Are fresh grads cooked?

I’ve spent a good part of this year building things that do junior work. An agent in Copilot Studio that fields the questions people used to ask a person. Claude wired into Power BI over MCP, renaming measures, drafting documentation and sanity-checking relationships while I answer email. None of it was hard to set up. I wrote walkthroughs.

Somewhere around the third build it occurred to me that everything I was automating was something I’d have been handed in my first couple of years of work. The tedious report changes. The “can you pull this number” requests. The documentation nobody wanted to write. That used to be a job. Quite often it was the whole job.

So when someone graduating this year asks me whether fresh grads are cooked (finished, done for, no way back, if the word hasn’t reached you yet), I can’t wave it off. The honest short answer: partly, and not entirely for the reason everyone thinks.

The numbers, briefly

Start with America, because that’s where the measurement is. The Federal Reserve Bank of New York tracks unemployment for recent graduates by major, and computer science currently sits around 7%, against roughly 4% for early-career graduates overall. A year earlier the same series read closer to 6%. Whichever way you squint, it isn’t improving.

Hiring data tells the same story from the employer’s side. SignalFire, a VC firm that tracks talent movements across hundreds of millions of profiles, found that new graduates made up 15% of Big Tech hires before the pandemic and make up 7% now. At startups it’s under 6%.

The strange detail buried in the Fed data is that computer science graduates who do get hired are less likely than almost anyone else to be underemployed. The ones inside are doing proper degree-level work at proper salaries. The problem is the door, not the room.

Malaysia has the inverted problem. Headline unemployment here hit a ten-year low of 3% in 2025, and graduate unemployment is barely above that, so on paper everything is fine. The number that matters is skill-related underemployment, which DOSM puts at around 36%: more than one in three tertiary-educated Malaysians works a semi-skilled or low-skilled job.

Those two columns are not close in pay. Per DOSM’s 2024 graduate statistics, graduates in skilled work earn around RM5,700 a month, while graduates in semi-skilled work earn around RM2,550. Tech has long been the reliable escape hatch, one of the few sectors where a fresh grad could walk straight into the skilled column. The pressure from AI is landing on the escape hatch.

How much of this is actually AI

The tidy story says ChatGPT arrived in late 2022 and ate the junior jobs. The tidy story has competition.

Tech spent 2021 and 2022 hiring like money was free, because it was. When rates rose, the correction came, and the half a million or so people laid off since 2022 (per layoffs.fyi’s running tally) went straight into the applicant pool. A company choosing between a laid-off engineer with four years at a known firm and a fresh graduate isn’t really choosing. Add a decade of “learn to code” pushing CS enrolment to record highs, add offshoring, and you have an entry-level squeeze that needs no robots to explain it.

The economists are still arguing over how to split the blame. Economists at Google pointed at interest rates. Apollo’s Torsten Slok reads it as a low-hire, low-fire market squeezing anyone without a foothold. Even the team behind the most-cited study of AI’s effect on young workers, Stanford’s “Canaries in the Coal Mine”, published an update in February conceding that under the broadest set of statistical controls, the employment decline in AI-exposed occupations only turns significant in 2024, with the earlier drop likely down to other factors.

My read: AI is real, and it’s roughly the fifth item on a five-item list of problems. It’s also the only item that won’t mean-revert. Rates come down. Overhiring corrections end. Enrolment falls once the market signal gets loud enough, and in the US it already has. The models don’t get worse.

What the machines took

A graduate hire was always a subsidised apprenticeship. The company paid someone to write boilerplate, scaffold tests, draft documentation and close the tickets nobody senior wanted, and in exchange it got a competent engineer three years later. The output was mediocre by design. The boilerplate was the tuition.

