AI · MBA

A single orange operator mark above a grid of agent nodes on a dark poster field

One operator, a fleet of agents

The MIT Project NANDA report of July 2025 handed the world the number 95 percent and dropped a far more interesting sentence on the way. In more than 90 percent of the firms surveyed, employees reported regularly using personal AI tools for work, although only about 40 percent of firms said they had bought an official language-model subscription. Access to the same models therefore lies on both sides of the line between success and failure. My claim is that the edge of a one-person operator with a fleet of agents grows out of something nobody invoices: the internal layer of coordination between employees drops away there. So more of the leverage goes into the output than in an organization. The decision that hangs on this is yours, and it is uncomfortable — hire the next person, or buy the next degree of autonomy for the agent you already have.

The number that changed its unit inside the report itself

The MIT Project NANDA team published The GenAI Divide: State of AI in Business 2025 in July 2025 (Aditya Challapally, Chris Pease, Ramesh Raskar, Pradyumna Chari). It rests on a review of more than 300 publicly disclosed AI initiatives. To that were added interviews with representatives of 52 organizations and survey responses from 153 senior leaders, collected at four industry conferences; the research period was January to June 2025. The authors call their findings preliminary and caution that the six-month observation period may understate success rates for longer-term implementations. In one paragraph of its summary the report says that 95 percent of organizations are getting zero return. It also says that only 5 percent of integrated pilots pull millions in value out of AI, while the vast majority show no measurable effect on the income statement.

The first sentence is about organizations, the second about pilots.

The distinction is not pedantry. A company running twenty pilots of which one earns money lands both among the 5 percent and among the rest, so the second sentence says little about the company itself. Further on, the report attaches a 5 percent figure to enterprise tools as well as to enterprises, and does not say what sample it was derived from. It also cautions that the figures are directionally accurate and based on individual interviews rather than on official company reporting, and that sample sizes vary by category. The findings rest on interviews and self-reported answers, the report is an initiative's document and did not appear as a peer-reviewed paper, and its appendix holds questionnaires but no data.

Let us name the mechanism straight away: the unit of the number changes within one paragraph of the summary, so whoever cites it has to pick one of the two, and nobody lies along the way.

The same document holds the finding my whole claim rests on. The authors write that workers at more than 90 percent of the firms surveyed reported regularly using personal AI tools for work, and they call it the shadow AI economy. The prose speaks of firms, while the chart on the same page labels the 90 percent as employees, a different unit, so I take the firm-level wording. I treat that figure as an order of magnitude, because it comes from a survey and interviews rather than telemetry. An order of magnitude is enough. If in the overwhelming majority of surveyed firms, including those whose initiatives fail to deliver, somebody reaches privately for the same models I do, then access to those models stops explaining anything at all. This does not prove those employees work better for it. The report claims that this zone often yields a better return than official initiatives, but it shows no measurement, and picking up a tool is not yet a result or even the promise of one. In my view the barrier does not sit on the access side.

There is a third observation in that report, and it is the most sardonic of the three. The report says back-office automation often returned more, and some of its most dramatic documented savings come from replacing outsourced services and agencies. Yet in the split of a hypothetical budget that executives were asked to make, sales and marketing took about half, and in another place about 70 percent. Spending was not measured, so this tells us where executives say the money would go, not where it went. It still points one way: the report itself ties the front office's budget lead to how easily its results can be attributed and seen. Technology settled nothing here. The stated split of the budget followed the visibility of the result, and visibility is set by whoever signs that result off.

The model: coordination effort grows combinatorially

Frederick P. Brooks Jr. laid out this arithmetic in The Mythical Man-Month (Addison-Wesley, 1975), drawing on his own work on OS/360. If every part of a task has to be agreed separately with every other part, effort grows as n(n−1)/2. Four people means six channels. Ten means forty-five. Fifty means one thousand two hundred and twenty-five channels, and nobody has yet been issued a day longer than twenty-four hours. Hence Brooks's law: adding people to a late software project makes it later.

