In May 2025, Dario Amodei told Axios that AI could wipe out half of all entry-level white-collar jobs and push unemployment to 10–20% within one to five years.
Many engineers laughed. "It writes buggy code." "It can't design systems." "It will never understand our domain."
A year later, the laughing stopped. The prediction, though, landed somewhere else.
Right direction, wrong target
Amodei looked in the right direction. Displacement is real. But he pointed at the wrong floor.
Entry-level is not where it hits hardest. Google and many other companies still open intern roles every cycle, and a share of those interns get hired. Juniors grow up with AI. They never learned to work without it. For them it is not a new tool: it is the default way to work.
The displacement is happening one floor up. It hits managers and directors who reject AI.
They built careers on coordination: planning, alignment, status, follow-ups. That layer is now largely automated. Jira drafts the tasks. Meeting tools write the summary and the next steps. Agents ship the feature. A leader who only coordinates is now competing with software that coordinates for free.
It was never about the craft
Here is the uncomfortable part.
We all know the stories from the trenches. The legacy module nobody wanted to touch. The file with a comment saying "do not refactor, it will break." The service written by someone who left six years ago. And that product shipped to millions of users, year after year.
What kept it alive? Security and compliance: the product had to be legal and safe. Good enough features. Market adoption. Customers who paid.
Not the craft.
Craft did not pay your paycheck. Shipped outcomes did.
You can tell yourself a thousand times that your code is clean and Uncle Bob would be proud. You are probably right. That is still not what pays the bills. It never was.
Quality moved into the pipeline
This does not mean quality stopped mattering. It means quality changed its place.
Today Opus 5.5 and GPT-6 Astra write production code just fine. In my own work, about 1 in 10 features built by Opus 5.5 trips a SonarQube quality gate. The other nine pass on the first run. I will share the details in a follow-up post.
Think about what that means. Quality is no longer a personal virtue you defend in code review. It is a gate a machine checks. Static analysis, tests, security scans, compliance rules: they run on every change, whoever or whatever wrote it.
The engineer who owns the gate owns quality. The engineer who only polishes code by hand is doing work the pipeline already does.
Build, fail, learn, repeat was always here
Fast delivery and failure are not new ideas. AI did not invent them.
Remember Scrum? The retrospective was built on this: learn within the iteration, do better in the next one. Proofs of concept and rapid prototyping were always part of serious engineering. Most prototypes were meant to be thrown away.
Look at it at company scale. Killed by Google lists 307 products and initiatives Google shut down. Google Reader. Google+. Stadia. Inbox. Hangouts. Each of them was built by strong engineers with real craft. Each was killed because it did not move the company forward.
It was always like that, at every level: team, product, company.
What AI changed is the cost. A prototype that took a team a quarter now takes one person a week. When failure is cheap, the organization that tries more wins. The organization that polishes one attempt loses.
The flattening
Companies see this, and they react.
Structures flatten. Pure coordination layers disappear. Spans of control grow. Leadership moves closer to production.
What companies expect from managers and directors now is different:
- Hands-on with AI. Using the tools daily, not reading reports about them.
- Shipping. Taking features to production with AI, not only approving plans.
- Leading adoption. Helping their teams change how they work, with real results.
- Judging speed and risk. Knowing what a team can deliver today, because the baseline moved.
A director who cannot do this cannot judge their own organization. They cannot tell a real gain from a demo. They cannot see where AI adds risk. They cannot set credible goals.
The pressure is clear: faster delivery, fewer layers, more ownership.
What catching up looks like
If you lead engineering, this is the practical list:
- Use AI for your own work every day. Planning, writing, analysis, code. You cannot lead a change you do not practice.
- Ship something yourself. A tool, a feature, an internal automation. Feel the real speed and the real failure modes.
- Move quality into gates. Invest in static analysis, tests and security checks. Let machines enforce the standard. Let people decide what to build.
- Measure outcomes, not activity. Lead time, adoption, customer impact. Not story points, not lines of code.
- Let go of the craft ego. Clean code is a means. The product is the end.
The real risk
The risk is not that AI replaces you.
The risk is that someone who uses AI better than you does your job faster. Maybe a peer. Maybe a new manager who never learned to work any other way. Maybe a junior who grew up with it and is ready to lead sooner than you think.
The market is not waiting.
Catch up, or get outpaced.