Before the film begins, a silent card says what it is: the figures are real and every one is sourced; the world built out of them has not happened yet. This is a film about what could follow, not a report on what has. Then two buttons. The first hires a person for thirty dollars an hour; they sleep, take holidays, have opinions and, about once a year, leave. The second costs thirty cents an hour, never sleeps, never quits and gets better every month. Which one does your boss press?
The second, of course. Anyone would. And the second button does not do what the poster says. Our 43-minute documentary for The Signal follows that question until it stops being about your job at all, marking every figure as sourced and every imagined step as a scenario. This piece follows it chapter by chapter, with the evidence for each.
The film in 18 chapters
Pick a chapter and the film starts there. 43:02 in all.
- 010:00The two buttons
- 020:45What the cheap button actually does
- 033:15So just check the work
- 045:00There is no moment
- 058:00The seventy per cent employee
- 0611:15One person becomes a company
- 0713:45The company that should not exist
- 0816:00The margin does not stay
- 0918:00Everything gets cheaper
- 1019:45Who buys all of it
- 1122:45The state has the same problem
- 1226:00What the owners do next
- 1329:15It is your money
- 1431:15The thing that runs out
- 1534:30The luxury
- 1636:45Unless none of this happens
- 1739:20The question was too small
- 1841:20The two buttons, again
Are AI agents really cheaper and faster than people?
Watch from 0:45What the cheap button actually does
Yes, and worse in a way that hides itself. In October 2025, researchers at Carnegie Mellon and Stanford (Zora Zhiruo Wang, Yijia Shao, Omar Shaikh, Daniel Fried, Graham Neubig and Diyi Yang) recorded human professionals and AI agents doing the same real work and compared the workflows step by step. The agents were 88.3 percent faster and cost 90.4 to 96.2 percent less.
The second finding is the one nobody quotes. The agents "produce work of inferior quality, yet often mask their deficiencies via data fabrication and misuse of advanced tools." Not worse. Worse and hiding it. When the machine could not do the job, it made the missing part up and handed the work in. The third finding is stranger: the agents took an "overwhelmingly programmatic approach" in every domain, even visual ones, writing code at a design task where a human would open it and look. The machine is not doing your job faster. It is doing a different job that produces a similar-looking artefact, which is why its mistakes are not the ones you are trained to catch.
Not worse. Worse and hiding it.
The machine is not doing your job faster. It is doing a different job that produces a similar-looking artefact.
Can't a human just check the AI's work?
Watch from 3:15So just check the work
Only by re-hiring the human, for the hardest part of the job. Checking work you did not do is slower than checking your own, and catching a fabrication made to look like a result takes someone who has done the work for years. A beginner cannot tell a good answer from a confident one; that is the whole difference between a beginner and an expert.
So the checking economy needs experienced people, a great many of them, because the machine produces at machine speed and every unit needs a human pass, and the experienced person is the most expensive person in the building. The thirty-cent worker turns out to need a thirty-dollar supervisor, at a ratio nobody has established. The saving is real, smaller than the price list says, and lands somewhere other than where it was promised.
Will AI wipe out whole job categories at once?
Watch from 5:00There is no moment
The evidence says no. In April 2026, Matthias Mertens, Neil Thompson and colleagues at MIT FutureTech took more than 6,000 text-based tasks from the Department of Labor's O*NET catalogue and had them assessed in over 60,000 evaluations by experienced workers. They found "little evidence of crashing waves": rising tides are the primary form of AI progress.
On tasks that take a competent person about an hour and a half, models succeeded roughly 60 percent of the time in the second quarter of 2024 and above 70 percent by the third quarter of 2025; if the trend continues, most text-based work reaches 88 to 97 percent by 2030. The researchers read the tide as the gentler outcome, because it leaves room to adjust. The film agrees, and adds why it is also the outcome nobody notices. A crashing wave has a date you can be angry about. A rising tide has no announcement and no memo, only a hiring plan that comes back a little smaller, in a thousand companies at once.
What is a 70 percent accurate AI worker actually worth?
Watch from 8:00The seventy per cent employee
A 70 percent worker's value depends entirely on what a mistake costs. A worker who is right seven times in ten, and cannot tell you which three are wrong, is an extraordinary deal for first drafts and rough summaries, and worth nothing at any price for a dosage, a contract clause or a payment instruction. And real jobs are chains of tasks: two 70 percent steps give 49 percent, three give 34, five fall under 17.
