Every few weeks someone on a leadership team asks whether AI will replace the DevOps engineers. It is the wrong question for the person holding the hiring budget. The better one is what the engineers you do hire now spend their day on, and whether your interviews test any of it.
DevOps Society's September hiring report found AI mentioned in nine out of ten infrastructure job postings. So the job adverts have moved. In most companies the interview loop has not.
What the data says the job has become
Indeed's Hiring Lab published a breakdown of seniority in US postings in July. In software development, senior roles made up 69.3 percent of postings in the first quarter of 2026, and entry-level roles just 4.5 percent. Since 2019 the senior share has risen about nine points, mostly at the expense of mid-level roles. Companies are paying for experience, which means they are paying for judgement. (We covered what this does to the junior pipeline in "Five Postings Out of 516: Nobody Decided to Stop Training Juniors", so we will not repeat it here.)
Then look at what that judgement is being spent on. New Relic surveyed 200 US technology leaders for its 2026 State of AI Coding report, released in June. Of those, 94 percent said AI-generated code looked higher quality than human code at review time. Yet 78 percent reported more incidents after deploying it, and 86 percent said senior staff were spending more time fixing code. It is a small vendor survey, so treat the exact numbers loosely.
Google's 2025 DORA report, based on nearly 5,000 technology professionals, found AI use now goes with higher delivery throughput but that delivery instability keeps rising. Stack Overflow's 2025 developer survey of around 49,000 people found 66 percent frustrated by AI answers that are almost right but not quite, and 45 percent saying debugging AI-written code takes longer.
Code arrives faster and looks tidier. The work that remains is deciding whether it is safe, finding out why it broke, and owning the result.
What has become more valuable
Five things are worth more in 2026 than they were three years ago.
Debugging a system you did not build. AI-written changes mean more engineers will be on call for code nobody on the team wrote line by line. The person you want can form a hypothesis, check it against logs and metrics, and throw it away when the evidence says so.
Judging a change before it ships. This is code review with higher stakes. Can the candidate spot that a Terraform change will destroy and recreate a database, or that a retry loop has no back-off? Can they say why a tidy-looking diff is still wrong?
Cost awareness. The FinOps Foundation's State of FinOps 2026 report, from 1,192 practitioners, found 98 percent of FinOps teams now manage AI spend, up from 31 percent two years earlier, and practitioners asked for tools that cost an architecture before it is deployed. That cost-before-build thinking has to live in engineers' heads, not just in a finance dashboard.
Incident judgement. Knowing when to roll back rather than fix forward, when to wake someone, and when to stop chasing the root cause and restore service first.
Writing. Postmortems, design notes, runbooks and the description on a pull request. When much of the code is generated, the human explanation of intent is what the next engineer has to work from.
What has become less useful to test
Memorised syntax. Nobody needs to recall exact kubectl flags when an assistant produces them in seconds. Such questions mostly measure who crammed the night before.
Tool trivia. "What is the difference between a DaemonSet and a StatefulSet" tells you someone read the documentation. It does not tell you whether they would notice a misconfigured one taking down a node pool at 3am.
There is also a practical reason to drop these. Gartner predicted last year that by 2028 one in four candidate profiles worldwide could be fake, and 6 percent of 3,000 candidates it surveyed admitted to some form of interview fraud. Recall questions are exactly what a hidden assistant answers best.
Interview formats that show the real thing
Two formats do most of the work.
Review an AI-written pull request. Generate a realistic change against a small sample repository: a Helm chart update, a CI pipeline edit, a Terraform module. Seed it with two or three real problems, one obvious and one subtle. Give the candidate 30 to 40 minutes and any tools they like. You learn what they question and whether they trust output they cannot explain.
Live incident walkthrough. Show a dashboard, a few log lines and an alert from a scenario you have actually lived through. Ask what they would check first, and reveal more as they ask. Stop halfway and ask what they would tell the incident channel right now. That last question tests writing and judgement together.
Allowing AI in the room is no longer unusual. Canva switched its engineering interviews to require AI tools in 2025 and said candidates without AI experience often struggled because they could not steer it, not because they could not code. Meta began testing AI-assisted coding interviews later that year, partly because it makes covert AI use less useful.
The mistake of copying a FAANG loop
Many smaller companies borrow the loop of a very large one: several algorithm rounds, a systems design round, a behavioural round. Those loops were built to screen huge numbers of applicants for general software roles, not to find someone who can keep a 40-service Kubernetes estate running with a team of four.
A platform or SRE hire at a mid-sized company needs breadth, comfort with mess, and the ability to explain trade-offs to product teams. An algorithm round tells you almost nothing about those. It also filters out experienced operators who have not practised puzzle questions in years, at exactly the moment senior people are what the market is short of.
The honest caveat
None of this is settled science. The New Relic numbers come from a vendor survey of 200 leaders. The DORA and Stack Overflow figures describe developers broadly, not infrastructure engineers specifically. Realistic tasks also take longer to design and are harder to score consistently than a puzzle, and a poorly built scenario can reward familiarity with your stack rather than ability. Without a scoring guide, the new format becomes a vibe check.
What to do this quarter
Pull your current interview questions into one document and mark each as recall or judgement. If more than half are recall, you have found your first job.
Build one AI-written pull request exercise from a real change your team shipped, with the problems you actually caught in review. Run it on two current engineers before any candidate sees it.
Turn your last good postmortem into a walkthrough scenario. You already have the timeline, the dashboards and the decisions. Write down what a strong answer sounds like before you use it.
Decide your AI policy for interviews and tell candidates in advance. Allowing it openly is easier to enforce than banning it.
