The AI Review Tax: Why Cutting Junior Roles Is Setting Your Company Up for Burnout

There’s a story that businesses across every industry keep telling themselves right now: AI lets one person do the work of three. Entry-level roles are obsolete. Lean into it or get left behind.

The data partially supports that story. A Harvard study examining 285,000 US firms and 62 million workers found that when companies adopt generative AI, junior employment drops 9–10% within six quarters (Hosseini & Lichtinger, Harvard, 2025). Venture capital firm SignalFire found a 50% decline in entry-level starts at the largest public tech firms between 2019 and 2024 (CNBC, September 2025). In the US, programmer employment fell 27.5% between 2023 and 2025 (IEEE Spectrum, December 2025). CS graduates now face a 6.1% unemployment rate. Computer engineers sit at 7.5%. Fine arts graduates are, statistically, more employed than both (Stack Overflow Blog, citing Federal Reserve data, December 2025).

These numbers are real. I don’t want to minimise the hardship of someone who studied for four years and can’t get a foot in the door. But I’m going to argue that most companies are looking at this from entirely the wrong angle. And the ones that figure out the right angle first will have a competitive advantage that lasts a generation.

A quick note: I’m going to use software development as the primary example throughout this piece, because that’s where the data is richest and the effects are most visible. But the argument applies far beyond tech. Any role where people are using AI to draft, research, design, analyse, or produce work — from marketing teams to law firms, to financial analysts to architects — is subject to the same dynamics. The review tax doesn’t care what industry you’re in.

A Signal Hiding in the Noise

To start we need to look more carefully at what the headline numbers actually describe.

That 27.5% decline in programmer employment? It’s specifically for “programmers”, i.e. people who primarily write and test code to specification. Employment for “software developers” in more design-oriented, systems-thinking role fell just 0.3% in the same period (IEEE Spectrum, December 2025). The jobs that translate business needs into architecture are holding. The jobs that translate specifications into output are contracting.

Meanwhile, a randomised controlled trial by METR tested 16 experienced developers on 246 real tasks. The developers predicted AI would make them 24% faster. They believed it made them 20% faster. The actual measured result? 19% slower (METR, July 2025). A separate Microsoft and MIT Sloan study of over 500 developers found that 90% reported increased productivity, but gains were concentrated in routine, repetitive tasks (Pure AI, August 2025).

Does AI make people faster or slower? It depends on the task, the experience level, and the complexity of the work. That ambiguity matters. We’re nowhere near the clean “AI replaces humans” narrative that dominates public discourse.

But all of this is the standard debate. You’ve read versions of it before. Here’s where the conversation shifts.

The Blind Spot

In February 2026, the Harvard Business Review published research that should have stopped every CEO mid-sentence. UC Berkeley researchers spent eight months embedded inside a 200-person tech company, conducting over 40 in-depth interviews across engineering, product, design, research, and operations. Their finding was stark: AI doesn’t reduce work. It intensifies it (HBR, “AI Doesn’t Reduce Work — It Intensifies It,” February 2026).

Workers using AI took on more tasks because the tools made more feel “possible.” Product managers started writing code. Researchers took on engineering work. Roles that once had clear boundaries blurred as people absorbed tasks that previously sat outside their remit. Nobody was told to do this. They just did, because AI made it feel doable.

Then came the invisible load. Engineers found themselves spending increasing amounts of time reviewing and correcting AI-assisted work produced by colleagues by people who were “vibe-coding” and submitting partially complete pull requests. This review work surfaced informally: in Slack threads, quick desk-side consultations, a “can you just take a look at this?” that multiplied across the day. None of it showed up on any dashboard but all of it added up.

The researchers found that workers felt more productive but not less busy. In many cases, busier. Prompting during breaks became habitual. Downtime no longer provided recovery. As one engineer told the researchers: “You had thought that maybe, oh, because you could be more productive with AI, then you save some time, you can work less. But then really, you don’t work less. You just work the same amount or even more” (Fortune, February 2026).

