AI Comes for the Grunt Work, Not the Judgment
You type a sentence and the AI writes working code, a whole component while you blink. If you write software, you have felt the jolt: if the machine writes the code, what am I here for? It is the sharpest version of the question hanging over almost every knowledge job right now, from coders to analysts to marketers. "Will AI take my job?"
I answer it the same way every time, with a story that already ran to the end: the accountant and the computer. Because the fear the developer feels today is the exact fear the accountant felt a generation ago, and we know how that one turned out. So follow the accountant through the last big shift. The visual pins above and moves as you read. Then we'll come back to you, by way of the two questions everyone is really asking: is this an AI bubble, and either way, what does your job become?
"If the AI writes the code…"
A prompt becomes a working function. An agent refactors the file, writes the tests, and opens the pull request. For a developer it lands somewhere between magic and dread. The core craft, turned into an autocomplete. Analysts, writers, and support staff are feeling the same thing.
The real question is never "will the job survive?" It is "what shape does the job take next?" And to answer that, we do not have to guess. We can rewind.
Before the coder, the accountant
A company of any size once did not have an accountant. It had many, often several per department, keeping ledgers by hand and reconciling everything on calculators. Accounting was arithmetic and record-keeping, done by people, at scale.
Then a general-purpose machine arrived and did the arithmetic. It was the exact jolt the developer feels today. This part of the story already ran to the end, so let's watch how it actually turned out.
Then the computer arrived
Accounting software landed on every desk. The paper ledger and the desk calculator, the two tools that defined the craft, were suddenly obsolete.
This is the moment the fear predicts a wipe-out. Software can add, reconcile, and report faster and more accurately than any room of people. So surely the accountants were finished.
The job changed, it didn't vanish
Companies still needed accountants, just fewer of them per company. And the work itself transformed. The bulk of it became entering data into the software and validating what came out, rather than calculating by hand.
The rote part, the arithmetic, is exactly the part the machine took. What remained was judgment, correctness, and interpretation.
They moved up the stack
Something more interesting happened to the people. Accountants became the domain experts who helped design and build the accounting software in the first place. Nobody else understood the rules well enough.
And as users, the software did not shrink them; it amplified them. One accountant with good software could now do what a whole floor once did, and reach for analysis that was never practical by hand.
And it created whole new industries
This is the part the fear always misses. The accounting software did more than reshape one role. It called entire professions into existence. Software developers to build it. QA engineers to test it. Operations teams to run it as a service. Infrastructure to host it.
None of those jobs existed on the old accounting floor. The revolution retired one kind of work and manufactured many kinds in its place. The AI wave is already doing the same, sprouting titles nobody had a decade ago: prompt and agent engineers, eval and safety specialists, AI-ops teams.
Where the work went
Draw it out and the picture is unmistakable. The one column that shrank is manual calculation. Everything around it grew: validation, building, testing, operating, hosting. Most of it did not exist before.
Automation did not subtract a job. It rearranged one job into many, and pushed people toward the parts that need a human.
The third wave is here
Energy and machines were the first wave. Compute and software were the second. Intelligence and agents are the third, and it is the one we are standing in right now.
AI is a general-purpose technology. That is exactly why it will follow the accountant's arc, not escape it, and why the developer is simply the accountant of this wave.
It creates more than it retires
If the pattern holds, and general-purpose technologies have a strong track record, AI will take the rote layer of knowledge work the way software took manual arithmetic. Analysts, support agents, marketers, and yes, engineers will spend less time on the mechanical part and more on judgment, direction, and validation.
And it will spin up categories of work we do not have clean names for yet: people who build and evaluate AI systems, who supervise and correct them, who operate them safely at scale. More created than retired, again.
But there's a second story
So far this is the optimist's case, and I believe it. But there is a second story running at the same time, and ignoring it would be dishonest. And the link is uncomfortable: the same software wave that promoted our accountant also blew the dot-com bubble. Real transformation and real mania arrived together, and this time is no different.
The four biggest US tech firms (Microsoft, Amazon, Alphabet, and Meta) spent roughly $246 billion on capital expenditure in 2024, up about 60% in a year, and are guiding well past $320 billion for 2025, most of it AI data centers. Their combined spend has more than doubled in two years. When the numbers get this big, the word "bubble" starts showing up.
Every boom runs the same loop
There is a rhythm beneath every boom, and it has four beats. High returns pull in capital. Capital keeps flowing until far too much capacity gets built. The overcapacity crushes returns and the thing collapses. And then a few survivors pick up the wreckage cheap and make a fortune when demand finally catches up.
The technology can be completely real and the bubble can still burst. Those are not contradictions. They are the same story.
We have seen this film before
Britain's canal mania of the 1790s funded dozens of waterways, many never finished, yet the ones that were became arteries of the Industrial Revolution. Its railway mania of 1846 authorized around 9,500 miles of track and left roughly a third of it unbuilt. The 1920s radio boom sent stocks like RCA to the moon and then to the floor in 1929, but the broadcast networks stayed.
The clearest rhyme is the dot-com fiber glut: hundreds of billions were spent burying cable, and by 2002 at most about 5% of it was actually lit. The rest sat "dark" for years. Each time the story was intoxicating and mostly true. That never stopped the money from getting wildly ahead of the demand.
Does the spend pencil out?
Run the same arithmetic on AI. Sequoia Capital laid it out plainly: to justify the hardware being bought, the industry needs on the order of $600 billion a year in revenue. Actual AI revenue is still well under $100 billion, and several of the leaders are losing money. That is close to a ten-to-one gap between what is being spent and what is coming back.
