January · Artificial Intelligence · 67 pages
What AI actually does for a business this size — and what it demonstrably does not.
The Machine Age · Technology
Thirteen years of forecasts, measured against the record. The jobs machines actually took, the ones they were said to be taking, and the number everybody quoted without reading.
In 2013, two researchers at Oxford put a number on the future of work, and the number went everywhere. About 47 percent of total US employment, they wrote, was at risk of computerisation. The figure was reprinted in newspapers, board slides and election speeches for a decade. Thirteen years on, the arithmetic is checkable. The Bureau of Labor Statistics counted 145.4 million jobs in the 2012 base year of its own projections. It counted 170.3 million in 2025, and expects 176.2 million by 2035. Employment did not fall by anything resembling half. That is not in itself a reason to dismiss the forecast. It is a reason to read it properly, because what the paper said and what the public heard were two different things.
The paper estimated a probability of computerisation for 702 occupations using a machine-learning classifier. To train it, the authors and a group of machine-learning researchers hand-labelled 70 occupations — assigning 1 if automatable and 0 if not — based, in their own words, on eyeballing the O*NET tasks and job description of each occupation. Occupations scoring above 0.7 went into the high-risk bucket. The sentence that produced the famous figure reads: 47 percent of total US employment is in the high risk category, meaning that associated occupations are potentially automatable over some unspecified number of years, perhaps a decade or two. The qualifier travelled considerably less well than the number.
Potentially automatable over some unspecified number of years, perhaps a decade or two.Carl Benedikt Frey and Michael Osborne, The Future of Employment, 2013
Three years later a second team asked the same question differently. Working for the OECD, Melanie Arntz, Terry Gregory and Ulrich Zierahn argued that treating whole occupations as automatable overstates the case, because occupations labelled high-risk still contain a substantial share of tasks that are hard to automate. Using worker-level data on what people in those jobs actually do, they put the share of automatable jobs across 21 OECD countries at 9 percent. Same question, same era, a fivefold difference — and the difference is entirely methodological. Anyone quoting either figure is quoting a modelling choice, not a measurement.
Against those projections sits the duller record of what machines have in fact taken. Bank tellers are the textbook case, and the textbook is wrong about them. Cash machines arrived in American branches in the late 1960s. In 2025 the United States still employed 339,200 tellers. BLS expects that to fall 13 percent by 2035, and attributes the decline mostly to customers banking on their phones rather than to the machines in the lobby. Office and administrative support is the occupational group projected to shrink fastest over that decade: down 4.0 percent, a loss of 752,100 jobs. Fast, in this literature, means four percent in ten years.
339,200 Bank tellers employed in the United States in 2025, roughly sixty years after the cash machine was supposed to end the job — BLS Occupational Outlook Handbook
The pattern in both cases is substitution inside the job rather than deletion of it. Tellers stopped counting cash and started selling products. Administrative staff stopped typing and started running systems. James Bessen's work on textiles, steel and automobiles makes the mechanism explicit: whether automation cuts employment or expands it depends on what happens to demand once the product gets cheaper. Where demand had room to grow, mechanisation added workers for a century. Where demand saturated, the same technology shed them. The technology does not decide the outcome. The market the technology serves does.
There is also a blunt aggregate test, and it has not yet been passed. If artificial intelligence were transforming American production, it would show up as faster output per hour. BLS reports nonfarm business productivity growing at a 2.1 percent annualised rate over the current business cycle, from the end of 2019 through the second quarter of 2026. That is better than the 1.5 percent of the previous cycle. It is also exactly the long-run rate since 1947. Whatever is happening inside individual firms, the national statistics currently show an economy improving its productivity at its historical average — no more.
Economists disagree about what to expect from here. Daron Acemoglu's task-based estimate puts the plausible gain at no more than 0.66 percent of total factor productivity over ten years, and he revises that down to less than 0.53 percent once harder-to-learn tasks are accounted for. Others expect considerably more. The point for a business owner is not which camp is right, but that the dispute is live among people with full access to the data. A vendor presenting a settled figure for economy-wide gains is presenting something the economics profession does not currently possess.
Some of the loudest promises, meanwhile, have delivered narrowly rather than not at all. Autonomous driving was the emblem of imminent displacement. Stanford's AI Index records Waymo providing more than 150,000 paid rides a week in the United States — a genuine commercial service, in a handful of metropolitan areas, more than fifteen years after the first public demonstrations. In medicine, the same report counts 223 AI-enabled devices cleared by the Food and Drug Administration in 2023, against six in 2015. Narrow tools cleared for specific purposes have multiplied. The general replacement of the professional has not arrived.
The honest summary is that nobody knows yet, and the statistical agencies say so in writing. In a 2025 article on how it handles artificial intelligence, BLS states that its projection methods are not designed to capture extremely rapid technological change and therefore assume the overall pace of technological change will be consistent with past experience. It adds that projecting future employment involves substantial uncertainty, especially in the case of evaluating the future impacts of a developing technology. That is the government's own caveat on its own numbers, and it is more candid than most of what a buyer will be shown this year.