TechCrunch, reporting on the same study, put it plainly: the research shows where AI augmentation actually leads, which is “fatigue, burnout, and a growing sense that work is harder to step away from, especially as organisational expectations for speed and responsiveness rise” (TechCrunch, February 2026).

This is the blind spot. And it’s enormous.

The AI Review Tax

Let me give this problem a name. I call it the AI review tax. The unavoidable cost of checking whether AI output is actually fit for purpose. It doesn’t matter if the output is good or bad. Someone still has to look at it, evaluate it, and make a call: is this usable, does it need rework, or do we start again? That decision takes time, context, and judgement. And it happens every single time AI produces something.

The review tax goes beyond just reading output. It’s the prompting that precedes it, the iteration when it misunderstands, the re-prompting when context drifts, the checking against reality, the reformatting, the testing. But the review itself as the moment a human has to decide whether the thing in front of them is right is the bottleneck. Everything else is overhead. The review is the gate.

The evidence for this tax is piling up from every direction.

Atlassian’s 2025 State of Developer Experience survey, covering 3,500 developers and managers across six countries, found what they called an “unexpected paradox.” Developers were saving over 10 hours a week with AI tools. But 50% were simultaneously losing 10 or more hours a week to organisational inefficiencies through poor coordination, unclear direction, finding information, and context-switching between tools. The net productivity gain for half the workforce was effectively zero (Atlassian, July 2025). The time AI saved at the individual level was being consumed by the review overhead it created at the team level.

Then there’s “workslop.” That’s the term researchers at Stanford’s Social Media Lab and BetterUp coined for AI-generated work that looks polished on the surface but lacks substance. The hollow memo, the report littered with confident-sounding nonsense, the code that passes a glance but fails at edge cases. They surveyed over 1,100 full-time workers and found that 40% had received workslop from a colleague in the past month. On average, each incident cost nearly two hours to decode, correct, or redo entirely. For a 10,000-person company, they estimated the annual cost at over £7 million in wasted productivity (HBR, “AI-Generated Workslop Is Destroying Productivity,” September 2025; Fortune, October 2025).

The burden of this tax falls disproportionately on experienced workers. They’re the ones with the context to spot when AI output is subtly wrong. They’re the ones reviewing pull requests from colleagues who prompted their way to something that “looks right.” They’re the ones fielding the Slack messages and the desk-side consultations. And they’re the ones whose strategic, high-value work gets squeezed out as a result.

Right now, in most organisations, the review tax is falling squarely on the shoulders of senior people.

That’s the part that doesn’t make sense. That’s the bottleneck.

The Lesson We Keep Refusing to Learn

There is a pattern in technological disruption so consistent it borders on a law of nature: when technology makes something dramatically cheaper or easier, we don’t do less of it. We do astronomically more.

When clean coal technology improved the efficiency of power generation, the worry was that the coal industry would shrink. Instead, cheaper energy meant more demand, more plants, more coal mined than ever. When the automobile replaced the horse, the concern was mass unemployment for the transport industry. Instead, cheaper transport meant more travel, more logistics, more roads, and orders of magnitude more workers in transport than the horse-and-carriage era ever employed. When open-source software threatened to make commercial code worthless, Microsoft’s market capitalisation grew roughly 20 times over the following two decades.

AI is making software cheaper to produce. That won’t mean less software. It will mean vastly more software, built for purposes that nobody currently considers viable, serving markets that don’t yet exist. And all of that software will need to be reviewed, tested, maintained, and understood by humans who grasp the business context behind it.

More software means more output to review. More output to review means more people needed to check it. The question is not whether humans remain in the loop. It’s which humans, doing what.

The Context Problem

Here’s where the “AI replaces juniors” narrative actually has a point and also where it fatally undermines itself in the same breath.

