In April 2025, Shopify CEO Tobi Lütke posted an internal memo publicly. The line everyone quoted: reflexive AI usage is now a baseline expectation at Shopify.
First Round Review later noted that the format "became a genre." Box, Fiverr and others put out their own versions. Before long, "we are now an AI-first company" was a sentence every engineer had heard at some all-hands.
A memo is a statement of intent. What actually happens to engineers is decided by what comes after it.
This is part 3 of a series on how engineers who learned to build software before AI see the AI era. Part 2 showed that executives and engineers describe the same rollout differently. This part asks a narrower question: when leadership announces the vision, does it actually get implemented, and implemented as what?
Here are three companies that wrote roughly the same sentence and did three different things with it, plus a fourth that skipped the memo entirely.
Version 1: Shopify Built the Road Before Putting Up the Sign
The Shopify memo wasn't the beginning of anything. It wrote down something that was already happening.
According to Shopify's VP and Head of Engineering, Farhan Thawar, speaking to First Round, the company brought in GitHub Copilot a year before ChatGPT launched, early enough that it couldn't even pay for the product yet. Adoption quickly reached 80%. Thawar's reaction at the time: "I thought it was terrible. 20% of engineers still aren't using Copilot?"
Two details stand out.
First, legal was told to say yes. When he wanted Copilot in 2021, Thawar's opening line to the lawyers was "we're likely going to do this. How can we do it safely?" His principle: "If you don't default to 'yes,' you're defaulting to 'no.'"
Second, access was open to everyone. Shopify lets anyone use every tool and model. "I ordered 1,500 Cursor licenses last year and quickly had to procure another 1,500," Thawar said. "The fastest growing groups using it are not engineering. It's support and revenue."
By the time the memo landed, engineers already had the tools, the legal cover, and three years of practice.
Version 2: Coinbase Held a Saturday Meeting
In August 2025, Coinbase CEO Brian Armstrong described his approach on Stripe co-founder John Collison's podcast, as reported by TechCrunch.
Coinbase had bought GitHub Copilot and Cursor licenses for every engineer. Some people warned Armstrong that it could take months to get even half of them using AI. His response, in his words: "I went rogue." He posted in the main engineering Slack channel that everyone had to onboard by the end of the week, and that anyone who hadn't would be invited to a meeting on Saturday to explain why.
On the Saturday call, some people had good reasons, such as having just come back from a trip. Others didn't. "And they got fired."
Armstrong himself called it "a heavy-handed approach" and admitted that some people in the company "didn't like it."
Set aside whether firing was fair. Look at what was measured: whether someone had set up an account. Not whether the code got better, not whether incidents went down. Onboarding.
Version 3: Meta Got a Leaderboard
In April 2026, The Information reported, and Fortune followed up, that a Meta employee had built an internal dashboard called "Claudeonomics." It ranked token usage across the company's 85,000+ employees and showed the top 250 users with titles. Meta said the employee took it down on their own. It was gone two days after the story broke.
The practice now has a name: tokenmaxxing. Treating AI consumption as a sign of productivity.
LeadDev asked engineering leaders about it. Ankit Jain, founder of the Hangar developer-experience community, called it a return to the pre-DORA era of measuring lines of code, and added: "You could just use tokens to run your OpenClaw!"
Honeycomb's SVP of Engineering, Emily Nakashima, said something any finance team should hear: "I really worry about companies trying to 2x or 3x their token spend, because there are ways engineers can do that that return no value to the company."
Three companies, one sentence, and three different things measured: readiness, compliance, and consumption. None of them is an outcome.
Version 4: Klarna Skipped the Memo and Went Straight to Replacement
One company went further than a mandate.
Klarna's CEO, Sebastian Siemiatkowski, spent 2024 publicly saying that AI could already do jobs humans do. According to a Fast Company piece by Jonathan Corbin, CEO of AI customer-service company Maven AGI (so keep in mind he has a stake in this argument), Klarna paused hiring for more than a year, went from about 5,500 employees to 3,400, and promoted a chatbot it said was doing the work of 700 customer service agents.
In May 2025, Bloomberg's headline read: "Klarna Slows AI-Driven Job Cuts With Call for Real People." Siemiatkowski was planning a hiring drive so that customers would always have the option of talking to a person.
The detail engineers should notice is in Corbin's account: when service quality slipped, Klarna asked software engineers, designers and marketing staff to help answer customer inquiries.
So the most aggressive version of the AI vision ended up with engineers doing support shifts. That isn't a point against AI. It's a point against putting AI in without redesigning the work around it, which is exactly what the research in the next section describes.
Why Mandates Measure the Wrong End
Mandates measure inputs because inputs are easy to count. Whether the input turns into anything useful is a separate question, and the research on that is sobering.
MIT's NANDA initiative (150 leader interviews, 350 employees surveyed, 300 public deployments analysed) found that only about 5% of enterprise generative AI pilots achieve rapid revenue acceleration. The report blamed a "learning gap" in how organizations integrate the tools, not model quality. Purchased tools from specialist vendors worked about 67% of the time. Internal builds worked about a third as often.
Google's 2025 DORA report, based on nearly 5,000 technology professionals, found that 90% now use AI at work and more than 80% believe it has increased their productivity. But:
- AI adoption still correlates with higher software delivery instability.
- It showed no measurable impact on workplace friction or developer burnout.
