What The Work Is

What the great AI jobs panic tells us about the future of work and about enterprise AI adoption.

The View From the Top

The conversation is always the same. I’ve had it countless times in my career.

Someone senior, usually in a corner office, leans back in their chair and describes a process. It all sounds so simple.

“So what we do is take data from system A, do process B to it, and then put it into system C. It really couldn’t be easier. It should be very straightforward to automate.”

Senior leader in any organisation

Then you go and talk to the person who performs the process day-to-day. You repeat the description you heard from their supervisor’s manager’s boss. They frown. They make that intake of breath through gritted teeth noise that always means the same thing.

“Yeah…so…that’s basically right…most of the time. Except at month end, for location X, or entity Y, or product Z, or when there’s a full moon on a Thursday in a month with 30 days.”

Frontline worker in any organisation

You see, they use the old version of the system, or do everything on paper because that’s the way they’ve always done it, or have this weird tax rule, or Gareth insists on doing it differently – everyone knows about Gareth (eyeroll), don’t they?

I’ve seen this play out many times.

I once had to explain to my incredulous boss in New York that yes, I was aware that it was 2012, and that nobody used fax machines any more, but that nevertheless, that was a hard requirement for clients in Japan and that we were going to have to build that functionality into the shiny new system we were designing, whatever the cost.

Another time, another company, another industry. As a (moderate) Spanish speaker I found myself flying for 30 hours, in Economy, from Perth to the Caribbean coast of Colombia to help figure out the vagaries of local tax law which didn’t seem to make sense back in head office and were going to derail the implementation of a globally standardised SAP system.

These were modern symptoms of a malady as old as human organisation, a complaint of every ruler in history: the way they do things in the far outposts of the empire is very different from the way we do things in the capital. The world is more complicated than it appears from the corner office.

This is not to rubbish the occupants of that corner office, or the rulers of the empire. The higher up the hierarchy you progress, the more removed from the frontline you get, and the greater the level of abstraction necessary for you to understand what your underlings are doing.

In my previous essay, This Time Is Different, I wrote about political scientist James C. Scott’s concept of “legibility”. In order to control the world, states (and companies) first need to make it simple. The problem is, the world isn’t quite so simple.

The End of Jobs

As the Generative AI boom has gathered pace since 2022, there has certainly been no shortage of wild panic-inducing headlines. Whatever their faults, it’s hard to deny that Sam Altman, Dario Amodei, Demis Hassabis, Elon Musk and the rest have a flair for publicity. One of their recurrent tropes has been the impending chaos as white collar workers are replaced en masse by GenAI. As Microsoft’s AI CEO put it:

“These are fundamentally labor-replacing tools…that shift value from labor to capital because an A.I. is essentially a capital form of labor.”

Mustafa Suleyman

Suleyman’s peers concur, although they haven’t all used the same proto-Marxist language. Instead, they’ve been very specific about the scale and timing of the impact. In May 2025, Dario Amodei initially predicted that half of all entry-level jobs would disappear within one to five years, causing a 20% increase in overall unemployment rates.

Dario and his old boss Sam Altman famously don’t get along so well, after Dario and his sister Daniela jumped ship from OpenAI and started Anthropic, but despite their differences, they seem to agree on this topic. Altman told the US Federal Reserve that whole categories of jobs would soon vanish as AI replaced them.

In fact, Altman said in July 2025, his company’s popular chatbot had already made the medical profession largely obsolete:

“ChatGPT today, by the way, most of the time, can give you better – it’s like, a better diagnostician than most doctors in the world”

Sam Altman

But wait, there’s more!

As Amodei sees it, AI isn’t just coming for entry-level white collar workers, it’s ultimately coming for everyone.

“We will eventually reach the point where the AIs can do everything that humans can. And I think that will happen in every industry.”

Dario Amodei

Is this necessarily a bad thing though?