Senior people hold the context. They know why the client changed their mind on Tuesday. They know the legacy billing integration has an undocumented quirk. They remember that the compliance team rejected a similar feature six months ago for reasons that never made it into a ticket. They can read the room in a stakeholder meeting and realise the brief is politically motivated rather than technically sound. AI has none of this. So yes, in theory, a senior person can feed that context directly into an AI prompt, cut out the junior entirely, and get a first draft back in minutes.

But think about what it actually means in practice. The senior provides the context, waits for the output, and then has to personally evaluate whether the AI’s first draft is good enough to use. Often it isn’t. The Stanford workslop research made this explicit: AI-generated work “shifts the burden downstream, requiring the receiver to interpret, correct, or redo the work.” Except now there is no downstream. The senior is the downstream. They’re prompting, reviewing, correcting, re-prompting, and polishing all by themselves.

Before AI, this is exactly what juniors were for. Not because they were cheap labour but because they were a relief valve. A senior would brief a junior, the junior would produce a first draft, and the senior would review it. If the draft was rough, the senior would explain what was wrong and why. Next time, the junior’s draft would be a little better. The time after that, better again. Over months and years, the senior’s review workload shrank as the junior absorbed more context, developed better judgement, and produced work that needed less correction. The relationship was an investment that paid down over time.

AI doesn't work like that. Yes, you can improve its output with custom instructions, guardrails, style guides, project context baked into prompts. Good teams are already doing this, and it helps. But here's the fundamental difference: a junior who gets corrected learns why they were wrong. They absorb that correction and eventually become the senior person providing the briefs, making the calls, and needing their own juniors to absorb the review tax. AI never makes that journey. No matter how good the guardrails get, it still needs someone to feed it the context and someone to check what comes out. It can get better at the middle part, but it can never move up the chain. The junior absorbed context and eventually gave it back as competence. AI absorbs context and gives back a first draft that needs to be checked every single time.

That’s not a productivity gain. That’s a trap.

Why I Think Juniors Become More Important, Not Less

The “one person does the work of three” framing assumes the experienced person doing AI-augmented work can sustain it indefinitely. The HBR research says they can’t. The review tax is real, it’s growing, and it’s burning people out.

Here is what I think the sustainable model actually looks like. And if it sounds familiar, that’s because it’s essentially how well-run teams have always worked.

You’re a product manager building three features simultaneously. You define the what and the why. A junior takes that brief and does the legwork by prompting AI, iterating on outputs, managing the context, waiting for generation, exporting, performing initial quality checks. They produce a first pass. A senior reviews their work, teaches best practices, catches what AI got wrong. The senior pushes refined work to you for final sign-off.

That’s not revolutionary. It’s the same brief-draft-review cycle that knowledge work has always run on. Senior provides direction, junior executes, senior reviews and teaches, junior improves. The structure isn’t new.

What’s new is the speed. A junior armed with AI tools doesn’t produce one first draft per day. They produce three, or five, or ten. The iteration cycle that used to take a week compresses into hours. Which means the volume of reviewable output per junior has increased dramatically, and that means each senior can now be supported by more juniors working at higher throughput, not fewer. The old model was one senior reviewing the work of two or three juniors. The new model might be one senior reviewing the work of four or five, because each junior’s output rate has multiplied.

In this model, the junior isn’t doing what juniors used to do. They’re not writing code from scratch or designing layouts by hand. They’re operating AI and absorbing the review tax. Checking output, flagging problems, iterating until it’s ready for senior eyes. The senior isn’t drowning in first-draft triage and pair programming. They’re doing high-value review and teaching. And the product manager isn’t trapped in iteration loops. They’re focused on strategy and context, which can't be replaced by AI.

Three features moving forward simultaneously. Sustainable workloads at every level. And critically, juniors learning through the review process by building the context and judgement that, over time, means their drafts need less correction and seniors can redirect that recovered time into higher-value work.

