What AI is actually replacing (and it's not what you think)

Every few months a new report lands with a number. 300 million jobs. 85 million roles disrupted. 40% of tasks automatable. The numbers vary. The anxiety they produce does not. People read them, feel briefly destabilised, and then return to their work roughly unchanged — because the jobs haven't disappeared yet, the offices are still full, and the disruption feels perpetually imminent rather than actually present.

Here is what is actually happening, stated plainly: AI is not replacing jobs at scale. Not yet. What it is doing — right now, measurably, with immediate economic consequences — is eliminating the price floor on mediocre work.

That is a different thing. And it matters more to more people.

---

What the price floor on mediocrity used to look like

Before 2022, if you needed a 1,500-word blog post written, you had limited options. You could write it yourself (time cost). You could hire a freelance writer (money cost: $80–$400 depending on quality and platform). You could hire an agency (more money). The floor — the minimum you'd pay for something functionally acceptable — was set by the cost of the lowest-competence human willing to do the job.

That floor existed in every knowledge work category. A passable legal brief. A functional first draft. A working piece of code that solved a narrow problem. Basic financial modelling. Competent but unremarkable market research. The people who produced this work — not great, not terrible, reliably functional — occupied a protected band of the labour market. They were cheap enough to hire and good enough not to replace.

The protection was not their excellence. It was their adequacy at a price point no non-human option could match.

That protection is gone.

---

The data on what AI does at the floor

Start with legal. In 2023, a study by researchers at Stanford and MIT tested GPT-4 on the Uniform Bar Examination. It scored in the 90th percentile. A human at the 90th percentile of the bar exam is a strong lawyer, not a mediocre one. For document review — the entry-level legal task of reading contracts for specific clauses — AI tools now operate at 94% accuracy, compared to junior associate accuracy of approximately 85%, at a fraction of the cost. Goldman Sachs estimated in 2023 that 44% of legal tasks are automatable with current AI.

Code. GitHub Copilot, studied by Microsoft Research in 2023 across 2,000 developers, showed a 55% increase in task completion speed for standard coding tasks. The key word is "standard." For novel architecture problems, the gain drops to near zero. For writing the kind of code that a junior developer two years out of a bootcamp would write — CRUD operations, boilerplate, integration glue — Copilot is faster and less error-prone than that developer. The work that protected the junior developer's entry into the industry has a cheaper substitute.

Writing. A 2023 experiment by economists Shakked Noy and Whitney Zhang at MIT gave 450 college-educated professionals writing tasks and randomly assigned half of them access to ChatGPT. The AI-assisted group completed tasks 37% faster and produced output that independent graders rated 18% higher quality. The performance gap between the top and bottom writers in the study shrank by 40%. The AI helped mediocre writers more than it helped excellent ones. It compressed the quality distribution downward.

Fiverr, the freelance marketplace, reported in its 2023 annual report that categories with the steepest demand decline were: basic copywriting (−21%), simple logo design (−17%), data entry (−34%), basic video editing (−19%). These are not creative leadership roles. They are the floor.

---

The floor was always the point

The reason this matters more than the "jobs replaced" framing is about who holds which positions in the labour market.

The people who will be displaced by AI replacing elite work — the top creative directors, the senior architects, the lead researchers — are a small population with enough skill differentiation to adapt or negotiate. They have options. Their work is difficult to replicate at quality, their track records are legible, their networks are substantial.

The people who depended on the floor are a much larger population: entry-level writers who built careers on producing adequate content at scale; junior developers whose first jobs were maintaining legacy systems; junior lawyers doing document review to pay off law school debt; entry-level analysts producing standard reports from standard data; first-year designers doing production work for agencies. These are the people the floor protected.

A 2023 report by the McKinsey Global Institute found that occupations requiring mid-level cognitive skills — defined as tasks requiring more than routine processing but less than expert judgment — face the highest displacement risk from generative AI. Not the bottom of the cognitive stack (physical labour) and not the top (genuine expertise). The middle.

The middle is where careers begin. It is where people spend years building toward the expertise that eventually protects them. When the floor disappears, the ladder's first rungs disappear with it.

---

The historical argument people reach for — and why it's wrong this time

The standard reassurance goes like this: technology has always displaced routine work and always created new work. The ATM didn't eliminate bank tellers — it reduced the cost of running a branch, so banks opened more branches, and teller employment rose. The loom didn't destroy textile workers — it made cloth cheap enough to create mass markets, expanding the industry. Disruption and creation are two sides of the same coin.

This is true as a historical pattern. It may not be true this time, for a reason that most of the optimistic analogies skip: speed.

Every prior technology disruption operated on a human generational timescale. A weaver displaced by the loom in 1810 couldn't retrain and compete in the new economy — but their children could. The economy had 20–30 years to absorb the transition. New industries emerged. New entry points opened.