Elon Musk doesn’t think so. In fact, he sees the end of jobs as cause for celebration, as humanity enters an unprecedented age of abundance. In 2023 he told then UK Prime Minister Rishi Sunak:

“There will come a point where no job is needed… You can have a job if you want to have a job for sort of personal satisfaction, but the AI will be able to do everything…It’s like a magic genie…You just have as many wishes as you want.”

Elon Musk

In a wonderful irony, Musk, darling of the political hard right, sounds eerily reminiscent of the right’s favourite bogeyman, Karl Marx. In the century-and-a-half since Marx died, his often strange and confusing philosophy has been routinely caricatured by supporters and detractors alike. It’s often been forgotten, but in much of his writing Marx was clear that technological progress was a prerequisite for bringing about the end of capitalism and the coming of socialist utopia.

At his most whimsical and optimistic, Marx imagined communism as a bucolic paradise where work was a thing of the past and all humanity basked in technologically-produced abundance, freed from the shackles of labour to do whatever we please. Does he remind you of anyone?

“…in communist society, where nobody has one exclusive sphere of activity but each can become accomplished in any branch he wishes, society regulates the general production and thus makes it possible for me to do one thing today and another tomorrow, to hunt in the morning, fish in the afternoon, rear cattle in the evening, criticise after dinner, just as I have a mind, without ever becoming hunter, fisherman, herdsman or critic.”

Karl Marx

Voodoo Economics

Amid huge excitement in certain circles about the imminence of this post-labour economics, the world of boundless plenty has arrived early for some, not least the founders and employees of the AI companies, whose valuations are soaring. Anthropic is widely reported to be preparing for an IPO in the fourth quarter of 2026, seeking a valuation of US$2 trillion. That figure would make it the largest IPO in history.

That’s some going for a company that lost US$42 billion (close to US$4 billion a month) in 2025 and has publicised spending commitments of more than half a trillion dollars. The industry as a whole is spending more than US$800 billion on data centres in 2026 alone.

A major element of the rationale for all this can be found in the AI companies’ pre-IPO filings, which include the companies’ own projections of their “Total Addressable Market” (TAM), a value representing the size of the market that the company serves.

According to the Wall Street Journal, Anthropic will reportedly claim a TAM of US$30 trillion. This is, conveniently enough, a couple of trillion higher than SpaceX’s projected TAM in their IPO filing; it seems the AI CEOs love a game of “my TAM is bigger than your TAM”.

For comparison, total US Gross Domestic Product (GDP) is approximately US$32 trillion a year. As Reuters reported it, “To quantify ​its TAM, Anthropic is looking at the ​full scope of work that could be completed with AI models, WSJ said.”

In other words, Anthropic’s extraordinary valuation ultimately rests on an extraordinary assumption about work: that AI models will eventually be capable of performing an enormous proportion of it. Such a projection chimes with remarks Dario Amodei has reportedly made about a near future where Anthropic is the only company in the world.

Crazy though this sounds, if the AI CEOs are right and AI will soon replace all human labour, or even a significant part of it, perhaps it’s much more plausible than it seems.

Crucially, the claim that AI can perform most human work isn’t merely a technological prediction, it’s a critical assumption embedded in the valuations of the companies making the predictions.

But where did these predictions come from?

Taskmaster

Despite their race to list on the stock market in the largest IPOs in history, both Anthropic and OpenAI began life as not-for-profits, and both continue to style themselves as research labs rather than companies. Self-contradictory though this might appear, it does at least result in both attempting to maintain a fig leaf of academic respectability by occasionally publishing quasi-academic style research papers on important topics.

One of those topics is AI’s impact on jobs. In March 2026, Anthropic published Labor market impacts of AI: A new measure and early evidence, followed a month later by OpenAI’s The AI jobs transition framework: Mapping AI’s near-term impact on jobs. The two papers have much in common, not least their underlying methodology, which they share with various previous papers from the same organisations and various others.