This isn’t a tech-specific model. It works anywhere AI is producing first drafts that need human review. A marketing team where juniors coordinate AI-generated campaigns while seniors review strategy. A legal department where trainees run AI-assisted research while partners review for accuracy and precedent. An architecture firm where graduates manage AI-generated plans while principals review for compliance and client context. The review tax exists in every one of these settings. The question is who absorbs it and how many of them you need now that the pace has accelerated.

I’m Not Entirely Alone

Chris Eldridge, CEO of Robert Walters, warns that removing junior roles will “starve the internal talent pipeline” and trigger a “talent doom cycle” of escalating hiring costs (CNBC, September 2025). Fabian Stephany at the Oxford Internet Institute argues firms cutting juniors risk becoming stale. AWS CEO Matt Garman called replacing juniors with AI “one of the dumbest things I’ve ever heard” (IT Pro, August 2025).

Ethan Mollick at Wharton comes closest to my thinking. He’s written that the implicit training pipeline in organisations has broken and needs reconstructing at the CHRO level, and argues leaders need to explicitly decide which work goes to AI and which builds judgement (Valence, August 2025; One Useful Thing, Substack).

But most of these voices argue from a future-facing position: we need juniors for the pipeline. My argument is different. I’m saying we need juniors right now, because the review tax is already large enough to justify dedicated roles and because loading it onto experienced workers is already proving unsustainable.

The Honest Counterarguments

The data is not on my side right now. Companies are cutting junior headcount, not expanding it. GitHub reports Copilot contributes approximately 46% of code across organisations. DBS Bank cut 4,000 jobs while creating only 1,000 new AI-specialist positions (Fortune, February 2025). The market is voting for fewer juniors.

There’s a plausible argument the review tax is temporary. Every new technology has an awkward adoption phase: early spreadsheets needed specialists, early websites needed webmasters, early cloud needed DevOps engineers. Then the tools matured and the overhead dropped. If AI interfaces become seamless enough, the review tax might evaporate before it justifies dedicated roles.

There’s also a legitimate question about whether I’m projecting my own workflow. I’m building multiple products simultaneously. I want juniors to delegate AI operations to. Most companies aren’t in that position.

And the offshore alternative looms. If the job is “operate AI and iterate,” companies might choose £10-an-hour global talent over £25,000-a-year domestic graduates.

Why I Believe It Anyway

The burnout trajectory is already unsustainable. The HBR research doesn’t describe a temporary adjustment. It describes a self-reinforcing cycle: AI makes more work feel possible, which leads to more work taken on, which leads to less recovery, which leads to cognitive collapse. An EY survey of 500 senior leaders found that more than half feel like they’re failing amid AI’s rapid growth, with companywide enthusiasm for AI on the decline (CIO Dive, December 2024). A DHR Global survey of 1,500 corporate professionals found 83% experiencing burnout, with overwhelming workloads as the top cause (Decrypt, February 2026). This isn’t theoretical. It’s already happening.

The historical pattern is undefeated. Every time technology makes something dramatically easier, total demand for human involvement increases. It always has. There is no documented exception. The question is never “will humans be needed?” but “what will humans be needed for?” And the answer, every single time, is review, judgement, and context.

Someone has to absorb the review tax. It can be senior professionals who burn out doing everything. It can be offshore contractors who lack organisational context. Or it can be juniors embedded in teams, learning from seniors through the review process, building the judgement that makes them senior professionals in five years.

I know which model I’d bet on.

Where this leaves us

The jobs aren't disappearing. They're transforming. But the companies that figure out the right human-AI team structure first; the ones that distribute the review tax instead of dumping it on their most experienced people, will have a head start that lasts a generation.

The question is no longer 'can I build this?' It's 'who operates the AI, who reviews the output, who holds the context, and who decides whether it was worth building in the first place?'

I know where I'd put my juniors. And I'd call them the AI Tax Accountants... though their LinkedIn profiles will probably say AI Coordinator.