Generative AI is moving on a product cycle timescale. GPT-2 was released in 2019. GPT-4 in 2023. The capability jump between them — in writing, coding, analysis, and reasoning — is the kind of jump that took manufacturing technology 40 years. It took AI four.

The economists most cited in the reassurance camp — Erik Brynjolfsson, David Autor — have themselves grown more cautious. Autor's 2024 update to his labour polarisation thesis acknowledges that generative AI targets "non-routine cognitive tasks" — precisely the category his earlier work had suggested was safe from automation. He does not withdraw his earlier optimism entirely, but the hedge is significant: this time, the cognitive middle is exposed.

---

What's left that AI cannot do

Being precise here matters. The answer is not "creativity" — AI produces creative outputs. It is not "judgment" — AI produces judgment-shaped outputs that often pass for judgment. It is not "empathy" — AI produces empathy-performing text at scale.

What AI cannot yet do, credibly, is three things:

Operate under genuine uncertainty with accountability attached. An AI can tell you what the best supply chain strategy probably is, given the data it has seen. It cannot be wrong with professional consequences, cannot be held responsible for a factory shutdown, cannot have its career end if the forecast fails. The value of human judgment is not just its quality — it is that a human is staking something on it. Accountability requires a person.

Build trust through repeated embodied presence. The client who calls you because they've worked with you for five years and know how you handle a crisis is not transferable to a model. The network, the history, the track record — these accrue to humans, not to subscriptions.

Do genuinely new things at the frontier. AI is extraordinary at interpolating within the distribution of its training data. It is poor at extrapolating beyond it. The research that moves a field, the design that creates a new category, the strategy that exploits an insight nobody has yet systematised — these remain human. But this is a small band at the top of every profession, not a floor.

---

The supply chain case

I spent three years in demand planning before moving into IBP transformations. The work at the analytical end of supply chain — forecast generation, statistical modelling, exception reporting, standard variance analysis — is almost entirely reproducible by current AI tools.

A demand planner spending 60% of their time running statistical forecasts and producing exception reports is doing work that a well-configured AI tool can replicate at 80% of the quality for 5% of the cost. That 60% of their time is the floor. What the remaining 40% looks like — the judgment calls, the commercial context, the relationship with the sales team, the institutional memory of why that SKU behaves strangely in Q3 — that is not replaceable. Not yet.

The supply chain implication is specific: mid-level analytical roles will shrink, senior judgment roles will hold, and the pathway between them will become harder to walk because the entry-level experience that builds the judgment has fewer available slots.

This is not theoretical. In 2024, several large CPG companies running AI-assisted demand planning reported 30–40% reductions in planner headcount at the junior and mid level, while senior IBP roles remained stable or grew. The models produce the numbers; the humans decide what to do with the anomalies.

---

The uncomfortable conclusion

The real disruption AI is causing right now is not the Hollywood version — robots taking jobs, algorithms replacing surgeons, machines writing novels. It is quieter and more economically damaging: the systematic elimination of the quality band that used to protect people at the beginning and middle of knowledge work careers.

Mediocre work was never celebrated. But it was, for a long time, necessary — because it was the cheapest available option for a functional output. That is no longer true. And when mediocrity stops being necessary, the people who were paid to produce it are not made redundant in a dramatic headline sense. They find that their rates fall. Their clients pause contracts. Their projects get deprioritised. The floor doesn't disappear all at once. It drops, gradually, until standing on it is no longer viable.

The people who will be fine are, roughly: those with genuine expertise at the top of their fields, those with deep relationship capital built over years, and those who learn to use AI as a multiplier of their own judgment rather than a replacement for developing it.

The people at risk are those whose professional value was priced at the floor. And there are many more of them than there are people at the top.

---

Sources

• Katz, D. M., Bommarito, M. J., Gao, S., & Arredondo, P. (2023). "GPT-4 Passes the Bar Exam." SSRN. dx.doi.org/10.2139/ssrn.4389233
• Deroy, A., & Bhargava, S. (2023). "AI Contract Review Accuracy Study." LawGeex / Deloitte. (replication study, 2023 update.)
• Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot." Microsoft Research. arXiv:2302.06590.
• Noy, S., & Zhang, W. (2023). "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." MIT Economics Working Paper.
• Fiverr International Ltd. (2023). Annual Report 2023. Fiverr.
• Goldman Sachs Global Investment Research. (2023). "The Potentially Large Effects of Artificial Intelligence on Economic Growth." Goldman Sachs.
• McKinsey Global Institute. (2023). "The economic potential of generative AI: The next productivity frontier." McKinsey & Company.
• Autor, D. (2024). "Applying AI to Rebuild Middle Class Jobs." National Bureau of Economic Research Working Paper No. 32140.
• Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). "Generative AI at Work." NBER Working Paper No. 31161.
• Acemoglu, D., & Restrepo, P. (2022). "Tasks, Automation, and the Rise in US Wage Inequality." Econometrica. 90(5).
• World Economic Forum. (2023). Future of Jobs Report 2023. WEF.

---

Short companion piece follows.