As the Introduction to the OpenAI paper puts it:

“Most analysis of AI’s impact on the labor market begins with the same core question: what jobs are most exposed to AI?”

OpenAI

Both papers acknowledge that real-world AI adoption and job impacts haven’t simply followed a linear path from AI capability to replacement and both seek to provide a more refined, nuanced framework to predict impact. However, it’s inescapable that the underlying thesis of this whole narrative is, as Amodei himself stated, the idea that “we will eventually reach the point where the AIs can do everything that humans can”.

Anthropic and OpenAI use strikingly similar approaches to come to this conclusion and to measure how far away from it we currently are. Even their more sophisticated models begin with an ontology of work inherited from the same place: work as enumerable tasks.

As the screenshots below illustrate, both companies’ job loss predictions are based on analysis originally published in 2023, in a paper by a team of OpenAI researchers led by Tyna Eloundou. Both Eloundou’s paper and both companies’ subsequent research use the same underlying approach of breaking jobs down into their constituent tasks, using an open source taxonomy known as the O*NET Database.

On the face of it, this methodology surely makes logical sense. A person doing a job completes a set of tasks; assessing whether or not those tasks can be automated by AI should tell you the extent to which that job can be automated by AI.

Alas, there are a couple of problems with this approach in practice. Firstly, the “tasks” in the O*Net Database are at such a high level, it’s impossible to make any kind of realistic assessment of whether or not they might be automated by AI. A few examples illustrate this clearly, but I’d encourage readers to click through to the database itself and download the relevant Excel file so you can confirm I’m not cherrypicking to suit my narrative.

Let’s start where O*NET starts, at the very top. Here are four of the thirty-one tasks listed for “Chief Executives”.

To what extent would you say AI can automatically, say, “Direct or coordinate an organisation’s financial or budgetary activities…”? Even without knowing anything about the current capability of frontier AI models, surely the answer is always “it depends”?

“Direct or coordinate an organisation’s financial or budgetary activities…” isn’t a task, but a high-level bucket of dozens, perhaps hundreds, or even thousands of smaller sub-tasks. Some of those might be able to be automated, some of the time, in some circumstances, some will never be, and some will be somewhere in the middle. It depends.

This isn’t just a function of the CEO role. Here’s a small subset of the thirty O*NET tasks for radiologists, an occupation at the forefront of the AI jobs panic ever since Nobel Laureate and AI pioneer Geoffrey Hinton claimed in 2016 that the medical profession should “stop training radiologists” because it was “completely obvious” that they’d be replaced by AI. (In the ten years since, the number of radiologists has increased considerably.)

Would you say it’s “completely obvious” that AI can automatically, say, “prepare comprehensive interpretive reports of findings”, or “recognise or treat complications during and after procedures”?

It depends. Once again, these “tasks” are incredibly superficial groupings of hundreds or thousands of smaller sub-tasks. Not for the first time, it turns out that the world is more complicated than the simplified logical models we use to make sense of it, a broader theme I wrote about in This Time is Different.

However, even if the AI labs used a much more detailed level of sub-tasks as their unit of analysis, there’s an even more fundamental issue with the reduction of a job to simple bundles of tasks: there’s a lot more to any job than just the tasks.

The reality of a job is much more complex than it superficially appears. We’ve seen this mistake before. It’s exactly where we started: the view from O*NET is strikingly similar to the view from the corner office.

The Sum of the Parts

What do you think of when you read the word “ethnography”?

If I had to guess, I’d say you’re probably imagining an eccentric European explorer type studying the strange rituals of exotic tribes in remote tropical locations.

How about photocopier technicians in corporate America? Me neither.

But photocopier technicians were precisely the subjects of a highly influential ethnographic study published in 1996. Julian Orr’s Talking About Machines is billed as “a story of how work gets done…a study of how field service technicians talk about their work and how that talk is instrumental in their success.”