The returns on the ground are shaky too. An MIT study of enterprise deployments found roughly 95% of generative-AI pilots delivering no measurable profit, and companies are already shifting to smaller, cheaper models just to keep their token bills under control.
The tech outlives the bubble
The twist ties both stories together. That dark fiber did not vanish when the bubble burst. It sat in the ground until streaming, cloud, and the smartphone arrived, and then it became the backbone of the modern internet, the very infrastructure that made Google, Netflix, and AWS possible.
The companies that overbuilt mostly died. The technology, and the transformation it enabled, absolutely did not. So even in the worst case for the AI market, the capability that has the developer worried is not going anywhere.
The developer is the new accountant
So the capability always outlives the crash. But surviving the bust is not the exciting part. What happens to the people is. The developer is walking the accountant's exact path. The AI takes the rote layer, the boilerplate and glue code and first drafts, the way software took manual arithmetic.
Picture the old org chart as a pyramid: a wide base doing the routine work, narrowing up through experienced owners, senior leads, and executives. AI eats the base. Smart teams do not simply lop it off. They let that pyramid bulge into a diamond, lifting people out of grunt work into roles that own more, direct more, and decide more.
The Jevons paradox
In 1865, William Jevons noticed something strange: as steam engines got more efficient, Britain burned more coal, not less. Cheaper power meant everyone found new uses for it, so total demand climbed. Efficiency did not shrink the market. It exploded it.
AI is the same story, and you are the coal. As the cost of doing a task collapses, we do not do fewer things; we do far more of them, and invent things that were never worth doing before. Microsoft's Satya Nadella invoked exactly this about AI; IBM, after once planning to automate roles away, is now tripling its entry-level hiring into redesigned jobs.
From writing the code to directing it
That is the shape of the new job. The developer stops hand-typing every line and starts deciding what to build and why: framing the problem, directing a team of AI agents, reviewing what they produce, owning the result. It looks a lot like the senior or architect work we used to reserve for years in, handed to people far earlier. The accountant became the person who ran the software; the developer becomes the person who directs it.
The skills that carry you up the diamond are stubbornly human: adaptability, curiosity, creativity, and above all critical thinking. Master the tool, help build it, and you do not get replaced. You get promoted.
The forecasters draw the same shape
Broadly, yes. The World Economic Forum's 2030 tally is roughly 92 million roles lost and 170 million created, a net gain of about 78 million. PwC sees headcount still climbing even in the most exposed occupations, with a steep pay premium on AI skills.
And McKinsey expects a big slice of today's tasks to become automatable this decade, but as an assistant to skilled work rather than a replacement for it. Three studies, one picture: the machine takes the mechanical part, people move to the judgment, and the total keeps growing.
A demo is not a rollout
I would not oversell the speed. Yale's Budget Lab finds the labour market broadly steady since ChatGPT arrived, and that fits how these shifts actually run.
A tool being able to do a task is not the same as an industry rebuilding itself around that tool. The gap between a slick demo and a company genuinely re-tooled is measured in years, not news cycles, and most companies are still at the demo.
The first rung is disappearing
One finding I won't wave away: Stanford's researchers see the youngest workers, those in their early twenties in exposed fields like software and support, already losing ground while senior colleagues hold firm.
That is the real hazard of this wave. Not mass unemployment, but the quiet removal of the first rung, the junior job where you used to learn the craft by grinding through the boring parts. The promotion still waits at the top of the ladder; getting a foot on it is what just got harder.
Climb, and hold the ladder
Which is why the advice is the same, whichever job worries you. Get fluent in the tool early, while everyone else is still debating it. Be the person who points it at the right problem, checks its work, and puts their name on the result. The judgment and the accountability are the parts it cannot hand back.
Do that and you don't get automated; you reach the senior work years early. And if you do the hiring: rebuild the bottom rung instead of deleting it. That pipeline is where your future seniors come from.
The journey, in one row
Where this series goes next
This post kicks off an eight-part series on AI as a human amplifier rather than a replacement. The posts that follow go deeper: the evidence that human-plus-AI teams beat either one alone, the "centaur" model of routing each task to whoever does it best, why AI lifts beginners the most, the cognitive-offloading trap that erodes the skills it was meant to boost, what it does to code and to creativity, the rising premium on the skills that complement AI, and a playbook for redesigning jobs around all of it.
A word on that beginner point, since it cuts against the entry-rung worry above. AI can make a junior productive in weeks, not years, so the danger was never that they cannot do the work. It is that companies stop hiring them at all. The pipeline is a choice, not a casualty.
- David Cahn, "AI's $600B Question," Sequoia Capital (2024).
- MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025."
- Big-tech AI capex: CNBC, Visual Capitalist, Epoch AI.
- Nvidia, FY2025 financial results.
- Historical parallels: Canal Mania (1790s), Railway Mania (1846), the 1920s radio boom, and the dot-com dark-fibre glut. On the fiber, see "Amid telecom ruins, a fortune is buried" (The Oregonian, 2002).
- Jevons paradox; Satya Nadella on AI and Jevons (GeekWire, 2025); IBM's entry-level hiring shift (TechCrunch, 2026).
- Future of work: WEF Future of Jobs 2025, McKinsey, PwC 2025 AI Jobs Barometer, Yale Budget Lab, and Stanford's "Canaries in the Coal Mine".
Figures are rounded and, where noted, approximate; several are analyst or investor estimates rather than audited totals.