Brooks gave an analytical model and an observation from a single undertaking, not a study with a sample.

A related link between organizational structure and code quality was studied by Nachiappan Nagappan with Brendan Murphy and Victor Basili in The Influence of Organizational Structure on Software Quality (ICSE '08, Leipzig 2008). They took data from Windows Vista and asked what predicts whether a single binary fails after release. A model built on eight organizational metrics reached precision of 86.2 percent and recall of 84.0 percent. The metrics included how many engineers who touched the binary were still at the company and how many had since left it. They also covered how deep the ownership of the code ran and how many organizations each contributed more than 10 percent of the changes. That beat models built on code churn, code complexity, test coverage, dependencies, and the count of pre-release defects.

Organizational structure predicted failures better than metrics of the code itself. It is a case study of one product at one company, and a predictive correlation rather than a causal experiment, and it does not test Brooks's law about late projects. The authors themselves write that little data existed on the relationship. The direction agrees with Brooks's claim that organizational structure affects quality.

The second frame came from an economist. Ronald H. Coase asked in The Nature of the Firm (Economica, 1937) why firms exist at all, if the market supposedly allocates resources more efficiently. His answer: using the price mechanism costs something. You have to find a counterparty, negotiate terms, write the contract, enforce it. A firm comes into being to move those transactions inside and pay less for them. It grows for exactly as long as the cost of organizing one more transaction internally stays below the cost of arranging the same one on the market.

Agents lower one side of that equation, because part of the coordination between people stops being needed. A one-person firm can therefore take inside work that used to require hiring, and that is the entire economic case my one-person product studio with a fleet of agents stands on. Coase has a second half, though, which writing about AI keeps quiet: he also described the entrepreneur's diminishing efficiency as the number of transactions rises. One head has a throughput limit. An agent does not add a head. It adds hands.

The missing link I found in Nicholas Bloom, Luis Garicano, Raffaella Sadun and John Van Reenen, The Distinct Effects of Information Technology and Communication Technology on Firm Organization (Management Science 60(12), 2014). They pulled apart two things usually thrown into one bag labelled "digitalization". Information technology gives knowledge to the person at the bottom and widens their decision autonomy. Communication technology carries information upward cheaply and centralizes the decision.

That settles the argument over whether a tool by itself shortens the decision loop. It does not. It shortens the loop when the knowledge lands with the person who decides. For a one-person operator those two people are the same person by definition, so every tool works there as information technology. Inside an organization, the very same agent very often works as communication technology: it produces material for the next round of sign-off, a slide for the next review, a justification for the next budget request.

The last piece of the puzzle was supplied by Oriana Bandiera, Andrea Prat, Stephen Hansen and Raffaella Sadun in CEO Behavior and Firm Performance (Journal of Political Economy 128(4), 2020). They gathered time diaries from 1,114 CEOs across six countries and classified their behaviour without imposing categories in advance. Two types fell out. One kind sits in cross-functional and senior meetings; the other in one-on-ones with core functions. Firms of the first kind do better, except that the effect only surfaces after three years. The authors themselves put the difference down to the fit between style and firm rather than to one type being superior.

What interests me is something the authors had no reason to stress, because to them it was obvious. A chief executive's working material is the meeting. Not out of laziness. Inside an organization, coordination simply is the work, and the most expensive hour in the company goes into letting two other hours begin.

The mechanism: what an agent does not do, and what it takes in exchange

Take any task that passes through three divisions in an organization. Count what happens to it along the way.

Someone has to define the scope. Someone else has to approve it, because he carries the budget. A third person checks that the task breaches nothing she is accountable for. A fourth asks why it was not done in her team. Each of these people brings real knowledge and a real interest to the matter, and none of them is a villain. Together they produce a cost nobody books, because coordination has no line in the chart of accounts.

Now the same task at an operator with an agent. The agent does not negotiate scope. It does not defend a budget, having no budget to defend. It has no career to secure in passing, no division whose interest it represents, and no need to look busy in a quarter with nothing to ship. It takes an instruction and either delivers a result or does not. None of which means coordination inside an organization is waste, because a good share of it exists so that somebody says "no" in time, and that function an agent does not perform.

