AI Isn't Replacing Software Engineers. It's Pricing Their Judgment.

A few months ago, in the same week, I read two headlines that had no business existing at the same time.

One was about a company bragging that its engineers had basically stopped writing code — AI was doing it for them. The other was about a different company quietly rehiring the senior engineers it had laid off a year earlier, because the AI-built version of their product kept shipping bugs a human would have caught in code review.

I run mock interviews for engineers trying to land their next role, and I teach backend development on this channel, so I hear both sides of this panic constantly — from people worried AI is about to make their job irrelevant, and from hiring managers quietly telling me the opposite. I wanted to stop going off vibes and actually dig into what was happening with real numbers. What I found didn't fit neatly into either the "AI is taking our jobs" camp or the "AI is overhyped nonsense" camp. It's stranger than both, and honestly more useful to understand than either.

Two completely opposite stories, same twelve months

Start with the side that got all the headlines. Salesforce's Marc Benioff said his company was "seriously debating — maybe we aren't going to hire anybody this year." Spotify's co-CEO told analysts his best developers hadn't written a single line of code since December. Google's Sundar Pichai said more than a quarter of Google's new code was already AI-generated. If you only read tech Twitter, you'd think the engineering job was basically over.

Now the side that barely made the news. Ford quietly rehired 350 veteran engineers after AI-driven design tools let quality problems slip through that a human reviewer would have flagged immediately. IBM tripled its entry-level hiring — not despite the AI, but because of exactly what the AI couldn't do. Its own HR assistant handled 94% of requests just fine, and then hit a wall on the judgment calls in the remaining 6% — the part that actually required a person. Commonwealth Bank reversed its AI-driven job cuts and publicly admitted it should have been more thorough before deciding which roles it actually needed. And Klarna's CEO, who had spent a year as the poster child for AI-driven layoffs, said it plainly: "We went too far."

Both of these are real. Both happened in the same window of time, often inside companies in the exact same industry. So the honest question isn't "is AI replacing engineers, yes or no." It's: why did so many smart, well-resourced companies get this wrong in the exact same direction, and what did the ones who corrected course actually learn?

Why 2025 looked like a free lunch

I don't think the executives making these calls were stupid. For a while, the numbers genuinely supported them. Model costs were falling month over month. Coding benchmarks were climbing in a way that felt almost rigged — models went from solving under 5% of real-world coding tasks to over 70% in about a year. If you were looking at a spreadsheet of capability-per-dollar, full automation looked like the obvious next move. A lot of smart people made that bet because the trend line told them to.

What the trend line didn't show was the bill that shows up after you actually deploy the thing. Gartner found that agentic AI workflows — the kind that don't just answer a question but actually carry out a task end to end, the way you'd want an AI "replacement" to — burn 5 to 30 times more tokens than a single chatbot reply. That's the part that got buried under all the "cost per token is falling" headlines. Your price per token can keep dropping while your actual bill climbs, because the task itself got more expensive to run, not cheaper. For a lot of teams, that's exactly what happened: the sticker price fell and the invoice went up.

Even the people measuring this got fooled

Here's the part of this research that actually changed how I think about AI tools, and it's not a corporate anecdote — it's a controlled study. METR, a group whose entire job is measuring AI capability rigorously instead of guessing from vibes, ran a randomized trial with experienced open-source developers using AI coding tools on real tasks. The result: those developers were 19% slower with the AI tools than without them.

That's not even the strange part. After finishing the tasks — after living through exactly how long everything took — those same developers still believed the AI had made them about 20% faster. Not a little off. Backwards. The tool didn't fail; the estimate of the tool failed, because nobody had accounted for the ramp-up time it takes a team to actually get good at using something new, versus how long everyone assumed that ramp-up would take. If experienced developers in a controlled study can be this wrong about their own output, it's a reasonable bet that a lot of the "AI made us 10x faster" claims floating around are measuring the same illusion.

Then there's the cost that never makes it into a sprint demo: security debt. Veracode found that 45% of AI-generated code samples failed basic security tests — nearly half. CodeRabbit's analysis of real pull requests found AI-written code carrying up to 2.74 times more security issues than human-written code in the same repositories. And Black Duck's 2026 open source security report found vulnerabilities per codebase had jumped 107% year over year — more than doubling. None of that shows up when the demo goes well in front of leadership. It shows up six months later, in an incident report, written by the engineer who got rehired to deal with it.

This isn't the first time we've heard this story

If any of this feels familiar, it should. Every decade or so, the industry convinces itself that some new shift means engineers are about to become optional — offshoring was going to do it in the 2000s, no-code platforms were going to do it in the 2010s. Both of those did change the job. Neither one eliminated the need for people who understood systems well enough to know when something was quietly wrong. AI is the biggest version of this pattern yet, mostly because it's genuinely more capable than anything before it — but the lesson from the last two cycles still holds. The tool changes who does which part of the work. It doesn't remove the need for someone to be responsible for the outcome.

The profession isn't shrinking. It's splitting.

Here's the part I find genuinely encouraging, and I mean that especially if you're early in your career and this whole topic has been stressing you out: the job isn't disappearing. It's dividing, and fast. Software engineering job openings actually rose about 30% in 2026, climbing past 67,000 listings — the highest in three years, in the same quarter that saw over 52,000 tech layoffs announced. Read that twice. Layoffs and open roles both spiking at once isn't a contradiction — it's the split showing up in the hiring data directly.

Almost all of that new demand is for people who already have judgment: engineers who can review AI output and catch what's subtly wrong with it, who can make architecture calls, who know which 6% of a task actually needs a human the way IBM's HR assistant did. AI didn't shrink this profession. It made experience worth more, and it made the entry-level rung a lot harder to reach — because the easiest, most mechanical parts of the job, the parts junior engineers used to learn on, are exactly what AI does fastest.

Gartner's own forecasts back this up from a different angle. They expect over 40% of agentic AI projects to be canceled by the end of 2027 over cost and unclear ROI — a direct echo of the token-cost problem above. Separately, they project that 80% of the engineering workforce will need to upskill through 2027 just to stay relevant — not because their jobs are vanishing, but because the job itself is changing shape underneath them. And after watching the research show that 55% of companies that cut staff for AI now admit they regret it, I don't read any of this as doom. I read it as the correction already happening in real time, company by company.

So what do you actually do with this?

If you're an engineer reading this and wondering what it means for you, here's what I'd actually tell someone in a mock interview right now:

If you're early career, don't try to out-type the AI — you'll lose, and it's the wrong contest anyway. Spend your energy learning to read code critically: why a pull request is wrong, not just whether it runs. That's the skill that just got more valuable, not less.

If you're mid-career or senior, lean into exactly the stuff this article is full of — architecture decisions, security review, knowing when an elegant-looking solution is going to be an incident six months from now. That judgment is the entire reason Ford, IBM, and Commonwealth Bank went back and rehired people. It's the one part of the job that didn't get cheaper.

And if you're hiring or making the call on an AI rollout yourself, ask the question these companies apparently didn't ask loudly enough before cutting staff: not "can AI do this task," but "what does it cost us when AI gets the 6% wrong, and who's left to catch it?"

The bottom line

AI isn't replacing software engineers. It's pricing, in public, what engineering judgment was always worth. The companies that fired first are the ones now paying rehire premiums to buy that judgment back — and the ones who never cut it in the first place just quietly got a head start.

If you found this useful, the video above walks through all of this with the actual charts and numbers on screen, and I'd genuinely appreciate the subscribe if you want more breakdowns like this on backend engineering and AI.