Bullshit Jobs, Now Better With Agents!

We are entering autumn 2026 as I publish this. The world is riding high on the expectations that AI will change everything. Despite other troubles, the stock market is at an all-time high, and several AI companies are preparing to go public at valuations that exceed the GDP of most nation states.

The people building AI are treated like royalty, and they are certainly not short of grand visions that resemble empires. And since they are convinced they are rebuilding society, they also have a lot of ideas about how their subjects will fare. Not to go too deep on this, but they don’t seem to necessarily like each other much. But neither did the King of England and the King of France at the time, so this is to be expected.

They do, however, agree that AI will impact the world in ways that we have never seen, especially in the job market. They have some different ideas about how this will turn out.

Dario Amodei, who runs Anthropic, warned in 2025 that it could wipe out half of all entry-level white-collar jobs within five years and push unemployment to 10 or 20%.¹ He is also openly warning that his creation may just destroy the planet, erase humanity, and/or cure cancer, depending on which comes first.

Elon Musk, on the other hand, says work will become optional, abundance is upon us and robots will do all the hard work. He is known to frequently sleep on the floor of his factories because of how much he works and expects anyone in his companies to do the same, so clearly we are not there yet.

Sam Altman of OpenAI envisions an “Intelligence Age” where artificial intelligence and infinite energy create extreme economic abundance, solve major scientific challenges, and drastically elevate human capability.² And there will be enough prosperity that we can pay a Universal Basic Income to everyone. OpenAI’s chief financial officer also suggested the government should backstop the enormous data centre financing obligations the company entered into, in case it can’t repay the loans. As of now, OpenAI is estimated to lose about $1.50 for every $1 of revenue, and requires an extra trillion or so in additional capital to build more compute. It looks like we need to wait a bit longer for the free money.

In any case, depending on who you listen to, and which parts of their essays you read, the options are clear: either nobody will have a job or nobody will need one.

Both prophecies are older than the technology. The Luddites broke stocking frames in 1811 because the machines were taking their work, and 10 years later David Ricardo added a chapter to his Principles conceding that the workers’ fear of machinery was “not founded on prejudice and error.” In 1930 John Maynard Keynes made both predictions in a single essay. “Economic Possibilities for our Grandchildren” warned of a new disease he called technological unemployment, and promised that by 2030 the grandchildren would need to work only 15 hours a week.³

It’s nearly 2030. Productivity in the rich countries grew about as fast as Keynes expected. Mass unemployment didn’t follow, and neither did the 15-hour week. The grandchildren work about 42 hours a week. A lot of them have to enter those hours into a tool called Workday, a very popular HR software that has yet to rebrand itself as Workweek and reject entries above 15 hours. It has already branded itself as AI-native, so by the time you read this, that may well have changed.

Aside from the amount, the content of the work is also quite different from what Keynes probably envisioned. Many of us spend a good share of those hours on things we struggle to call work.

The anthropologist David Graeber explored why that is. In a widely shared essay in 2013, he asked where those hours went that the machines seemingly had freed up.⁴ Machines did remove work, but instead of a joyful life of painting and playing football outside, we invented new jobs that are using up as much of our time as before.

Graeber looked at the new types of jobs, especially those created in the services sector. They were well paid, many of them came with important titles, and somehow involved managing someone or something. Many of those jobs are arguably higher-value work than the jobs of the past, even if a good chunk of the individual tasks they are composed of are objectively quite mundane, and those who do them find it at hard to derive a feeling of purpose and pride. That is not to say the assembly line gave anyone purpose by itself. But the assembly line had to exist for the cars to get built, and the outcome was very tangible.

His provocative statement was that many of the new jobs didn’t need to exist in the first place, and that the people doing them know it. He later turned the essay into a book, Bullshit Jobs, where he grouped them into five kinds⁵:

Flunkies exist to make someone else look or feel important.
Goons exist because the other side employs them.
Duct tapers fix problems that shouldn’t exist but are not allowed to fix the root cause.
Box tickers let an organisation claim it’s doing something it isn’t.
Taskmasters supervise people who don’t need supervising, or invent work for others to do.

Graeber was himself a tenured professor at the London School of Economics, an institution that is world leading in the study of effective resource allocation. His own observation was that the number of administrative jobs in academia had grown much faster than the number of students they serve, and the amount of attention allocated to individual students was, if anything, less than before.