A service technician himself before becoming a corporate anthropologist for Xerox, Orr’s somewhat eclectic professional background made him uniquely suited to observing and documenting the esoteric world of people servicing photocopiers. His basic research question was the painfully mundane sounding: “what do Xerox photocopier technicians do all day?”

The answer seems straightforward enough. Typical of a large multinational corporation, Xerox had extensively documented their service procedures, with well-defined processes to diagnose issues, classify faults, order new parts etc. The workflow was straightforward: identify symptom, run diagnostic procedures identify fault, make repair and/or order parts.

What Orr found was entirely different. Rather than the straightforward linear workflow in the service manual, the reality of technicians’ day-to-day work as Orr saw it was “continuous, highly skilled improvisation within a triangular relationship of technician, customer, and machine”.

Perhaps you’re thinking that this sounds like precisely the kind of overly-pretentious twaddle an anthropologist would produce. But what Orr witnessed was technicians constantly trying to interpret and triangulate between the often ambiguous signals from the machine itself and a customer who couldn’t necessarily describe the problem in a technically useful way.

Orr describes a fundamentally narrative element to the diagnostic process. Based on what they see from the machine, its own diagnostics, and what they hear from the users, the technician constructs a story to explain the machine’s behaviour.

Orr observed a broader role for those narratives, which technicians enthusiastically shared with each other. These “war stories” are crucial components of knowledge transfer between technicians, but also act as status markers, as technicians subtly compete to tell the most interesting or unusual story, and demonstrate their own prowess in solving particularly tricky problems.

It’s not hard to see a similar situation playing out in any number of occupations: doctors, nurses, mechanics, call centre operators, lawyers, to name just a few.

Even in an ostensibly highly technical role, technical knowledge isn’t the sole preserve of individual technicians, or organisational manuals, it’s a socially-distributed resource, transmitted through stories.

What’s more, even in a technical role, there is a significant human, relational component. Fixing the machine is only one arm of the triangle Orr described. Sometimes the problem isn’t the machine, it’s the users. Perhaps they’re using the machine in an unexpected way, or misinterpreting its behaviour and misdiagnosing the problem, or perhaps the machine isn’t malfunctioning but is still failing to meet the users’ needs.

As Orr puts it, “in the process of maintaining and fixing photocopiers, technicians maintain and fix social relationships”. This is a very long way from the O*Net view of work as a series of mechanical tasks.

It’s quite easy to imagine a company like Xerox analysing what their technicians do all day and identifying significant time wasted talking to customers, having coffee or lunch with other technicians and gossiping about photocopiers. It’s also quite easy to imagine such a company deciding that some kind of AI diagnostic tool trained on all the service manuals could eliminate wastage and increase productivity. But without the time “wasted” telling war stories, crucial knowledge is lost. Six months later, technicians are struggling to fix weird faults.

LLMs can ingest documented knowledge remarkably well, but it’s much harder to ingest what’s not written down: experience, context, practice. So much knowledge is tacit rather than explicit. As philosopher Michael Polanyi described, there is an inherent paradox:

“We can know more than we can tell.”

Michael Polanyi

LLMs are increasingly surfacing another inherent paradox. So often in the rush to automate, we conflate the process with the output, the activity with the purpose. LLMs provide the tantalising prospect of instantly creating the artefacts that are the currency of white collar work: reports, spreadsheets, powerpoint presentations. Often, however, the value of the artefact isn’t the document itself, but the difficult and time-consuming process of creating it, the thinking, argument, learning, and consensus that appeared as by-products.

Sometimes the product of work is not only what gets produced, but what the worker becomes by producing it. Writing is perhaps the classic case in point. I write essays and books for a few reasons, but maybe not the ones you imagine.

Writing is hard. It forces you to structure your thoughts, refine them into coherent arguments, and stress-test them as you prepare to put them out into the world. A very astute old boss of mine used to say, “if you write well you think well”. If you outsource the writing to an LLM, you run a significant risk of outsourcing the thinking too.

