Companies Replaced Junior Developers With AI. Here Is What Happened Next.
AI didn't just change how we work. It changed who gets hired, who gets fired, and how skills are built. The data is finally starting to reveal the consequences.
Senior Developer
In 2023, Klarna CEO Sebastian Siemiatkowski went on record with a claim that made headlines around the world.
His AI chatbot, built with OpenAI, was now doing the work of 700 full-time customer service employees. Queries resolved faster. Costs down. Efficiency up. The investors loved it.
In early 2025, Siemiatkowski went on record again.
"Cost unfortunately seems to have been a too predominant evaluation factor. What you end up having is lower quality."
Klarna is now rehiring humans.
That arc — confident announcement, quiet reversal, reluctant admission — is the story of what happened when the industry tried to replace humans with AI at scale. Not in every company. Not in every role. But enough to see the pattern.
Here is what the data actually shows.
The numbers first
Entry-level developer job postings dropped 60% between 2022 and 2024. Tech unemployment jumped from 3.9% to 5.7% in a single month in early 2025. Internship postings in tech dropped 30% since 2023. Google and Meta hired roughly 50% fewer new graduates compared to 2021. Salesforce announced it would halt junior hiring for 2025 entirely.
The experiment was real and large-scale. Millions of dollars in wages were cut. Hundreds of thousands of entry-level positions evaporated.
And the consequences are arriving now.
What actually broke at Klarna
The Klarna case is worth studying in detail because the company was unusually public about both sides of the story — the initial triumph and the subsequent reversal.
In 2024, Klarna's AI handled two-thirds of all customer service chats. Average conversation time dropped from 11 minutes to 2 minutes. The company projected a $40 million improvement in profit.
Six months later, customer satisfaction scores were falling. Repeat inquiries were climbing. The CEO was asking software engineers and marketing staff to help answer customer inquiries.
What broke was not the easy stuff. The AI handled routine queries exactly as advertised — fast, consistent, cheap. What broke was everything else. Emotionally charged interactions. Complex multi-step problem resolution. Edge cases that fell outside the training distribution. The cases where a customer had three problems at once, was frustrated, and needed someone to understand the situation rather than match it to a category.
Full AI replacement of customer service failed on quality, not cost: Klarna's AI agents handled the volume but not the complexity. Customer satisfaction scores dropped as edge cases, emotionally charged interactions, and multi-step problem resolution overwhelmed AI trained to handle routine queries. Rehiring costs exceeded the original savings estimate — reversing the layoffs required recruiting, onboarding, and training new customer service staff, an expensive process companies rarely model in their AI replacement business cases.
The cost savings calculation missed the cost of unwinding a failed replacement. Almost every company running this experiment made the same omission.
The problem that nobody is talking about loudly enough
Here is what makes the developer case different from customer service.
When you replace junior developers with AI and ask senior developers to review the output, you are not just changing a workflow. You are removing the pipeline that creates future senior developers.
Senior developers don't appear from nowhere. They were junior developers once. They learned by doing the exact boring boilerplate work that AI now handles — writing CRUD code, fixing small bugs, working through problems slowly and making mistakes and understanding why those mistakes happened. A developer who spent three years reviewing AI code has not become a senior developer. They have become a very experienced reviewer of AI code.
The skill is not in producing the output. The skill is in the process of struggling with the problem, hitting the dead ends, and understanding why certain approaches fail. Reviewing AI output gives you none of that.
An Anthropic randomized controlled trial published in January 2026 made this concrete: developers who relied on AI tools to write code scored 17% lower on coding comprehension tests than those who coded manually. Nearly two letter grades lower. The AI raised their output speed. It reduced their understanding of what they were outputting.
That gap between production speed and actual comprehension is where codebases go wrong quietly.
The codebase problem is already showing up
Code churn — the amount of code that gets rewritten or deleted within two weeks — has doubled. Duplicate code is up 4x because AI doesn't refactor, it copy-pastes patterns. Up to 30% of AI-generated code snippets have security issues — SQL injection, XSS, authentication bypass. Technical debt is compounding at a rate not seen before.
A March 2026 academic study analyzed 1,154 posts across Reddit and Hacker News threads about AI-assisted development. It found three consistent patterns among teams that had adopted AI heavily: review friction (AI output required more time to review, not less), quality degradation (architectural coherence declined over time), and skill atrophy.
One incident from the study: a designer used AI to build a full React application. The code was a mess so they hired a freelancer to fix it. He deleted 90 of 100 files.
Another: a senior developer asked a junior who had submitted an AI-generated PR whether they understood what their code did. They didn't.
A senior engineer quit voluntarily — not because they found something better. Because they watched their company fire eight junior developers and replace their work with GitHub Copilot. Then watched management ask the remaining senior engineers to supervise AI output for the workload of eleven people. Then watched the codebase slowly become a disaster nobody fully understood anymore. He said he didn't recognize the job.
That is not an isolated anecdote. It is a pattern. And it compounds.
The pipeline problem has a delayed fuse
Here is the part that will matter most in 2028 and 2029.
The junior developers who did not get hired in 2024 and 2025 are not going to become senior developers in 2027 and 2028. The pipeline that has historically produced experienced engineers — years of writing real code, making real mistakes, building real systems — was interrupted at scale.
Ten years from now, who are the senior developers reviewing AI output if nobody got hired as a junior in 2025?
Entry-level developer postings dropped 60% between 2022 and 2024. Stanford Digital Economy Lab research showed entry-level developer employment for people aged 22 to 25 dropped nearly 20% from its peak. The pipeline for future senior developers was disrupted at scale — and nobody seems to have noticed yet.
The consequences of that disruption will not be visible for several years. By the time they are visible — when companies find themselves short of experienced engineers who actually understand their systems — the gap will be years wide and expensive to close.
What is actually working
Not all of this is doom. Some companies are navigating it better.
The ones doing well share a specific approach: they are using AI to accelerate junior developers, not replace them. The junior developer's role changes — less boilerplate, more ownership of what the AI produces — but the learning pipeline stays intact.
Companies are reshaping onboarding programs with modules like "How to Work with AI Assistance" and pairing juniors with mentors who specifically review AI-generated code with them. Tech lead Enrico Piovesan puts it this way: "Today's juniors are not yesterday's juniors. They emphasize breadth over depth and orchestration over authorship."
A junior in 2026 who understands what good architecture looks like, can prompt an AI to produce a first draft, and can evaluate and improve that draft is genuinely more productive than a junior who writes everything from scratch. The skill set has shifted, not disappeared.
The companies cutting junior positions entirely are making a short-term optimization with a long-term cost they have not yet paid.
The honest version of what this means
The narrative that AI would simply replace junior developers and everything would be more efficient was wrong in a specific way: it modeled AI as a substitute for human output rather than as a change in how humans develop.
Code is not the product. Engineers are the product. Companies that understood this kept their junior pipelines. Companies that didn't are going to feel the consequence.
Klarna's CEO, who in 2024 said "AI can already do all of the jobs that we as humans do," said in 2025: "Really investing in the quality of the human support is the way of the future for us."
That reversal took two years and cost more than the original savings.
The experiment ran. The results are in. The companies reading them correctly are the ones keeping their junior developers — while making sure those developers are learning how to work alongside AI rather than being replaced by it.
Has your team changed how you hire or onboard junior developers because of AI? What worked and what didn't? The ground-level experiences are more useful than any think piece.
Sources:
Comments (0)
Login to post a comment.