Then the tuition became free. Boilerplate, CRUD endpoints, test scaffolding, first drafts of anything: this is precisely the work large language models do fastest and best. The Stanford study, built on payroll records covering millions of American workers, found a 16% relative decline in employment for 22-to-25-year-olds in the most AI-exposed occupations since late 2022, while older workers in the same occupations held steady or grew. The losses concentrate in roles where AI automates the work outright rather than augmenting it. Erik Brynjolfsson, one of the authors, put it plainly: “what younger workers know overlaps with what LLMs can replace”.

James O’Brien, a Berkeley computer science professor who advises startups, describes founders asking why they’d pay an undergraduate when the model is cheaper and quicker. His sharper observation is that neither AI code nor new-grad code is very good. The difference is the revision loop: minutes for the model, days for the graduate.

I’ve watched the same thing from the inside. When I point Claude at a Power BI model, the tasks it hoovers up are the first-week tasks: the renames, the descriptions, the measure that’s a small variation on an existing measure. Nobody asks it to design the star schema. The work still needs doing, and it still gets done. What’s gone is the salary that used to be attached to learning on it.

The ladder problem

This is the part that makes the whole arrangement unstable. If AI does the work juniors learned on, and companies stop hiring juniors, where do the seniors of 2032 come from? SignalFire’s 2026 report warns about exactly this: an industry optimising this quarter’s balance sheet into a leadership vacuum five to ten years out. The same firm’s data shows the average age of a technical hire rising by three years since 2021. Every company is quietly assuming somebody else will do the training.

What I’d tell a fresh grad

Most of what you’ll hear is either doom or a LinkedIn carousel. Here’s my attempt at neither.

Compete on verification, not production. Producing plausible code is now free, so the scarce skill is knowing when it’s wrong. Debugging. Reading code you didn’t write. Writing the test that would actually catch the failure. Anyone can generate a DAX measure in 2026; the person who notices it double-counts because a relationship filters in the wrong direction is the one who gets kept. Every hour the model saves on writing gets partly reinvested in checking, and the checking is increasingly the job.

The old deal was that you were handed tasks and grew into outcomes. The tasks are what got automated, so outcomes are what’s left to hand out, and you want to arrive already able to carry a small one. This is what a portfolio is for. One deployed thing with real users, a pipeline that’s survived six months of contact with reality, a report an actual decision depended on. Any of those beats a stack of certificates, because employers have stopped screening for potential and started screening for evidence.

Training still happens, just not where the queue is. Big Tech’s graduate intake is a lottery at 7% of hires. Meanwhile the unglamorous employers, the insurers and logistics firms and banks and every enterprise still held together by spreadsheets, need technical people and still expect to grow them. In Malaysia that includes the GLCs and the whole data-centre and semiconductor buildout, a world that needs far more cloud, data and operations people than the glamorous end of tech ever advertised. Adjacent doors count too: support engineering, QA, solutions, analyst roles. They route back into engineering, and unlike the front door, they open.

Learn to use AI at practitioner depth. “Proficient in ChatGPT” on a CV is worth nothing. Run an agent on a real task and watch where it fails. Learn to review model output the way a suspicious senior reviews a pull request. SignalFire’s report advises companies to reframe junior hires as agent operators, people who manage, test and audit AI workflows. Be the person that description fits before the job ads learn the phrase.

And one observation rather than advice: graduates from the top 20 American CS programmes are now 45% less likely to join a Big Tech firm than a few years ago, and twice as likely to be founders as at the 2022 peak. I wouldn’t tell anyone to start a company because the job market is bad. I’d just note that the people with the most options have stopped queueing and started routing around the queue.

So, cooked?

Partly. The escalator is switched off and I don’t think it’s coming back on. The stairs still exist; they’re steeper, they start in stranger places, and nobody stands at the bottom handing out maps.

But hold two facts together: the industry has stopped training juniors, and it will still need mid-level engineers in five years. Those can’t both stay true. Around 2031, the companies currently sawing off their own bottom rung will be paying painful money for anyone with five years of real experience, and the only way to have that by then is to start now, anywhere that lets you.

Start building.