Work that decomposes into coordination inside an enterprise programme stays work for a one-person operator.

In that single sentence I see the cause of the disproportion visible in the NANDA numbers. I need no assumption that worse people work in corporations. The people there are usually better. They have better data, better infrastructure and larger token budgets than I will ever have. They also have more people who must be aligned before a decision becomes a result, and the number of possible pairs grows as n(n−1)/2.

Let us measure that loop, because without a number it stays a pretty sentence. Changing a price in my studio takes checking two things: whether the new price covers the cost to serve, and whether it breaks commitments to existing customers. An hour, sometimes two. In an organization the same decision passes through the product owner, finance, legal and communications, none of whom works on it continuously. Calendar time then runs in weeks, even where the sum of person-hours comes close to mine. The effort comes out comparable. The calendar does not come out comparable at all.

Honesty demands the other side of that ledger immediately, because it is just as mechanical. A firm does not keep those channels out of fondness for meetings. It keeps them because they buy things a single person cannot purchase at any price. Capital that lets you survive eighteen months without revenue. Distribution, meaning a list of customers you only have to call. Institutional trust, meaning that procurement at a bank will open your offer at all. And continuity: if the person running the project ends up in hospital, the project carries on.

The same property that lifts the internal layer of coordination off a one-person studio takes all four away. You cannot have one without the other, because this is a single parameter seen from two sides. Whoever does not do that arithmetic beforehand rather than afterwards is selling himself a fairy tale about freedom.

The leverage is narrowband, and the bandwidth can be measured

Both numbers go on the table, the uncomfortable one included.

Sida Peng, Eirini Kalliamvakou, Peter Cihon and Mert Demirer described a controlled experiment (The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, arXiv:2302.06590, 2023). The task was to write an HTTP server in JavaScript. Participants received a template repository with the task description and a code skeleton, and twelve tests decided whether the task was complete. Among those who finished, the group with an assistant was 55.8 percent faster, with a confidence interval from 21 to 89 percent. Worth recording alongside: the co-authors include employees of the company that makes the tool.

Now the second study. A METR team led by Joel Becker and Nate Rush ran a randomized experiment (Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, arXiv:2507.09089, 2025). Randomization ran at the level of the task: sixteen experienced open-source developers, 246 tasks in mature repositories where they had an average of five years behind them. Before the study, the participants forecast that their time would fall by 24 percent. After it, they judged that they had worked 20 percent faster.

The measurement showed they were 19 percent slower.

The gap between self-report and measurement came to thirty-nine percentage points. The confidence interval for that slowdown reached from plus two to plus thirty-nine percent, and the sample was sixteen people drawn from one population, so this is no verdict on the industry. It is, though, the best-documented case of practitioners being wrong about their own speed while doing the work with their own hands.

The most honest move in the whole story was made by the authors themselves. In February 2026 METR published a note on changing the design of the experiment. It judged the data from the newer round, begun in August 2025, an unreliable signal, mainly because more developers did not want to take part in the study without AI. It also wrote that developers are probably more sped up now than in 2025. The raw results of that round are a change in working time of minus 18 percent for those of the original developers who took part in the newer round and minus 4 percent for newly recruited ones. Minus means a shorter time here, and in both cases the intervals cross zero. The authors called it very weak evidence for how much the speed-up has grown since 2025. In addition, between 30 and 50 percent of surveyed developers admitted to not submitting some tasks, unwilling to do them without AI. So tasks with a high expected gain are missing, and the note treats these estimates as probably understating the effect.