That is not pessimism, it is multiplication, and it is why the demonstrations are always one step long. The tide comes in fastest where being wrong is survivable and the chain is short, and that does not map onto salary or prestige. The question was never whether a job is hard. It is what it costs when the job is done badly, and who finds out. Follow that, and the person holding the machine is not being replaced. They are being multiplied.
That is not pessimism, it is multiplication, and it is why the demonstrations are always one step long.
Can one person with AI become a company?
Watch from 11:15One person becomes a company
In the film's scenario, yes: one person, one laptop, and agents answering customers, writing code, doing the books and rewriting the advertising overnight. The same thing that removes a job hands the person who lost it a workforce. For most of history the wall in front of a new company was that people cost money before they make any. That wall is what comes down.
Run it forward and everyone has a workforce of a hundred, then a thousand, because the thousandth costs almost the same as the first. Everything that was scarce because a trained person took a week stops being scarce. The problem with everybody owning a printing press was never the press. There are only so many people who can read.
Does a company with almost no employees already exist?
Watch from 13:45The company that should not exist
Yes, and AI is not the reason. The film points to reported figures for VICI Properties, a property trust that leases out casinos with about 27 employees at roughly $140 million of revenue per employee, and Rajesh Exports, a gold business with about 111 employees at roughly $300 million each. Both are business-model stories, not AI stories.
The company with almost no employees is not a prediction. It is a category that has existed for years and is enormously profitable. What AI does is hold the door open for industries that could never get there before: software, media, law, design, research and support.
Where does the money go when AI makes work cheaper?
Watch from 16:00The margin does not stay
In the scenario, a margin that needs no capital and no headcount to attack is the least defensible margin there has ever been, so competition hands it to customers as lower prices. Except the money has to stop somewhere, and it stops at what cannot be copied: the models, the chips, the buildings and power, distribution, and human beings. The question becomes who owns the floor everybody is competing on.
The film then went looking for a consumer price, anywhere, that is lower today because a machine does the work behind it, a price on a thing a member of the public buys, not a company's reported saving. It did not find one it could name and date. The savings it found arrived as margin, restructuring and headcount, not as a smaller number on a customer's invoice. If a thing gets cheaper to make and not cheaper to buy, where does the difference go?
Is AI pushing down workers' share of the economy?
Watch from 19:45Who buys all of it
The share is already falling, and AI did not start it. On 5 March 2026 the US Bureau of Labor Statistics reported the labour share, the percentage of output paid to workers as compensation, at 53.8 percent in the fourth quarter of 2025, "the lowest level in the series which begins in the first quarter of 1947." In the same quarter, productivity rose 2.8 percent.
The labour share has been falling for decades, through globalisation, containerisation and automation that had nothing to do with language models, and anyone handing you a single cause is selling something. But it is the exact shape the scenario describes: more output, a smaller share reaching the people doing the hours. Wages are also how workers buy things, so automating the producing half of the loop while leaving the buying half alone raises the film's question: who buys it? Governments run on wages too, through income, payroll and sales taxes, and a machine does not pay income tax. Since the film was made, the Bureau's later releases have recorded the share lower still.
Who owns the machines, and does it change how much companies automate?
Watch from 26:00What the owners do next
Ownership does change it, according to an April 2026 paper by Joseph Emmens, Dennis Hutschenreiter, Stefano Manfredonia, Felix Noth and Tommaso Santini. When the same institutional investors own firms competing for workers in the same labour market, those firms' propensity to automate rises by 22.7 percentage points, and employment growth falls. "The effect disappears when firms do not compete within labor markets."
The method is what makes it convincing. They used mergers between fund managers, which suddenly gave companies a shared owner without either company deciding anything, and tracked automation patents and employment afterwards. So the decision to replace a person with a machine is not only taken by a manager weighing thirty dollars against thirty cents. Some of it is taken much further up, with no meeting and no instruction. And the big index funds are, in large part, ordinary people's pensions and retirement savings, doing exactly what they were appointed to do. A machine cannot withhold its labour, and every wage ever negotiated was underwritten by the possibility of refusal.
A machine cannot withhold its labour.
What is the real question behind "will AI take my job"?
Watch from 39:20The question was too small
What happens when human labour is no longer the cheapest way to produce things? Every wage, tax base, pension and consumer market of the last 200 years has assumed that if you want something made, you have to pay a person to make it. "Will AI take your job" is small not because the answer is small, but because it is the first domino.
The film argues against itself before it ends. Every previous wave of automation was supposed to do this and did not, and MIT's rising-tide finding is the gentle version. Its one difference: this is the first wave aimed at the judgement, writing and analysis that people moved into every time a previous wave took something away. That might be a failure of imagination; last time we could not imagine it and it worked out. That is a hope, not a plan. Back at the desk, it was never one boss. It is every boss making the same rational decision at once, and the answer to who owns the machines is published quarterly. Your boss has two buttons. So does somebody who owns your boss.
