About two years ago a friend of mine (I’ll call her Ana) worked in the back-office team of a midsize Brazilian bank. Her desk was buried in onboarding forms, ID scans, and the daily tedium of copying numbers into spreadsheets. Then her manager rolled out an AI tool that could read the documents, fill in the right cells, and check signatures in seconds. Six months later half the team had quietly left – and the seats stayed empty. “Nobody actually got fired,” she told me. “They just… stopped replacing people.”
Since ChatGPT showed up in late 2022, bosses everywhere have been asking the same thing: can fewer people do more if we bolt AI onto the workday? The data says yes, up to a point, with some nuance and the occasional fiasco along the way.
The quiet hiring freeze
A headline about thousands of layoffs grabs attention, but what almost nobody talks about is the hiring freeze. IBM’s CEO paused 7,800 back-office positions betting that attrition plus AI would cover the load, and cyber-security firm CrowdStrike cut 5% of its staff for the same reason. Even the companies that are still growing keep whispering that their next cohort of junior analysts or HR coordinators might be the last big one.
Why the caution? Because AI is good at exactly the boring part: data entry, scheduling, basic report generation. Those tasks are what used to justify an entry-level job, and today a chatbot or a document-reading model gets through them before your coffee cools. Companies don’t have to fire armies of assistants. They just stop hiring the new ones.
The World Economic Forum’s Future of Jobs 2023 survey found that 40% of employers expect to trim staff where AI can automate tasks. I treat that number as an illustration rather than a forecast. It’s not every job, but it’s a big chunk of white-collar work.
And the quality?
Doing things faster is worth nothing if what comes out is bad, and there have been a few well-documented failures along the way.
CNET quietly published dozens of AI-written finance explainers, and more than half had factual errors. Readers did not take it well. A pair of New York lawyers got fined after ChatGPT invented case citations in a court brief, which turned into a very expensive lesson about trusting a model with your eyes closed.
Customer service has its own horror stories. Klarna swapped 700 support agents for an AI bot in 2023 and, by early 2025, the CEO himself admitted the bot had left shoppers furious, and the company rushed to hire humans back. When refunds are on the line, speed doesn’t save you. (In cases like this I’m never sure if the problem is the bot or the rush of whoever decided to make the swap, but anyway.)
The pendulum swings the other way too. MIT researchers gave ChatGPT to a group of white-collar volunteers and watched tasks finish 40% faster with quality scores up 18%. At a Fortune 500 call center, AI suggestions lifted rookie agents’ productivity by 35%, and customer satisfaction scores went up along with it. In both studies a human stayed in the loop: AI drafted, someone checked, and only then did it reach the customer.
Upskilling doesn’t cover it
Platform-economy thinker Sangeet Paul Choudary argued recently that the feel-good slogan “AI won’t take your job, but someone using AI will” papers over much less feel-good shifts in power. He’s right, and you can see it in the cases above.
Some tasks change hands and others simply vanish. IBM’s freeze shows how whole areas of back-office work can be written out of the org chart once AI takes over the forms and the approvals. That’s the first fallacy Choudary points to: automation and augmenting the human aren’t a tidy either/or, and the task itself can lose its value overnight.
The MIT and NBER studies show AI can make a person faster and sometimes even improve the quality, but Choudary reminds us that the gain doesn’t automatically land in the worker’s pocket. If every analyst starts finishing ten decks before lunch, the going rate for a deck collapses, and the value concentrates with whoever owns the workflow, often the platform or the firm that has the data.
Rewriting a workflow opens up cracks that nobody sees at the time. Klarna’s support fiasco and CNET’s error-ridden articles are what happens when a company rips out the human checkpoint without redesigning who is accountable. Choudary calls this the “workflow isn’t sacred” fallacy: AI skips a step or pushes that step somewhere else, and sometimes it dumps the whole liability on the last human in the chain.