The two studies differ in nearly everything at once. The 2022 Copilot study used the Codex model, and 95 freelancers from Upwork did one standard task from a ready template, and 35 completed it. The 2025 METR study mostly used Cursor Pro with Claude 3.5 and 3.7 Sonnet, and sixteen experienced maintainers of large repositories did 246 tasks of their own. The model, the tool, the participants, the task, the measure and the year all differ, so the comparison alone cannot say what flipped the sign. METR suspects familiarity with the repository and its size. It calls that evidence suggestive but uncertain. I have a hypothesis of my own, though nothing here proves it: the class of task may have changed the sign. Fresh and verifiable work differs from work embedded in a mature system with scattered context, where the experienced developer simply knows better. So before you count your leverage, name the class of task you are measuring.

How to build a fleet

A fleet of agents does not reduce to a count of subscriptions. What decides it is the split between tasks you release without looking and tasks you check before every use. The rest you do not hand over at all.

In practice my fleet looks simpler than the word suggests. A few parallel sessions, each with its own scope and its own rules file, none of them aware of the others. That is deliberate. The moment two agents have to agree with each other about a shared context, I am back inside Brooks's arithmetic, only paid in tokens and in my own time spent reconciling contradictory output. So I split the work along boundaries that are separable anyway: a separate repository, or a separate document.

The numerical footing comes from METR's measurements in Measuring AI Ability to Complete Long Software Tasks (arXiv:2503.14499, 2025). The metric is called the task horizon: the longest task a model finishes with fifty-percent reliability, measured in the time it takes a human expert. The horizon has doubled roughly every seven months since 2019. For a model from early 2025 it stood at around fifty minutes.

The number that actually drives the decision about autonomy is a different one. The horizon at eighty percent reliability is about five times shorter — for that same model we are talking about ten minutes or so. In a note on limitations from January 2026 the authors supply the rest. The error bars typically stretch by a factor of two in each direction, horizons differ by orders of magnitude between domains, and tasks that are critical for reliability and poorly verifiable demand success rates above 98 percent. Doubling the horizon does not double the degree of automation, because failures on longer tasks are more tangled and cost more human work to repair.

Autonomy, then, does not scale with the power of the model. It scales with the reliability the class of task demands — and that class is set by you, not by your vendor.

Out of this comes the delegation criterion that replaced every earlier heuristic I used. Reversibility times verifiability. The question is not whether the model can manage it, but what the rollback costs if it cannot, and whether I will even notice. Graduated autonomy and the mechanics of demoting an agent that has started to get things wrong I wrote out separately in agent demotions; here I leave the decision skeleton alone.

Class of taskReversibilityVerifiabilityDegree of autonomy
Draft, research, a variant of a conceptfull, rollback costs nothinghigh, you see the result at onceagent works alone, you read the result
Code in a project with tests and continuous integrationhigh, rollback is one commandhigh, but automated, not yoursagent works alone up to the green gate, you read the diff
Data migration, schema changelow, rollback costs a daymedium, the failure can be silentagent prepares, you execute
Customer communication, an offer, a pricenone, the word has been saidlow, the effect returns next quarteragent does not touch it
Decision under legal supervisiondepends on the statuterequires your signatureagent prepares the material, you answer for it

The criterion for taking a task back off an agent is shorter than the table. If verifying the output costs you more than doing the work yourself, take it back. Do that thing with your own hands. It is the same trap the experienced developers fell into in the METR study: they were given help with execution and paid for it, among other things, in time spent checking. The bill went unnoticed. The rule runs in the other direction too, and there it is more pleasant. Where verification is cheap and immediate, the fleet gives the most.

That leaves verification, the place where a one-person studio lies to itself most often. The operator is a poor sensor of his own speed; the thirty-nine-point gap at METR showed as much. He is a poor sensor of his own quality as well. Veracode, in its 2025 GenAI Code Security Report, put more than a hundred models through eighty programming tasks in four languages. The result: 45 percent of code samples introduced a vulnerability from the OWASP Top 10, the canon of the most common security defects in web applications. Java came out worst, failing in 72 percent of cases. Cross-site scripting — the injection of someone else's script into a page — was left unguarded in 86 percent of the relevant samples. This is a benchmark from a company that sells code scanners, so it has an interest in a result of that shape; the methodology is public, though, and the direction matches what I see in practice.