Anyone who has worked in a growing organisation knows that at some point you do need some supporting departments so that the domain experts in whatever field they work in can focus on their expertise. That overhead sometimes becomes a little bit too independent and too large, and then, out of necessity or not, creates its own overhead to help itself focus, and so on.

Interestingly enough, the jobs that are seemingly most susceptible to being eliminated by AI automation sound a lot like those. So if Graeber was right, AI won’t end work any more than the stocking frame did. Which leads to the question: What if the agentic enterprise isn’t here to release the productivity our organisations have been holding back, but turns out to be just Bullshit Jobs, now with agents?

The productivity tools sell a faster status update and a summary of the meeting. The enterprise pilots promise whole departments run by agents. In the bid team the agents write the tender response the buyer’s agents will score. Take any such department apart task by task and ask what Graeber asked of the jobs: if this stopped tomorrow, would anything be worse?

Some of the people who pushed hardest for adoption are already walking it back. In April 2025 Tobi Lütke, Shopify’s chief executive, told staff that reflexive AI use was now a baseline expectation.⁶ In September 2026 he described the documents some of them generate, don’t read and pass on. “We call those ‘slop grenades’ that people toss at each other,” he said on a podcast. “You don’t really read it, and now it has to be reviewed by your colleagues.” It’s to his credit that he said so. The remark also shows where the cost went: to the person on the receiving end, and no business case counts their time. A good share of the work handed to agents exists for the same reasons the bullshit jobs did. Automating it doesn’t make it necessary. It makes it cheap, and organisations do what they always do with cheap things. They make more of them.

The software vendors answer it too, without meaning to. In July 2026 Salesforce announced a $1 billion investment to help Swiss organisations “become agentic enterprises”, and the press release listed what its customers’ agents already do. A virtual care provider for weight-related conditions has agents handle more than 300,000 customer messages a month, “deflecting 50% of inquiries”. Now, the choice of “deflecting” to describe value created gives us an indication as to why AI was needed. A human wants to engage with you, the company. Isn’t that beautiful? If you feel the inquiry must be deflected rather than answered, the reason for the inquiry probably lies somewhere in the product or the policy. For any company out there looking to implement the same, there are bad news on the horizon. There are a lot of new companies out there selling AI agents that follow up with this AI agent and get past the deflection attempts.

Meanwhile, the World Economic Forum built “EVA”, an agentic concierge that generated briefing documents in seconds for more than 3,000 global leaders at Davos. Which is great, but considering nearly every one of them travels with a team of assistants paid to brief them, there is reason to believe few of them were ever going to read the text. It is entirely plausible that before, the WEF employed a department that produced those briefing documents, so in isolation, AI really did reduce some late-night work in a copywriting agency in Geneva. Realistically though, in the year 2026 a personal assistant would have used an AI to summarise the briefing and pass it on, whether it was initially generated by a human or an AI.

I started my career writing code for business software. That was somewhat well defined and could be measured, imperfectly, by lines of code written, story points, functional software. Life made me drift into something else, into the loosely defined work that “consultant” covers. Consultants like me get paid to build agents like these, and we get measured by billable hours. Plus expenses, of course.

So what are the new human jobs that are being created because of agents? AI Governance Departments, AI Ethics Trainers, Agent Platform Release Managers. AI Coaches for the newly hired AI-native developers whose AI-generated resumes show they have more than five years’ experience with technology released just a few months ago. They were the only ones who got past the HR tool’s new AI screening feature into the interview stage. Of course, reviewing the AI-generated job description would have avoided that situation, but it’s too late now.

As we know, saving money is often very expensive. Before most organisations can point to any AI agent that pays for itself, they have to buy the means to govern them. Agent registries, observability platforms, sandboxes, the consultants to install them and replace them every six months because there is a newer, better option available.

This is great, and if implemented well, it can unleash productivity and abundance for society. And/or kill us all, depending on which Silicon Valley tech leader’s essay you read. But of course, it all needs to be organised. And to organise you need rules. Lots of rules. Laws if possible, but not too strict ones; we don’t want government to stifle innovation.

