Written by: Agata Wojtas, Chief Commercial Officer, Digital Colliers
The hook on LinkedIn was short so here's the longer version. When first drafts of code got cheap, the bottleneck moved. It didn't disappear. It just walked one room over, and a lot of hiring pipelines are still knocking on the old door.
If you run an engineering team in 2026, the risk isn't that you hired too slowly for AI. The risk is that you hired for the wrong shape of person, kept doing it for two years, and now your team can produce a mountain of plausible code that nobody on the inside can vouch for.
The skill that got cheap
Writing a first pass of a function used to be the job. Now a model does it in eight seconds. Fine. The problem is what that first pass costs downstream.
One recent empirical study found that more than 15% of commits from every mainstream AI coding assistant introduce at least one issue, ranging from 17.4% for Copilot up to 29.1% for Gemini. The same body of work tracked unresolved technical debt from AI-generated code climbing from a few hundred surviving issues in early 2025 to over 100,000 by February 2026. That debt didn't appear because the tools got worse. It appeared because teams got faster at merging things nobody deeply read.
And it shows up in the humans too. A study of experienced developers using AI assistants found they reviewed 6.5% more code after adoption but their own original code productivity dropped 19%. Read that twice. The mix of the job changed. Less authoring, more judging. If you hired for authoring, you now have people doing a job they weren't screened for.
What we actually screen for now
The skills that got scarce aren't glamorous. They're the ones that were always undervalued because they were hard to see on a CV.
- Reading code you didn't write, at speed, and forming a real opinion on it.
- Knowing when to throw a working draft away because the shape is wrong.
- Naming the thing the model got subtly incorrect, not the thing it got obviously incorrect.
- Holding a system in your head well enough to spot when a change breaks an assumption three files over.
- Writing tests that would actually catch a plausible-looking hallucination.
- Saying no to a PR that passes CI but shouldn't ship.
None of that is new. What's new is that these skills used to be a nice bonus on top of raw output. Now they are the output. The typing is done by something else.
The left-behind risk
Here's the shape of the trap. Most AI initiatives don't survive contact with production anyway. IDC reports that 88% of AI proofs-of-concept never reach widescale deployment, and only four out of every 33 pilots graduate. S&P Global's 2025 survey found 42% of companies abandoned most of their AI initiatives, up from 17% the year before, with the average org scrapping 46% of POCs before production.
Those numbers get quoted as an AI problem. They're not, or at least not only. They're a judgement problem. Somebody had to look at a working demo and decide whether it was actually good enough to run against real users, real money, real regulators. When your team is optimised to produce demos and under-optimised to judge them, you get the S&P Global chart.
The teams that get left behind in the next two years won't be the ones that were slow to adopt. They'll be the ones that adopted fast, hired for the old skill mix, and ended up with a pile of shipped-looking work that nobody senior can defend. That's a hard team to fix, because the fix is people, and the people you need weren't in the pipeline.
Signals in an interview
A few things that are working, from the pattern I keep seeing across teams that ship real AI-adjacent systems:
- Give the candidate AI-generated code with a subtle bug and ask them to talk through it. Not find it. Talk through it.
- Ask them about the last time they threw away a week of work. If they can't remember, that's the signal.
- Ask what they'd refuse to ship even if the tests passed. The answer tells you whether they've been burned yet.
- Look for people who have shipped in regulated or security-sensitive contexts. Only about 3-5% of publicly disclosed vulnerabilities get patched within 30 days, and the people who lived through that reality tend to have the judgement muscle already built.
The old screen selected for speed. The new screen selects for taste. If your loop still rewards the first, you're building a team for a job that stopped existing.