Hence the rule I do not negotiate with myself: I judge the output. I take no interest in whether the agent "sounded confident" or how many steps of reasoning it displayed. I take an interest in a green test, a diff I have read, a number I have checked, and a security scan that passed. Why measurement beats ceremony I took apart in governance without theatre, so the operational conclusion is enough here: a gate you cannot fail is not a gate.

My own rule is owed two caveats. First, verifying the output works only where the output takes a form that can be checked in finite time, and a large part of what people call strategy takes no such form. There the fleet is useless. Second, judging the output alone is enough only where an error can be undone after the fact. If during the work the agent can expose data or take an irreversible step, judging the output comes too late, so the row of the table applies in which the agent prepares and I execute.

Finally, the thing I never hand to an agent. That list I did not have to invent, because the legislator wrote it for high-risk systems. Article 14 of Regulation (EU) 2024/1689 on artificial intelligence requires such a system to be designed for oversight by natural persons. Those assigned that oversight must be able, as appropriate and proportionate, to understand the system and stay aware of a possible tendency to rely automatically on its output. To the same extent they must also be able to interpret that output correctly. To the same extent they must be able to decide not to use the system or to reverse its output, and also to stop it. The requirements for high-risk systems will apply from 2 December 2027 (Annex III) and 2 August 2028 (Annex I), after a postponement in 2026. Not every agent is such a system. I therefore treat the provision as a pattern of reasoning and derive no obligation of my own from it. Oversight there is assigned to natural persons. No procedure and no second agent takes that role over.

The ceiling, concentration risk, and the falsification condition

Start with the distribution my claim does not describe. The US Census Bureau reported in May 2025 that in reference year 2022 there were 29.8 million businesses without paid employees in the United States, with combined receipts of 1.7 trillion dollars. That is an average of about 58 thousand dollars per firm per year, calculated from the unrounded data. In the Census Bureau's combined count of firms with and without employees for the same year there were 35.7 million firms and 51.7 trillion in receipts. Firms without paid employees are therefore about 83 percent of firms and produce about 3.3 percent of revenue.

In 2022 five in six American firms had no paid employees. Together they did one thirtieth of the turnover.

The one-person model scales in count, not in volume. In these tables the Census Bureau gives no median for receipts, but its size classes show that about 61 percent of such firms took in under 25 thousand dollars and about 38 percent under 10 thousand. Some 87 percent are sole proprietorships, and the rest are partnerships and corporations. A firm without paid employees is therefore not always one person. Census adds that the business need not be the owner's main income. The typical member of this group is not a product studio with a fleet of agents. This essay describes the upper tail of the distribution and has to say so outright, or it becomes a winner's story about himself.

The second ceiling comes from regulation. Regulation (EU) 2022/2554, known as DORA, has applied since 17 January 2025. Articles 28–29 oblige a financial entity to assess, before signing a contract for information and communication technology services, whether it reinforces concentration risk. For services supporting critical or important functions, the entity must also assess whether the provider can be easily replaced. Added to that come a register of all contracts concluded and, for such services, exit strategies. The regulation does not exclude a one-person provider. The assessment of substitutability applies to that provider too, though, and my interpretation, which the provision does not contain, goes further: a provider who is one person is an obvious risk factor in that assessment. In parts of the market continuity is the product, so being one person may decide the outcome of a procurement conversation.

Article 14 of the artificial intelligence regulation, which a moment ago handed me the list of things never delegated to an agent, cuts the other way here. It does so if the studio is a provider or deployer of a high-risk system. Duties of oversight (Arts. 14 and 26) and documentation (Art. 11, among others) are then a fixed cost. A fixed cost is regressive with respect to size: a corporation spreads it across a department, while a one-person studio pays it out of the same day it is supposed to make the product in. The more regulation enters my class of task, the narrower the band in which I hold an advantage grows.

