7 min read ·
An AI-first memo is a hiring freeze
The most important sentence in Shopify's AI memo isn't about AI. It says a team has to prove a chatbot can't do the work before it's allowed to hire a person. Take the vocabulary out of this spring's AI-first memos and most of them say the same thing: stop hiring, but never write the word freeze.
Tobi Lütke posted the memo on X on April 7, after he sent it to staff and it started to leak. CNBC covered it the same day. The title is "Reflexive AI usage is now a baseline expectation at Shopify."
I'll give him the title. I use coding agents every day, and an engineer who still refuses to try them in 2025 is being stubborn and calling it taste.
But two rules sit under the title, and they're the parts that will outlive the memo. The first: "We will add AI usage questions to our performance and peer review questionnaire."
The second: "Before asking for more Headcount and resources, teams must demonstrate why they cannot get what they want done using AI." Right after it comes a question for every team: "What would this area look like if autonomous AI agents were already part of the team?"
That's a fine question for a planning offsite.
It's a terrible question to staple to a hiring request.
Start with the review rule, because it's the smaller one.
Goodhart's law says that when a number becomes a target, it stops being a good measure: people hit the number whether or not the thing behind it got better. Put "how much do you use AI?" in a performance review and you'll get engineers who route every ticket through a chat window where the usage shows up, whether or not it helped. You'll learn a lot about the tool. You'll learn nothing about the work, because you never measured the work.
It's like grading a writer by how often she opens the thesaurus.
Still, a bad review question does limited damage. People learn to answer it, managers learn to skim it, and in two cycles it's furniture.
The headcount rule is the real memo.
Picture a manager with a backlog that keeps growing and a team that's tired. She writes up a request for one more engineer. It goes two levels up and comes back with one line: "Have you tried doing this with AI?"
She has. The model handles some of it, not all of it. But the rule asks her to demonstrate why AI can't do the work, which means proving a negative in a year when a new model ships every few weeks. Whatever she shows today, someone can say "wait for the next release."
So the request sits. Next quarter it's still sitting. Nobody said no. Nobody signed anything with the word "freeze" in it.
I call this headcount theater. It's a freeze, staged as a process.
Why would a company prefer the theater to the freeze? Because a freeze is a decision. Someone puts their name on it, the board hears it, and it tells the world that money is tight. A rule that says "show us why AI can't do it" tells the world the company is so modern it barely needs people. The outcome is the same. The press release is better.
Shopify wasn't first, just most quotable.
Back in December, Marc Benioff told the 20VC podcast, "We're not adding any more software engineers next year," meaning 2025, and credited AI with raising engineering productivity by more than 30%. He said it again on the February earnings call. Measured how? He didn't say.
In the same interview he said Salesforce would "probably add another 1,000 to 2,000 salespeople in the short term." So engineering got 30% more productive, by his own number, and the new seats went to the people who sell the AI. That tells you what the freeze was for.
On April 28, Duolingo posted an all-hands email from its CEO, Luis von Ahn, on LinkedIn. Duolingo is going "AI-first." It will "gradually stop using contractors to do work that AI can handle." AI use becomes part of hiring and of performance reviews. And then the Shopify rule again, in slightly different words: "Headcount will only be given if a team cannot automate more of their work."
Read that word "more" twice. There is always more you could automate. No team can ever show it has finished automating, so no team can ever pass the test. It's a door with no handle on your side.
Then there are two dates I keep putting side by side.
On April 29, on stage at Meta's LlamaCon, Satya Nadella said that "maybe 20%, 30% of the code that is inside of our repos today" is written by software, meaning AI. That's a range ten points wide, offered from a stage, and it says nothing about who owns that code once it ships. I wrote about that in March.
Last Tuesday, May 13, Microsoft said it would cut about 6,000 jobs, less than 3% of its workforce, and the company said one aim was to remove layers of management. In Washington state, where Microsoft had to file advance layoff notices, more than 40% of the roughly 2,000 local cuts were software engineers.
Microsoft didn't say AI caused any of it. It didn't have to.
I'm not claiming the memos cause the layoffs. Companies cut people for a hundred boring reasons, and I can't see inside Microsoft's spreadsheet. What the memos do is supply the story. Every manager now has a sentence to quote, and every finance team has a reason to send a request back.
And the requests that bounce first are the junior ones.
A first-year engineer's work is small and a bit dull. Fix the validation on this form. Add a column to this report. Write the test nobody wrote. That's exactly the kind of task a chatbot demo handles well, because demos are built from tasks like that.
A manager who asks for a senior engineer can still argue for judgment: someone to own the migration, someone to say no to the bad design. A manager who asks for a new grad loses the argument to a two-minute screen recording.
But nobody hires a junior for the tickets.
The tickets are the tuition. You give someone small things that matter a little, they break a few, someone shows them why, and five years later they're the person who says no to the bad design. The headcount rule scores the request against the tickets, and the tickets were never the point. Stop hiring juniors this year and you'll find out around 2030, when you go looking for senior engineers and discover that nobody made any.
The New York Fed keeps a table of outcomes for recent college graduates by major. In the edition it published this February, unemployment for recent computer science graduates was 6.1%, and for computer engineering 7.5%. Both were above the rate for recent graduates as a whole. That should embarrass an industry that spent a decade telling teenagers to learn to code.
But that edition runs on 2023 data, from before most coding agents existed. So I won't pin it on a chatbot. Junior hiring was already cold. The memos land on a market that was shrinking for new graduates before anyone wrote them, and they give the next new-grad request one more reason to bounce.
A company can freeze hiring for honest reasons. Money gets tight. It hired too fast in 2021 and has to stop. Then someone signs the freeze, stands up at the all-hands, and answers for it. It hurts, and it has a name on it.
Headcount theater is the same freeze with nobody's name on it. The graduate who never gets the offer hears that a machine made the call, and the person who actually made it never has to stand up.
If you mean freeze, write freeze.