The Engineer Who Trained His Replacement
The Samsung workers who got a share of this had to threaten a strike to get it in writing. I’ve been tracking the pattern on X @MangoAggro if you want the running commentary.
There’s a version of this story that makes the company look incompetent.
They hired a senior engineer to build internal tooling. He spent two years on it. The tooling worked. Within eight months, his role was redundant and his team was cut from six to two.
The incompetent version says: they didn’t plan for this. They didn’t think through the consequences. Classic corporate short-termism.
That version is wrong.
They planned for exactly this. The engineer built what they asked him to build. The productivity gain landed where it was supposed to land. His role was always going to compress once the system ran without him. Nobody lied. Nobody made a mistake. The system worked as designed.
That’s what makes it worth examining.
The productivity goes somewhere
In April, Anthropic published a survey of 81,000 Claude users. The average user reported being “substantially more productive.” Ninety percent said the gains stayed with them, not the company.
Read that alongside what GM reported the same quarter. They beat earnings estimates by 40%. They also cut another 500 to 600 IT workers in Texas and Michigan. In total, GM has eliminated over 10,000 white-collar salaried roles since 2022.
Both numbers are true. The Anthropic survey and the GM layoffs are not contradictions. They are the same equation viewed from two different positions in the org chart.
The worker experiences the productivity gain. The company captures the headcount reduction. Those are two separate transactions and they don’t happen at the same time or to the same people.
The senior engineer who is 40% more productive this year is not getting a 40% pay increase. He’s getting a performance review that says he’s performing well. The company is getting a business case to not replace the two people who left his team last quarter.
This is the redistribution problem in its most practical form. Not an abstract policy debate. A spreadsheet decision made by someone in finance who has never met the engineer in question, working from headcount targets approved in a planning meeting three months ago. The engineer isn’t in that meeting. His productivity numbers are.
Who is running the AI taxation conversation
Governments in the EU, UK, and parts of the US are now actively discussing AI taxation. The proposals vary. Some target compute. Some target companies above a headcount reduction threshold. Some propose redirecting a share of productivity gains into retraining funds.
The proposals share a structural problem. They’re written for the displacement that happened in 2023. This newsletter called the second wave in Issue 9: cuts between late 2026 and early 2028 at companies that announced AI restructuring without completing deployment. By the time any of this legislation passes, that wave is already over.
The consultation process itself is worth examining. When governments open AI taxation consultations, they invite submissions from affected parties. The parties with the most resources to submit detailed, technically credible responses are the same companies that will be taxed. The parties most affected by the displacement are represented by unions where they exist and by nobody where they don’t. The outcome of a consultation shaped primarily by the entity being consulted about is predictable.
The other structural problem is the measurement gap. A lot of the displacement doesn’t show up cleanly. Roles that stop being backfilled. Teams that quietly compress over 18 months. Junior positions that disappear without a layoff announcement. You can’t tax what you can’t see and the most significant displacement in knowledge work right now is designed to be invisible.
This is not an accident. When a company announces a layoff, it generates headlines, regulatory scrutiny, and sometimes legal exposure. When a company simply stops hiring for roles made redundant by AI, none of that happens. The headcount compresses silently. The people affected aren’t laid off. They just don’t get hired. Their jobs never existed in a way that generates a statistic.
The companies lobbying hardest against AI taxation this year cited AI efficiency in their last round of layoff announcements. That is not ironic. That is a coherent position: we captured the gains, we are not interested in redistributing them, and we have the resources to ensure the policy conversation moves slowly enough for the fait accompli to be complete.
The one deal that worked
In May 2026, Samsung and its union reached an agreement after months of standoff. Seventy-four percent of workers voted yes. The deal gave workers a share of AI chip profits, a pool covering 78,000 employees, structured around operating profit thresholds over ten years.
It worked because the workers had direct leverage over what the company was selling. Samsung makes the chips that power AI. The workers in the chip division could threaten to stop making them. That leverage is specific, credible, and hard to replace with tooling.
Most tech workers don’t have that. A senior software engineer’s leverage is real but diffuse. The company can absorb losing one person. It cannot absorb losing the workers who make the physical infrastructure that the whole AI economy runs on.
The Samsung deal is the floor, not the model. It required a near-strike to get in writing. It required workers with direct leverage over an irreplaceable part of the supply chain. Most of the people reading this don’t have either of those things.
There is one other case worth examining. In April 2026, the Intermediate People’s Court of Hangzhou ruled that a company could not cut a worker’s salary simply because AI had absorbed his role. The court applied existing Chinese labor law: AI adoption doesn’t qualify as the kind of extraordinary event that justifies unilateral changes to employment terms. No new legislation. Standard labor protections applied to an AI case.
That’s a different mechanism than taxation. It doesn’t redistribute the gains. It says the cost of the automation decision stays with the employer, not the worker. Small protection in practice. But a different way of thinking about where the obligation sits.
What the engineer actually did wrong
Back to the engineer who trained his replacement.
He made one mistake and it wasn’t technical. He built a system that ran without him and then stayed in the role that built it rather than moving to the role that owned what the system produced.
The system generated output. Someone had to be accountable for whether that output was good. Someone had to make judgment calls when the system produced something wrong or incomplete. Someone had to sit in the meetings where the output was used and be the person who caught the problems before they became expensive.
That’s an outcome job. He was doing a process job and he automated the process. He should have automated himself into the outcome role before someone else noticed he’d done it.
The process versus outcome distinction is the frame this newsletter returns to because it’s the one that actually holds. A process job executes known steps to produce a deliverable. AI does this well and gets better at it every quarter. An outcome job owns a result you’re held accountable for when it breaks. AI assists. The human owns it. Those two things are not the same and the companies currently cutting roles understand the distinction even if they don’t use the same words for it.
I’ve watched this happen to people I know. Not the dramatic version. The slow version where the role doesn’t get eliminated. It just stops growing, stops getting backfilled when someone leaves, stops coming up in planning meetings. You’re still there. The role just has less gravity than it used to. That version is harder to fight than a layoff notice.
The honest read on policy
AI taxation will arrive. Some version of it will pass in some jurisdictions in the next five years. It will be underfunded, partially captured by the companies it’s meant to tax, and four to five years behind the displacement curve it was designed to address.
That’s not cynicism. That’s the track record of every previous attempt to legislate the distributional consequences of a technology wave in real time. The EU’s GDPR took six years from proposal to enforcement. By the time it had teeth, the data practices it was designed to prevent were already the infrastructure of the entire industry.
AI taxation will follow the same arc. The gains will have been distributed, the roles will have been eliminated, the new equilibrium will be established. Then the legislation will arrive to govern a situation that has already resolved itself in favor of whoever moved fastest.
The person reading this newsletter has a career window that doesn’t overlap with that timeline. Policy is not the map. The map is something else.
Paid subscribers get the specific move below.
The founding tier is still open. First 50 lock in the rate forever. Link below.
Keep reading with a 7-day free trial
Subscribe to The Displacement Audit to keep reading this post and get 7 days of free access to the full post archives.