Bookkeeping clerks, junior lawyers, even graphic designers may keep the same job title, but as AI takes more of the heavy lifting, the pay premium erodes. The World Economic Forum’s job-churn numbers hint at that quiet dilution: the roles carry on existing, but a bigger share of the value ends up further upstream, in the hubs where the AI is concentrated.
Learning to use AI is necessary, sure, but it’s a life jacket, not a way out of the water. The harder part is seeing how your company is redrawing its workflows and the places where value collects, so you can put yourself where the new value ends up.
Who feels the squeeze first?
It starts with admin and clerical work. If your job is moving information from one box into another, AI is already looking at your chair: the people who key in data, run payroll, and audit travel expenses show up at the top of any list of the most automatable jobs. Customer service is similar. A chatbot answers the FAQs around the clock for pennies and handles a simple refund on its own, but a five-step warranty claim still needs a person, at least for now, so what’s left is fewer Tier-1 agents and more specialists working the exceptions.
In finance and accounting, between automated loan underwriting and AI-powered reconciliation, the number-crunching side is being transformed: bookkeeping clerks carry the biggest risk, and forensic accountants, the ones who can interpret the red flags the AI raises, are doing fine. Legal research splits much the same way. A language model that has swallowed an ocean of case law spits out a draft contract in minutes, and the junior associate who used to spend nights proofreading clauses now supervises an algorithm. You have to check everything, because the model can hallucinate, but the amount of work that comes out is enormous.
In creative and design, generative art shook illustrators in 2023, when Netflix Japan’s “The Dog & The Boy” used AI-generated backgrounds. Agencies now produce thirty ad mock-ups in Midjourney before an art director picks one, the market for average artwork is collapsing and demand for good storytelling and brand voice is going up. In software development GitHub Copilot writes boilerplate in seconds, senior engineers love it and juniors worry, and the best devs now architect systems and review AI output for edge-case bugs instead of typing out every for-loop by hand.
Roles that didn’t exist
Notice that in all of these cases the routine slides to the machine, and what keeps earning money is the work that needs judgment and creativity. But every wave of automation also creates jobs nobody imagined ten years earlier, and AI is no different. A few that barely existed before 2021:
- Prompt engineer, who is part poet and part programmer, and gets better answers (or images) out of the models.
- AI ethicist, the conscience of the algorithm, who finds the bias before the thing reaches production.
- Model risk manager, because at some point a regulator will ask why your AI denied that mortgage.
- Human-in-the-loop supervisor, to look after the blended workflows where people sort out the exceptions AI can’t.
The World Economic Forum forecasts a 30% to 40% jump in demand for AI and data specialists by 2027. Somebody has to build, tune, and police these systems.
How to stay (happily) irreplaceable
- Lean into the messy parts. Diplomacy with an angry client, strategic brainstorming, mentoring a confused junior: it’s all subjective and full of gray areas, and that’s exactly where AI trips.
- Get better at steering AI. You don’t need to code a transformer, but you do need to know how to drive one, because the people who can drive the tools tend to keep their seats.
- Cultivate taste. Picking the best design prompt, spotting a bogus legal citation: I’d say judgment is the skill that matters most today. And the call on when not to use AI also has to stay with the people.
- Bet on cross-discipline knowledge. Someone in finance who understands UX, or a marketing person who can write a basic Python script, stands out in any hiring freeze.
It won’t be painless, and I’m not going to pretend otherwise. Transition periods rarely are. But looking back, the people who partnered with their machines have usually come out better than the ones who fought them.
So, are the robots taking all our desks? Unlikely. What’s more likely is that they rearrange the office. The gray, repetitive cubicles go off to one side, and the humans get the better-lit rooms, where creativity, empathy, and long-range thinking fit.
If we get the balance right, maybe Ana’s next new colleague is a tireless AI sidekick that handles the grunt work while she takes on the projects that actually make her proud, instead of another overwhelmed intern. That “if” is a big one, though, and I’ve got no guarantee we’ll get it right.
P.S.: yes, I used AI to help me structure this article.