The third objection is technical and concerns the fleet itself. Anthropic described in June 2025 how it built a multi-agent research system: an arrangement with a lead model and subagents beat a single agent by 90.2 percent on an internal evaluation. The bill for that is stated openly in the same text. Agents burn around four times the tokens of an ordinary conversation, and multi-agent systems around fifteen times. The authors also point to the class of task where parallelization does damage: anywhere all the agents need the same context, or the dependencies between them are dense, one agent with a long context beats five.

Brooks's arithmetic does not vanish in the move to agents. It comes back as an invoice for tokens.

The difference against an organization is real, but it has to be named precisely. In my case the coordination cost is monetary and settled the same day, because I watch it on the bill for the programming interface. In an organization it is temporal and political, and the accounts never see it. That is a difference of kind and of visibility, and coordinating with agents still costs something.

The hardest objection concerns my own head. A one-person decision loop is fast and has no defence against the error of a single person. In an organization, objection is at times the only quality gate that works, even when the motive behind it is political — bad intent can still catch a hole in the reasoning. Nobody objects at my place, because there is nobody to. I substitute three things for that objection, and I know of each one that it is weaker than a live opponent. A measurement gate returns a number independent of me. An agent launched purely in the role of opponent gets the finished material and is told to break it. An irreversible move I do not make on the day I thought of it. Anyone who claims that set is a full equivalent of a team is lying.

The last objection strikes at the genre of this text itself, and I cannot rebut it. Manifestos are written by the ones it worked out for. Had my studio folded last year, you would not be reading an essay about what dropping the internal layer of coordination buys, but one about why being a company of one is a trap. That second essay would have equally good sources, because the material for both claims lies in the same studies you have just read. Survivorship bias cannot be removed by better writing. It can be named, together with the condition by which you will recognize that the author was wrong.

And so I close the manifesto with a condition that lets me be wrong in a checkable way. The claim falls if, within twelve months, a team inside an enterprise programme delivers — in the same class of task, verifiable and reversible, with no institutional signature required — a result of comparable quality in comparable calendar time. It falls by a second route as well: if my studio loses because the leverage of agents demanded more verification than it saved in execution. That second clause is not decoration. It is the same kind of result METR measured on experienced developers in 2025, although METR now marks it as out of date; I have no grounds for assuming it passes me by.

Why I am writing this

I do not run a one-person studio because I have better models. I have the same ones any corporation with a budget has access to, and probably worse infrastructure around them. I run it because in the narrow band of tasks I can verify and undo, my decision loop lasts as long as my decision lasts, and nobody has to approve it along the way.

Outside that band I lose to resources, and I lose deservedly.

If you run a programme in a large organization, do not take from this essay the conclusion that your people are the problem. They are not. Count instead how many people must agree on your typical decision, and put that number into Brooks's formula as n: you get the number of pairs that can fall out of step. Then check one thing: does the agent you are rolling out give knowledge to the person who decides, or does it only produce material for the next review?