- Some organizations saw work intensification, where perceived gains simply raised output expectations.
DORA's summary line is worth remembering: AI works as a mirror and a multiplier. It amplifies what's already there. A company with good foundations gets faster. A company with chaotic ones gets chaotic faster.
That explains the four versions above. Shopify's memo multiplied years of groundwork. A Saturday deadline or a leaderboard has much less underneath it to multiply. And a chatbot replacing 700 agents multiplied a support process that nobody had redesigned first.
The "Resistance" Numbers Aren't What They Look Like
This is the part of the data I find most provocative.
In the WRITER 2026 survey, 29% of employees admitted to "sabotaging" their company's AI strategy. Among Gen Z it was 44%. That sounds like a revolt until you read the definition. The examples include entering company information into public tools and using unapproved tools, alongside refusing to use AI.
So some of the people counted as "sabotaging" AI are using more AI than their company approved. MIT's report points the same way, noting widespread "shadow AI" use of unsanctioned tools like ChatGPT.
My read, and it's a theory: a lot of what leadership calls resistance is engineers rejecting the company's version of AI (slow approvals, the wrong tool, a leaderboard) while happily using their own. That's a different problem from people refusing AI, and it needs a different fix.
How many engineers do you know who complain about the official Copilot rollout and then open Claude or ChatGPT in another tab?
What "Implemented" Should Mean
DORA identified seven capabilities that amplify AI's benefits. The first one on the list is not a tool. It's "a clear and communicated AI stance." Explicit policies, rather than vague guidelines that leave developers unsure what's allowed.
The rest are unglamorous: healthy and AI-accessible internal data, strong version control, working in small batches, a user-centric focus, and quality internal platforms. The report adds a warning: teams without a user-centric focus actually saw negative effects from AI adoption.
Here's a test you can run on your own org. Ask five engineers to explain your company's AI policy in one sentence. If you get five different answers, or "I think we're allowed to use Copilot?", you have a memo, not an implementation.
The Takeaway
A mandate tells engineers what to do on Monday. Implementation is whatever still holds on the Friday of an outage.
Tell me about yours:
- Did your company send "the memo"? What actually changed in the 90 days after?
- What does your org measure: accounts, tokens, PRs, or outcomes?
- Honest question: are you using AI tools your company hasn't approved? Why?
Next in the series: why a startup engineer, a mid-size SaaS engineer, a Big Tech engineer and an IT services engineer are effectively doing four different jobs with the same AI.
Sources
- First Round Review, From Memo to Movement: Shopify's Cultural Adoption of AI (July 2025) · original memo on X
- TechCrunch, Coinbase CEO explains why he fired engineers who didn't try AI immediately (Aug 2025)
- Fortune, A Meta employee created a dashboard so coworkers can compete to be the company's No. 1 AI token user (April 2026)
- LeadDev, Tokenmaxxing and the search for AI metrics that matter (April 2026)
- Bloomberg, Klarna Slows AI-Driven Job Cuts With Call for Real People (May 2025, paywalled)
- Fast Company, Klarna tried to replace its workforce with AI (contributor opinion piece by Jonathan Corbin, CEO of Maven AGI)
- Fortune, MIT report: 95% of generative AI pilots at companies are failing (Aug 2025)
- Google DORA, 2025 State of AI-assisted Software Development · summary via IT Revolution
- WRITER / Workplace Intelligence, 2026 AI Adoption in the Enterprise
Top comments (3)
The shift from tracking lines of code to tracking token volume is just Goodhart's law changing its currency. When leadership treats compute burn as proof of velocity, the incentive for the team is to offload validation to the model instead of doing it upfront. That shows up on the quarterly ledger as higher SaaS spend, but the real cost lands during an outage when nobody on call has the mental model to debug what the agent generated.
Thanks, Dean — "Goodhart's law changing its currency" is exactly the frame. That's the spine of the piece: Shopify measured readiness, Coinbase measured compliance, Meta measured consumption, and none of them is an outcome. The DORA 2025 data you'd expect to disagree actually agrees with you — 90% use AI, but adoption still correlates with higher delivery instability, and your outage-night mental-model point is what Honeycomb's Nakashima meant by token spend that "returns no value to the company." Appreciate you reading it closely.
Memo ra đời dễ, nhưng biến nó thành thói quen hàng ngày của team mới là bài toán khó. Ở chỗ mình, leadership cũng từng "ban hành chỉ thị dùng AI" — kết quả là tháng đầu everyone mở Copilot/ChatGPT thử vài lần, rồi quay về workflow cũ vì: (1) prompt engineering chưa có best practice chung, ai cũng tự mò; (2) code review vẫn bắt convention cũ, PR dùng AI sinh ra bị nitpick style thay vì logic; (3) không có metric nào đo "AI adoption" ngoài vibe.
Điều thực sự chuyển hóa được là khi engineering managers ngồi xuống viết internal playbook cụ thể: pattern prompt cho từng loại task (refactor, test, docs, migration), convention review cho code AI-generated, và quan trọng nhất — allocate 10% sprint capacity cho "AI experimentation" có structure, không phải "tự tìm hiểu". Six months sau, velocity đo được tăng ~18% trên task lặp, nhưng complexity task vẫn do human lead.
Memo chỉ mở cửa. Culture, tooling, và incentive mới giữ cửa đó mở PS: the tool I meant is on labagent .tech