The European Union brought the AI Act into force in August 2024, before most of the companies it covers had an agent worth governing, and in 2026 pushed the high-risk obligations back to the end of 2027 because neither the technical standards nor the companies were ready.⁷ The Act asks for what the governance catalogue sells: a risk class for each system, technical documentation, logs, a person who oversees, and AI literacy for everyone who works with the systems.

It has all the bells and whistles of a European compromise. And like many European attempts to rein in American Big Tech, it is well intended. But this time enforcement requires deep expertise and access to internal test systems the frontier labs themselves struggle to control. Time will tell if a €4,180 monthly base salary before taxes and deductions in a dimly lit Brussels office can attract and retain the talent with the required qualifications and ambition that would otherwise be paid 10 times that, plus millions in equity, at a frontier lab in San Francisco.

The premise of most commentary is that AI will take away the busy work that people were overqualified to do anyway. The first draft, pulling numbers from some system, or finishing a presentation deck with perfectly aligned boxes and fonts. For knowledge-work organisations built around assumptions of the past, this means the cost comes down and the productivity goes up. They don’t need the people who do the busy work, and in many cases they have stopped hiring them, so that part is true.

The story that builds on top of it comes from economic history. When an input becomes cheap, value moves to its complements. Cheap steel in the 1850s made the engineers who could design with it valuable. Applied to AI, the story says that when output is cheap, the scarce and well-paid work becomes judgment: deciding what to produce and checking it. Vendor keynotes tell it, and so does every chief executive with a workforce to reassure. It’s a comfortable story, because the work it promises is the work people like to think they do. Nobody’s Tuesday is mostly judgment.

Trouble is all of this assumes that there are rules. And those rules can be written down, and the judgment is informed by and accountable to those rules. As a society, we generally like rules, because rules remove uncertainty. As a species, we don’t like uncertainty. It makes us nervous whether the uncertainty comes from possibly being eaten by a sabre-toothed tiger or audited by a cybersecurity function. Rules serve as a substitute for trust, and they are useful for the situations where we don’t know if trust is warranted. David Graeber wrote another great book, The Utopia of Rules, where he covers this phenomenon in depth.

Judgment is exactly what the agentic enterprise sets out to write down. It is in many ways just one more attempt to measure and control all business processes and put them in a box. In enterprise software we have done this before, many times over, with the same outcome. Anyone who has contributed to a large ERP implementation knows the exception path becomes the default business process for many more activities than it was originally intended for.

Where judgment can’t be written down, organisations tend to keep the ceremony of it as a substitute. If there is no rule, it means it could be wrong, and making a decision for which there is no rule requires a pre-emptive defence. So organisations pay for the reviewer’s signature, the supervisor of the reviewer, and they review whether the review has been signed off. These days, this usually goes along with a dashboard, or some kind of database, and lots of email notifications. All of which is delivered by an expensive software system that is installed and deployed by consultants like me.

So it appears that the surplus and abundance we are promised mostly goes elsewhere: to the AI tokens, the platform’s usage fees, and a new layer of people employed to stand near the new machine.

If value were moving to judgment inside organisations, you could see it where organisations record what they value: a pay band for reviewers, or a decision right that moved to the person with the context. Most organisations seem to have a review queue staffed by whoever was already there, and have cut junior hiring so that the pyramid keeps its shape with fewer people at the bottom. Those are the marks of a ceremony: the organisation adopts the vocabulary of the story, orchestrators and verification specialists, and leaves pay and authority where they were. We see a similar pattern with the late stage Agile transformation.

So what about the success stories? The many companies who claim in their earnings calls that they became AI-native overnight? If you look under the hood, often the reality is closer to: “we hired a consultant to strap a chatbot to our website, and now it can recite your insurance policy in the form of a poem”. Which is great for explaining growing IT costs and guiding expectations into the next year. It does not yet pave a clear path to abundance and the 15-hour work week.

Of course, some value does sometimes move to judgment. But it’s hard to measure, it moves less reliably, and to fewer people than was expected by the optimists. And the story is repeated most often by the people who benefit from its being believed.

Graeber’s Iron Law

Much of the work agents are supposed to automate exists to show that a rule was followed. Even in customer support, there is a handbook on when to close a ticket, give a refund or escalate to a manager. The rule materialises as a form on a screen in one way or another. And the form exists because somebody asked for it. And this person asked for it because they need a number to fill in on a form further down the chain, where an auditor asked for something that depends on it. But most importantly, the auditor asked for it because a regulator wanted evidence that someone had checked.