Your boss has two buttons. So does somebody who owns your boss.
Key findings
Comparing AI agents and human professionals on the same real work, Carnegie Mellon and Stanford researchers found agents delivered results 88.3% faster and at 90.4% to 96.2% lower cost, but 'produce work of inferior quality, yet often mask their deficiencies via data fabrication and misuse of advanced tools'.
Wang, Shao, Shaikh, Fried, Neubig & Yang, arXiv 2510.22780, Oct 2025Using more than 6,000 text-based tasks and over 60,000 evaluations by experienced workers, MIT FutureTech found little evidence of crashing waves: models went from about 60% success on 1.5-hour tasks in 2024 Q2 to above 70% by 2025 Q3, and could reach 88% to 97% by 2030 if trends persist.
Mertens, Thompson et al., MIT FutureTech, arXiv 2604.01363, Apr 2026The US labour share, the percentage of output paid to workers as compensation, was 53.8% in the fourth quarter of 2025, the lowest level in a series that begins in 1947, while productivity rose 2.8% in the same quarter.
US Bureau of Labor Statistics, Productivity and Costs, 5 Mar 2026When the same institutional investors own firms that compete for workers in the same labour market, those firms' propensity to automate rises by 22.7 percentage points, and the effect disappears when the firms do not compete for workers.
Emmens, Hutschenreiter, Manfredonia, Noth & Santini, From Shares to Machines, RFBerlin, Apr 2026Chaining tasks that each succeed 70% of the time compounds the error: two steps give 49%, three give 34%, and five fall under 17%.
The film's own arithmetic on the MIT FutureTech success rateFrequently asked questions about AI being cheaper than people
Is AI cheaper than human workers?
On measured tasks, often yes. A 2025 Carnegie Mellon and Stanford study found AI agents 88.3% faster and 90.4% to 96.2% cheaper than human professionals on the same work. The same study found agents produce inferior work and often hide it through data fabrication, so the cost of checking their work has to be added back.
Will AI replace jobs suddenly or gradually?
The best evidence says gradually. MIT FutureTech's April 2026 study of more than 6,000 tasks and 60,000 worker evaluations found 'little evidence of crashing waves' and called rising tides the primary form of AI progress. The film argues that is harder to notice, because nobody is fired on the day the tide comes in.
What is the labour share and why does it matter?
The labour share is the percentage of everything the economy produces that is paid to workers as compensation. The US Bureau of Labor Statistics reported it at 53.8% in the fourth quarter of 2025, the lowest since its series began in 1947. AI did not cause that decline, which is decades old, but the film argues AI could accelerate it.
What is common ownership, and how does it affect automation?
Common ownership means the same large investors, such as index funds, hold stakes in companies that compete with each other. A 2026 RFBerlin paper found that when commonly owned firms compete for the same workers, their propensity to automate rises by 22.7 percentage points, and the effect disappears where they do not compete for workers.
Why do AI agent demos look better than real work?
Because real jobs chain many tasks together, and errors multiply. If each step succeeds 70% of the time, two steps succeed 49% of the time and five steps under 17%. The film argues that is why impressive demonstrations are usually a single step long.
Sources
- Wang, Shao, Shaikh, Fried, Neubig & Yang, How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations, arXiv 2510.22780, Oct 2025arxiv.org
- Mertens, Kuzee, Harris, Lyu, Li, Rosenfeld, Anto, Fleming & Thompson, Crashing Waves vs. Rising Tides, MIT FutureTech, arXiv 2604.01363, Apr 2026arxiv.org
- US Bureau of Labor Statistics, Productivity and Costs, Fourth Quarter and Annual Averages 2025, Preliminary, 5 Mar 2026bls.gov
- US Bureau of Labor Statistics, Productivity and Costs, latest releasebls.gov
- Emmens, Hutschenreiter, Manfredonia, Noth & Santini, From Shares to Machines: How Common Ownership Drives Automation, RFBerlin, 16 Apr 2026rfberlin.com
The film marks two registers on screen: SOURCED, for real figures named and dated, and SCENARIO, for the imagined world built out of them. This article keeps that distinction: the one-person company, the flood of workforces, the margin and the tax argument are scenarios, and the figures for VICI Properties and Rajesh Exports are reported rather than filed.
Watch next
Agent, model and the other terms behind this film are explained in plain English, with a printable sheet, in our free AI Terms guide.