As many have observed, this severing of the relationship between outcomes and outputs is particularly pronounced for junior workers automating entry-level tasks. Although often unspoken, a crucial outcome of juniors doing menial things wasn’t the things themselves, but the learning the juniors gained.

Just as the work isn’t only the tasks, the outcomes aren’t only the outputs. How do juniors learn the skills and judgement to progress up the career ladder if Claude or ChatGPT do their work for them?

The View From the Ground

When it comes to work, the whole is so much greater than the sum of the parts. Although we often don’t realise it, the problem with breaking jobs down into tasks isn’t simply that the tasks aren’t granular enough, it’s that much of what makes work work isn’t a task at all. So much of so many occupations is relational and relationship-driven, social rather than mechanical. So much of work isn’t the work itself, but work required to enable the work: coordinating, scheduling, chasing, negotiating, explaining, reminding, prioritising, reprioritising, navigating.

“Life at work is a staple in our conversation, but we rarely talk about what we really do in the doing of the job.”

Julian Orr

I strongly suspect this is a major reason that the AI CEOs’ dire predictions of impending white collar jobs apocalypse have so far proved inaccurate.

The mythical AI jobs armageddon is only the latest and loudest instance of the underlying tendency, however. The “Gareth” conversation with which we started is one I’ve been having for decades, long before the current LLM boom. The failure to properly grasp the reality of work has long hampered the successful adoption of digital technology by organisations. Again and again, the seductive promise of tech vendors’ glossy demos gives way to disappointment as the messy human reality of work grinds the gears of imagined smooth, linear, mechanical processes.

As I wrote in my book BS AT Work, much of the pervasive bullshit that makes much of modern work so miserable for so many, is rooted in a naively rationalistic, mechanistic, techno-centric view of humans, organisations, and the world. Above all, it’s rooted in a vast oversimplification of what humans are and what work is.

So many flawed organisational interventions, technological or otherwise, fail because they confuse the map with the territory, the theoretical organisation with the real one.

Most technology implementations start with the question “how can we use the technology?” The right question is surely, “what is the work?” This is even truer for AI than for previous digital technologies.

Of course, if you’re familiar with my work, you might be thinking “well he would say that, wouldn’t he”. It’s not just me though. As I finished writing this essay in mid-September 2026, Microsoft published an article entitled What we’ve learned from Microsoft’s own AI transformation, which eerily echoes my thesis.

Their first lesson? Start with the business outcome, not the technology. They learned this the hard way:

“We initially treated AI like a traditional technology rollout: deploy the tools, provide training, drive adoption. We learned that access and usage do not equal transformation: a tool licensed to over 200,000 people does not change how the work gets done.”

Microsoft

They also observed the false promise of fixating on individual tasks. Their second lesson: Redesign the entire workflow, not just individual tasks.

Before we try to automate work, we need to understand what the work is. Not what the procedures manual says it is, not what the process map describes, or the job description articulates, but what it really is and how it really gets done. That can’t be ascertained merely by reading documentation, or even just asking people, it requires careful, detailed observation. As Microsoft put it:

“The people who do the work know where processes break down, where judgment matters and where AI could help — insights that no process map can fully capture.”

Microsoft

From 30,000 feet, things look simple. From the corner office or the ivory tower of a Silicon Valley AI Lab, work looks like a straightforward mechanical process, a series of O*Net tasks. Break jobs into tasks, measure which tasks AI can perform, automate the jobs away.

The trouble starts when you come down to earth.

This is where the edge cases live, where the work actually gets done. Where Gareth does the thing, where they still have the old version of the system, still rely on paper processes, still use fax machines. Where people negotiate and improvise, tell stories and build relationships, learn things they can’t explain and do things that aren’t written down.

Down here, the world is messier than it looks from up there, and more human. Before we can predict which work AI will replace, or decide which work it should, we need to look closer.

We need to understand what the work is.