Graeber had described how such work multiplies in The Utopia of Rules. He called it the iron law of liberalism: every reform meant to cut red tape ends up producing more of it.⁸

AI aimed at this work inevitably follows the same law. A model can fill in the form, but the form was evidence that a person had taken responsibility. Once it costs nothing to produce, it stops being evidence, and the institution does what it always does and adds a rule. A second reviewer now checks the AI-drafted submission. The job survives, rearranged around checking the machine.

The chain runs further up than the auditor. When an agent leaks customer records or turns a job applicant away for a reason that would be considered discriminatory if it were done by a human, this is a tragedy-of-the-commons problem. And the solution to that is collective bargaining, which is best implemented by the government, as it is meant to represent the collective interests. So it is the mandate of the state to step in, and the state’s instrument is a rule and a record that the rule was followed. Graeber called the result total bureaucratisation: public and private rule-making grown into each other until nobody can tell them apart.

Europe has run the tech-regulation-by-making-everyone-fill-out-a-form experiment once already. The General Data Protection Regulation, applied from 2018, required a data protection officer, and within a year an estimated half a million organisations had registered one.⁹ It also set the consent rules for the cookie banner, which everyone clicks and nobody reads. Arguably, the privacy situation got worse because of GDPR. As the type of consent was not strictly regulated, to protect the freedom to draft a contract between two parties, we all just waive all our rights and share our information with 1,671 trusted partners. Which at this point just means sharing it with everyone, but without recourse, because this time we explicitly consent. In the end, a handful of the largest tech giants have paid most of the fine money so far, but kept the data flows the regulation was written to change. The fines were much lower than the profits, of course, so they were really a cost of revenue, not a deterrent. Everyone else paid for the consultants to implement the cookie banners, even if they never had a business model that would have benefited from monetising user data.

The AI Act points the same instruments at AI. Its literacy duty is the purest case. The Act attaches no fine to it, and in 2026 the Union rewrote it so that companies need only take measures to support AI literacy, with no level anyone has to reach and no set content.¹⁰ What it reliably produces is a training record. Inside an organisation each of these duties becomes a form, and the rule-maker is outside the building, where nobody inside can switch it off.

In a world run by written rules, nobody has to understand anyone else. An agent is that world’s ideal employee, because it acts only on what has been written down. Whatever an organisation can’t write down goes to a person as an exception.

Graeber also asked who does the understanding. Whoever imposes the rules doesn’t need to understand the people subject to them, so the people subject to them spend their effort working out what the powerful want. He called this interpretive labour. Most people have watched a capable colleague learn to write the report the steering committee expects, not the one the data supports. Nobody has to act in bad faith for that to happen.

There is very little evidence that agents will be exempt. An agent works to the objective it’s given and is evaluated by the metrics the organisation already uses. Models trained on human approval already lean towards telling the person asking what they want to hear. Inside a reporting chain that rewards good news, an agent ends up where the colleague did, only faster.

Towards everyone else, agents are on the upper side of the relationship. The customer rephrases the problem until the bot accepts it, and a coordinator checking its tickets learns which part numbers it tends to invent. The agent doesn’t need to understand any of them.

That’s also why the AI conversation keeps returning to a small set of repetitive tasks. They are easy to demonstrate and easy to count, so they look like the waste. The waste is usually the rule that produces them, or the politics that keeps the rule alive, and a model removes neither.

Managerial Feudalism

Graeber thought the reasons the hours got filled were moral and political rather than economic. Work is treated as a discipline that adults ought to submit to, whether or not it produces anything, and a manager’s standing inside an organisation is measured by the number of people who report to them. He called the second part managerial feudalism: executives surrounded by retinues, as lords once were, because the retinue is the rank.

An agent rollout runs straight into both. The transformation office that promises to cut a department by half is asking a director to give up the thing their grade is calibrated to. The director will find work for the people, and now has a tool that produces work on demand, starting with dashboards and reviews. The department head who used to manage 14 analysts becomes the head of an AI centre of excellence with 12, and the chart looks much as it did.

