What three CEOs said about AI, and what was actually running

An Authority Report is where we read what people with real operating responsibility said in public, in their own words, and work out what it means for a company that isn’t them.

This one takes three statements about AI from three CEOs across fifteen months. Two of them are about telling people to use AI. One is about building something that runs whether or not anyone is using it. That difference turns out to be the whole story.

April 2025: Lütke makes it a condition of employment

Tobi Lütke’s internal Shopify memo leaked, so he published it himself on X. The line everyone quoted:

Reflexive AI usage is now a baseline expectation at Shopify.

The word doing the work is reflexive. Not “encouraged”, not “available”. The memo’s argument for why is the part that got quoted less and is more interesting:

What we have learned so far is that using AI well is a skill that needs to be carefully learned by… using it a lot. It’s just too unlike everything else.

And on what he’d observed:

AI acts as a multiplier. And what’s even more amazing is that, for the first time, we see the tools become 10X themselves. I’ve seen many of these people approach implausible tasks, ones we wouldn’t even have chosen to tackle before, with reflexive and brilliant usage of AI to get 100X the work done.

Read it as an operating document and it’s a bet on a specific mechanism: capability comes from volume of use, so mandate the volume and the capability follows. The memo also required teams to demonstrate why AI couldn’t do a job before requesting headcount, which is the part that made it a policy instead of a pep talk.

What it actually specifies: a behavioural expectation on individuals, plus a gate on hiring. It does not specify a single system.

April 2025: von Ahn says the same thing louder, then walks it back

Days later, Duolingo’s Luis von Ahn went “AI-first”. Contractors would be phased out for work AI could handle, starting with content creation and translation, and new hires needed a justification that the work couldn’t be automated.

His framing was a platform shift:

We’re making a similar call now, and this time the platform shift is AI.

He also wrote, in the same memo:

This isn’t about replacing Duos with AI. It’s about removing bottlenecks so we can do more with the outstanding Duos we already have. We want you to focus on creative work and real problems, not repetitive tasks.

Both sentences were in the original. Only one of them travelled. The internet read “AI-first” plus “phase out contractors” and supplied the rest, the backlash ran for weeks, and by August von Ahn was saying he “did not give enough context” and pointing out that Duolingo had never laid off a full-time employee.

The lesson isn’t that he was wrong. It’s that “AI-first” is a phrase with no operating content, so it gets filled in by whoever reads it. Inside a company that’s just as true: tell a team you’re going AI-first and each person will privately decide whether it means “use ChatGPT more” or “your job is next”, and you’ll be managing the second interpretation for a quarter.

What it actually specifies: a hiring gate and a contractor policy. Again, no system.

June 2026: VandeHei describes a machine

Fourteen months later, Axios CEO Jim VandeHei published a first-person account of a year of using AI as hard as he could, under the title Confessions of an AI lab rat.

One to two hours a day in ChatGPT or Claude, for a year, self-described as being a lab rat.

Jim VandeHei, “Confessions of an AI lab rat”, Axios, June 2026

He fed it his medical records and blood tests, his diet, his workouts, his supplements. He’s reported as believing AI is smarter than 95% of people, on 95% of topics, 95% of the time, which is the kind of line that gets an article shared.

But the interesting sentence is the boring one. He describes controlling a personal operating system from his phone, connected to an always-on computer running several agents, including one that scans daily for data and trends relevant to a CEO.

Read that again as an architecture and not as a personality trait. There’s a machine that stays on. There are processes that run on a schedule without him. There’s an interface he touches when he wants to, and work that continues when he doesn’t. At Axios itself, the same piece describes testing AI coding tools across every layer of every department over the year.

What it actually specifies: a system with a trigger, a runtime, and an output that arrives whether or not the CEO opens an app.

The gap

Three statements, fifteen months apart, from people with nothing to prove.

Two of them describe human behaviour change. Use it more. Use it reflexively. Justify not using it. Both are reasonable, both were expensive to implement, and both put the entire burden on individuals to remember, in the moment, that a tool exists and would help here.

One describes something that runs on its own. The agent scanning for trends doesn’t need VandeHei to feel motivated on a Tuesday.

Almost every company we walk into is standing in the first category, wondering why the results aren’t showing up. They’ve bought licences. They’ve run training. Some have a mandate, in writing, from someone senior. And the honest measurement, when anyone does it, is that a few enthusiasts got much faster and everyone else went back to how they worked before, because the tool sits in a tab you have to decide to open.

Mandates change what people do while they are thinking about the mandate. Systems change what happens by default.

The first decays, and the decay curve is measured in weeks.

The uncomfortable follow-up

Notice what VandeHei had to do to get his machine: he built it. A personal always-on computer, running agents he assembled, wired to his phone. That’s a CEO with unusual technical appetite and unusual time, doing a year of full-time experimentation and calling himself a lab rat.

That’s not a reproducible strategy for a 60-person Mittelstand company. But it does tell you what the gap looks like from the top. He didn’t reach for a product that did this, because he’d have bought it. He built it, because the useful thing wasn’t a chat window and nobody was selling him the alternative.

The lesson for everyone else isn’t “become a lab rat”. It’s that the thing worth having is a system that produces output on a schedule, into a place someone already looks, without depending on anyone’s motivation that morning. Whether you build it, buy it, or have someone build it for you matters much less than whether it exists.

What we’d take from this

A mandate without a system is a request. If your AI initiative is a list of things people should do differently, you’ve asked for a favour repeatedly and you’ll get compliance for about six weeks. Pick one recurring task, automate it end to end, and let it run. One thing that runs beats a policy that everyone agrees with.

Say the operating detail, not the slogan. “AI-first” cost von Ahn a quarter of communications work because it doesn’t specify anything. “The weekly report is generated Monday at seven and lands in your inbox, and here’s who owns it when it’s wrong” specifies everything and generates no backlash.

Judge by what runs when nobody’s watching. The measurement that matters isn’t licences bought or people trained. It’s how many processes produce useful output on a schedule with no human trigger. For most companies with an AI strategy, that number is zero, and it’s the only number that would have shown up in the results.

If you want to move from a mandate to something that runs, book a strategy call.

Cover photo: Tobias Lütke, by Union Eleven, CC BY-SA 3.0, via Wikimedia Commons.

Sources: Tobi Lütke on X, April 2025 · First Round on Shopify’s AI adoption · TechCrunch on von Ahn’s clarification, August 2025 · Jim VandeHei, Confessions of an AI lab rat, Axios, June 2026

Frequently asked questions