in the way Lütke did.
The Backlash Is Also Real
What I didn’t expect when I started looking at this was how much pushback companies are getting internally. A study from early 2026 found that 31% of workers have refused to use AI tools at work. Not “they’re bad at it,” not “they don’t know how” — they’re actively refusing. Some of them said they were worried about job security and didn’t want to train the thing that would replace them. Others just found the tools frustrating.
The Duolingo backtrack makes more sense in that context. von Ahn admitted the AI-written stories for their app were not consistently good — “difficult to debug,” he said in one interview. Turns out some of the content AI generates is fine and some of it is kind of bad, and you still need people to tell the difference. That’s true across a lot of industries.
I’ve seen this in my own small way. I use AI tools a lot for writing-related work, and there are tasks where it saves me maybe two hours. There are other tasks where I spend more time fixing what it produced than I would have spent just doing it myself. The productivity gains are real but they’re not uniform across every kind of work.
So What Does This Mean for Regular Workers
The honest answer is: it depends on your job, and the situation is still developing. I know that’s not satisfying.
What seems clear is that companies are going to at minimum expect you to use these tools. Shopify has made it part of performance reviews. Other companies will follow. If you’re not using AI tools at your current job, you’re probably going to be asked to start in the next year or two. Whether that’s a good or bad thing depends on the tools and your specific workflow.
The Klarna model — where the company actually reduces headcount significantly — seems like the more aggressive end of what could happen. Not every company will go that far. But the contractor model is clearly changing. A lot of freelance and contract work in content, data entry, customer service, and basic coding is getting absorbed into AI systems. Those jobs are not coming back.
For full-time roles, the pattern so far has been to not replace people directly with AI, but to not hire replacements when people leave. That’s a quieter version of the same thing, and it’s happening across tech, finance, and increasingly other industries.
What Nobody’s Saying Out Loud
The companies running the AI-first experiments are not publishing their results. Klarna talks about its AI chatbot doing the work of 700 agents, but it doesn’t publish side-by-side customer satisfaction numbers. Shopify doesn’t say how much faster features are shipping since they made AI prototyping mandatory. Duolingo doesn’t say whether AI-generated content actually works as well as human-written content in helping people learn languages.
The data will probably come out eventually. But for now, a lot of the AI productivity claims are coming from the same people who benefit from the narrative — CEOs who’ve already made the investment and need it to look smart. That doesn’t mean they’re wrong. It means we should maybe wait before treating their claims as settled facts.
And look — von Ahn’s own words are useful here. He said AI-written code can be hard to debug. That’s not a minor thing. Debugging is a huge part of software development. If you use AI to write code faster but then spend twice as long debugging it, the net gain is smaller than it looks. He still thinks the overall direction is right, but the honest version is more complicated than “AI does everything better now.”
The next 18 months are probably going to tell us a lot more about whether these mandates actually work, or whether companies quietly scale back when the results are more mixed than expected. Right now it’s too early to know, and anyone who tells you they’re certain is probably oversimplifying.