why ai adoption mandates backfire (and what works instead)
AI adoption mandates backfire because they measure the wrong thing. Telling people to use AI makes usage the goal, so employees optimize for looking like AI users instead of doing better work — and they read the memo as a forecast of their own redundancy. WalkMe’s 2026 survey found 54% of workers had bypassed their company’s AI tools in the past 30 days and done the job by hand.
Two companies ran the experiment in public, in opposite directions. One of them had to walk it back.
what happened at duolingo
On April 28, 2025, CEO Luis von Ahn posted a companywide memo saying Duolingo would be “AI-first.” Three commitments in it did the damage: the company would gradually stop using contractors for work AI can handle, AI fluency would factor into hiring and performance reviews, and new headcount would be approved only when a team couldn’t automate more of its own work.
The public reaction was bad enough that Duolingo’s own social accounts got buried. But the interesting failure was internal, and von Ahn described it himself a year later. Employees started asking a question the memo had no answer for: do you just want us to use AI for AI’s sake?
In April 2026 he told the Silicon Valley Girl podcast that they’d given up on the metric. His words: “At the end, we backtracked, and we said, ‘No. Look, the most important thing in your performance is that you are doing whatever your job is as well as possible. A lot of times AI can help you with that. But if it can’t, I’m not going to force you to do that.’”
Read that carefully, because it’s a precise diagnosis from the person who wrote the original memo. The mandate created a conflict between doing your job well and demonstrating compliance. Once employees named the conflict out loud, the metric couldn’t survive it.
Worth being fair about what didn’t happen: this wasn’t a company using AI as cover for layoffs. In his first walk-back, von Ahn said Duolingo was “continuing to hire at the same speed as before,” and the contractor policy and the AI-first framing both stayed in place. The thing that got reversed was specifically the part that turned a tool into a scoreboard.
what moderna did differently
Moderna rolled out ChatGPT Enterprise and, within roughly two months, employees had built more than 750 custom GPTs across legal, research, manufacturing, and commercial teams. 40% of weekly active users had built at least one themselves. Average usage ran to 120 conversations per user per week, and the legal department — not the team anyone would pick in a betting pool — hit 100% adoption.
No mandate produced those numbers. Two other things did.
First, the groundwork. Before ChatGPT Enterprise there was mChat, an internal tool already used by more than 80% of employees. People weren’t being handed a foreign object; they were being upgraded on something they already trusted.
Second, and this is the whole ballgame: employees built the tools themselves rather than receiving them from IT. A central AI team can produce maybe a dozen good use cases a quarter, bounded by how well it understands other people’s jobs. Forty percent of your workforce building their own is bounded by nothing except what they know about their own work, which is a much bigger number.
the difference isn’t top-down vs bottom-up
The tidy version of this story is “top-down bad, bottom-up good,” and it’s wrong. Moderna’s AI push was extremely top-down. It had executive sponsorship, a dedicated AI products organization, and a CEO who reorganized the company around the idea.
What differed is what leadership handed down. Duolingo handed down a requirement: use this, and we’ll grade you on it. Moderna handed down capability: here’s a tool, here’s permission, build what you need. Both came from the top. Only one of them made the employee’s own judgment the enemy.
| Duolingo, April 2025 | Moderna | |
|---|---|---|
| what leadership supplied | a requirement and a review metric | a platform, prior tooling, and permission to build |
| what got measured | whether you used AI | what people built and used it for |
| employee’s role | comply | author |
| result | metric dropped a year later, AI-first branding kept | 750 GPTs in ~2 months, 40% of weekly users building |
the number that explains why mandates keep failing
WalkMe’s fifth annual State of Digital Adoption report surveyed 3,750 executives and employees across 14 countries in 2026, and the useful part isn’t the adoption rate. It’s the gap between what the two groups believe is happening.
88% of executives say their people have adequate tools. 21% of workers agree. 61% of executives trust AI for complex, business-critical decisions. Among workers, that’s 9%. And 78% of executives want to discipline unauthorized AI use, while only 21% of workers report ever being warned about an AI policy — they’re being judged against rules nobody told them.
Mandates get written from inside that gap. An executive who believes the tools are adequate and the output is trustworthy sees no reason for resistance except stubbornness, so the fix looks like enforcement. The worker refusing the tool isn’t being stubborn. They tried it, it produced something they’d have to fix anyway, and nobody trained them or told them what the rules were.
Add the fear, which is rational and rarely stated in the memo. When a company announces it will stop hiring contractors for work AI can handle, every contractor and every employee doing adjacent work now has a documented reason to worry. Asking that person to enthusiastically train their replacement is a big ask, and no performance metric will make it a smaller one.
what to do instead
The pattern that separates the two cases is short enough to act on.
Measure output, not usage. The moment “did you use AI” appears in a review, you’ve told people the tool matters more than the result. Duolingo’s own reversal is the cleanest available proof.
Let people build, not just consume. The single most useful Moderna number is that 40% of weekly users built something. Adoption stops being compliance the moment the thing being adopted is theirs.
Start from a tool people already trust. mChat at 80% adoption is why ChatGPT Enterprise landed the way it did. Rollouts that begin cold usually stay cold.
Say what happens to jobs, specifically. Silence gets filled with the worst available guess. If contractors are affected and employees aren’t, say that. If you don’t know, say that too — it’s less corrosive than a memo that implies an answer it won’t state.
Tell people the rules before you enforce them. 78% of executives wanting to punish shadow AI against 21% of workers who’ve heard a policy isn’t a discipline problem. It’s a communication failure with a discipline plan attached.
Von Ahn didn’t fail because he was wrong about AI being useful. He was right about that, and Duolingo is still an AI-first company by its own description. He failed at the one part every mandate gets wrong: he made the tool the objective instead of the means, and a workforce full of smart people noticed within a year. Moderna’s advantage wasn’t a better memo. It was never needing one.