Sources

  1. Frederick P. Brooks Jr., The Mythical Man-Month: Essays on Software Engineering, Addison-Wesley 1975 (anniversary edition 1995) — Brooks's law and the n(n−1)/2 arithmetic of pairwise coordination. An analytical model drawn from observation on OS/360, not a study with a sample.
  2. Ronald H. Coase, The Nature of the Firm, Economica, N.S. 4(16), 1937, pp. 386–405 — transaction costs as the reason firms exist and as the marker of their boundary, together with the claim about the entrepreneur's diminishing efficiency. Nobel Prize 1991, in part for this text.
  3. Nachiappan Nagappan, Brendan Murphy, Victor Basili, The Influence of Organizational Structure on Software Quality: An Empirical Case Study, ICSE '08 (also Microsoft Research TR-2008-11) — a model of eight organizational metrics predicting post-release failures of Windows Vista binaries. A case study of one product at one company; prediction without causal identification, and not a test of Brooks's law about late projects.
  4. Nicholas Bloom, Luis Garicano, Raffaella Sadun, John Van Reenen, The Distinct Effects of Information Technology and Communication Technology on Firm Organization, Management Science 60(12), 2014, pp. 2859–2885 — the separation of information technology from communication technology and the opposing directions of their effect on decision autonomy. Econometrics on World Management Survey data, instrumented identification.
  5. Oriana Bandiera, Andrea Prat, Stephen Hansen, Raffaella Sadun, CEO Behavior and Firm Performance, Journal of Political Economy 128(4), 2020, pp. 1325–1369 — time diaries of 1,114 CEOs across six countries, unsupervised classification, the "leaders" and "managers" types.
  6. METR — two measurement papers together with the authors' own caveats. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (Joel Becker, Nate Rush et al., arXiv:2507.09089, 2025; METR post): a randomized experiment, forecast −24%, self-report −20%, measurement +19% working time; in the note We are Changing our Developer Productivity Experiment Design of February 2026 the authors judged the data from the newer round an unreliable signal because of selection (some developers and tasks dropped out as they did not want to work without AI), and the 2025 post carries a notice that its results are out of date. Measuring AI Ability to Complete Long Software Tasks (arXiv:2503.14499, 2025) together with the January 2026 note on the limitations of the metric: the task horizon at 50% and 80% reliability, and the 98% threshold for tasks that are critical for reliability and poorly verifiable.
  7. Sida Peng, Eirini Kalliamvakou, Peter Cihon, Mert Demirer, The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, arXiv:2302.06590, 2023 — a standard task (an HTTP server in JavaScript, from a template repository with twelve tests) completed 55.8% faster by the Copilot group among those who completed it; 95 freelancers (35 completed the task), the Codex model. The tool maker's co-authorship recorded in the body of the essay.
  8. US Census Bureau, Census Bureau Releases Demographic Characteristics of Nonemployer Business Owners, press release of 8 May 2025, reference year 2022 — 29.8 million businesses without paid employees and 1.7 trillion USD in receipts, and a combined count of 35.7 million firms with 51.7 trillion USD in receipts (Annual Business Survey and Nonemployer Statistics by Demographics). The average, the shares by legal form and by receipts size class are my own calculations on the data of Nonemployer Statistics 2022 (December 2024 release, file nonemp22us); definitions after the FAQ. Administrative data from tax filings, United States coverage; Census notes that the business need not be the owner's main source of income, and a threshold of 1,000 USD in receipts (1 USD in construction) helps define the population.
  9. Regulation (EU) 2022/2554 of the European Parliament and of the Council (DORA), Arts. 28–29 — assessment of ICT concentration risk, the register of contracts and exit strategies; applicable since 17 January 2025.
  10. Regulation (EU) 2024/1689 (the artificial intelligence act), Art. 14 — the requirements for human oversight of a high-risk system; in force since 1 August 2024. Regulation (EU) 2026/1744 (Digital Omnibus on AI) amended the dates of application (Art. 113): for high-risk systems from 2 December 2027 (Annex III) and 2 August 2028 (Annex I), see the consolidated version. Status as of 11 October 2026.
  11. Aditya Challapally, Chris Pease, Ramesh Raskar, Pradyumna Chari (MIT Project NANDA), The GenAI Divide: State of AI in Business 2025, July 2025 (file v0.1, preliminary findings) — a report of a research initiative, not a peer-reviewed paper; interviews, a survey and a review of initiatives, with results largely self-reported; the appendix holds questionnaires, not data. The figures of 95% and 5% as worded on p. 3, and of 90% and 40% as worded in the prose on p. 8; the chart on p. 8 labels the 90% as employees.
  12. Anthropic, How we built our multi-agent research system, engineering post, June 2025 — the 90.2% advantage of the multi-agent arrangement, the fifteenfold cost in tokens, the class of task unsuited to a fleet. The maker's internal evaluation: direction credible, magnitude not reproducible from outside.
  13. Veracode, 2025 GenAI Code Security Report, October 2025 update — 45% of code samples carrying an OWASP Top 10 vulnerability, more than 100 models, 80 tasks. A benchmark from a maker of code scanners: methodology public, commercial interest in a result of that shape recorded.