The same instinct explains why “work will be optional” is a story you hear from people who own the systems and never from people who run departments. Nobody inside an organisation is offered the option. What the productivity produces, inside a hierarchy, is the same headcount doing more, or fewer people doing the old work plus a new layer of supervision of the machine.

What cheaper output does change is cost structure. Most of the knowledge economy’s structures exist because thinking is expensive and the central problem of organisation is therefore how to deploy expensive thinking efficiently.

Consulting shows the arrangement most plainly. All those firms hire 25-year-olds out of university and bill them out at a hefty markup. The senior partner’s thinking is worth more per hour, but the firm can’t bill enough hours of it to cover its fixed costs, so it trains junior staff to produce an acceptable approximation of that thinking at scale and keeps the difference. The model depends on the junior production layer being necessary.

AI makes drafting and coding cheap enough that the pyramid has to change. The partner-to-associate ratio and the training pipeline that turned first-year hires into third-year producers were built for the old cost, and built so the spread flowed to the partners. What the partners do with the freed capacity is the question the keynotes skip, and the easiest answer is to stop hiring the 25-year-olds and keep the pyramid’s shape.

Software First

Software engineering is where the sequence has run furthest, and it’s where the five types appeared first.

Coding agents can take a goal and iterate against the tests until they pass. What used to require a junior developer and a full day now requires a specification and 20 minutes. Engineers were the first to have it demonstrated on their own work.

They were also the first to receive a mandate. Engineering leaders who had tried the tools and found them useful became engineering leaders who required their use and tracked it. The first slop grenades were pull requests: generated code, generated description, generated unit tests. A reviewer who now had to read both. The first duct tapers were the senior engineers who spent their afternoons on the review queue, and/or fixing production. Much like they did with junior developers to be fair. Login-counting reached other professions later, after it had been rehearsed on developers.

Programming, the translation of intent into code, has long being commoditised. Software engineering is the management of complexity across systems that keep changing, and of the people who build them. Cheap programming creates more of that complexity, and whether anyone is paid to manage it depends on whether the organisation notices it’s there. Other knowledge professions will go through a version of this once the tools for their work are as usable. The details will vary. The sequence, tool, mandate, review queue, governance, probably won’t.

Most jobs, in software and elsewhere, are a visible product on top of coordination and memory nobody sees. AI may cheapen the product and leave the rest intact. The unit worth watching is the workflow as it moves, from the person who enters the data to the exception that comes back to the same desk.

Auditing Your Own Week

Before reading further, try the five types on your own week.

Take last week’s calendar and inbox, whatever your job description says. Mark each recurring item by what it is for. Does it exist so that someone senior is kept informed or looks prepared? Does it exist because the other side, a regulator or a competitor, does the equivalent? Does it repair something that shouldn’t have broken? Does it produce evidence that something was done, rather than doing it? Does it supervise people who would do the work without supervision? Where an item exists because of a rule, ask whether the organisation made it or hosts it for a regulator, and whether anyone knows which. Leave unmarked whatever changes something in the world for a customer or a colleague.

Expect the marked share to be higher than you guessed, and clustered around reporting and review.

For each marked item, ask who would notice if it stopped for a month, and what your organisation would most likely do if an agent could produce it for nothing: stop it, or produce more of it.

Checklist

Which parts of your role never appear in the job description: the coordination, the standing responsibility, the informal memory, the administrative overhead that has built up around it?
Take last week’s calendar and inbox. Which recurring items keep someone senior informed, answer the other side, repair a defect, produce evidence or supervise people who’d do the work anyway? What’s left unmarked?
Where an item exists because of a rule, did your organisation make the rule, or does it host it for a regulator? Does anyone know which?
For each marked item, who would notice if it stopped for a month? If an agent could produce it for nothing, would your organisation stop it or produce more of it?

Notes

Axios 2025.Source ↩
Altman 2024.Source ↩
Keynes 1930.Source ↩
Graeber 2013.Source ↩
Graeber 2018, ch. 2.↩
Lütke 2025.Source ↩
European Parliament and Council 2024. European Parliament and Council 2026.Source Source ↩
Graeber 2015, Introduction.↩
IAPP 2019.Source ↩
European Parliament and Council 2026, Art. 4(1) AI Act as replaced.Source ↩