What the evidence supports is narrower and more useful than either the headline or the backlash. Machines take tasks, not careers, and they take them unevenly and slowly enough that a business usually has time to respond. The forecasts that failed did not fail because their authors were foolish; they failed because turning a judgement about tasks into a percentage of employment requires assumptions that nobody can check in advance. A reader deciding whether to spend money this year is better served by that pattern than by any percentage.
The Machine Age · Technology
Two federal surveys, a Danish payroll study and four randomised trials now describe what this technology does inside a working business. The answer is reallocation, not reduction — and it arrives first in the tasks.
The question a business owner actually asks is narrower than the one the forecasts answer. Not whether some fraction of American jobs is at risk, but whether this particular company will need fewer people next year. For the first time there is a government answer drawn from firms rather than from occupational models. In April 2026 the Census Bureau published a study of its own artificial intelligence supplement, covering November 2025 through January 2026, matching firm-level adoption against what firms reported happening to their payrolls. The findings are unglamorous, which is precisely what makes them worth reading.
Eighteen percent of US firms reported using AI in a business function. Weighted by employment the figure rises to 32 percent, because large firms adopt first and employ more people. Among adopters, 57 percent confined it to three or fewer functions — most often sales and marketing (52 percent), strategy and business development (45 percent) and IT (41 percent). On the question that matters most, 66 percent of users reported deploying the technology to augment tasks rather than to substitute for workers, while employment decreases were reported by 2 percent of firms. Those reductions clustered among large firms that had adopted explicitly in order to substitute.
2% Share of US firms reporting employment decreases from AI adoption, in a Census Bureau study of its November 2025 to January 2026 AI supplement — Census Bureau Center for Economic Studies, CES-26-25
An earlier federal instrument reached the same place by a different route. The Annual Business Survey asked businesses about technologies adopted between 2020 and 2022. For artificial intelligence, the share reporting an increase in worker numbers was 9.5 percent and the share reporting a decrease was 8.1 percent — a difference the Bureau noted was not statistically significant. The same held for robotics. Most businesses reported little or no impact on worker skill level either. Two separate federal surveys, different years and different questions, both landing on roughly nothing in aggregate.
The cleanest evidence comes from Denmark, where researchers can link survey responses to administrative payroll records rather than relying on recall. Anders Humlum and Emilie Vestergaard found that most employers in exposed occupations had adopted chatbot initiatives, and that workers reported productivity gains — and then found precise null effects on earnings and recorded hours at both the worker and workplace levels, ruling out effects larger than 2 percent two years after ChatGPT's launch. Adoption was real. The self-reported gains were real. Pay and hours did not move.
What moved instead was the shape of the work. The Danish authors describe reorganisation: workers shifting into roles where the tools applied, firms rearranging who did what, none of it yet large enough to register in aggregate earnings. Their conclusion is the one most forecasts miss — that technological change reshapes work well before it surfaces in earnings or hours. For an owner, that is the practical warning. The first effect of adoption is not a smaller payroll. It is a different allocation of the payroll already there, and it tends to be invisible on a profit and loss statement.
Technological change reshapes work well before it surfaces in earnings or hours.Humlum and Vestergaard, Still Waters, Rapid Currents, NBER, 2025
The best-documented example of that reallocation comes from a study of 5,179 customer support agents given staggered access to a generative AI assistant. Published in the Quarterly Journal of Economics, it found productivity — measured as issues resolved per hour — rising 14 percent on average, but 34 percent for novice and low-skilled workers, with minimal impact on experienced and highly skilled ones. The authors read the tool as distributing the practices of the best agents to the newest. It also improved customer sentiment and increased employee retention. The headcount was not the variable that changed.
A randomised experiment with 444 college-educated professionals on mid-level professional writing tasks found the same compression: time taken fell by 0.8 standard deviations and output quality rose by 0.4, with the largest gains going to the weakest writers. That is a specific, replicated-looking finding about a specific class of task, and it says something useful about hiring and training. It says nothing at all about whether a firm needs fewer writers, which is a question about demand for writing rather than about the speed of producing it.
The counter-evidence is just as instructive. In 2025 the research group METR ran a randomised trial with 16 experienced open-source developers across 246 real issues in repositories they themselves maintained. With AI tools available, they took 19 percent longer. They had forecast a 24 percent speed-up beforehand, and after the fact still believed the tools had made them roughly 20 percent faster. The gap between measured and felt productivity may be the most expensive thing in this field, because enthusiasm inside a company is almost always assessed by asking the people using the tool how it is going.
BLS occupational detail fits the same picture of slow reshaping. In its case studies of AI-exposed occupations, the agency projects claims adjusters, examiners and investigators down 4.4 percent over a decade; credit analysts down 3.9 percent; paralegals and legal assistants down 1.2 percent; and automobile damage appraisers, the most exposed on its list, down 9.2 percent. Over the same period it projects software developers up 17.9 percent and personal financial advisers up 17.1 percent. These are not disappearances. They are gentle slopes, running in both directions at once.
Taken together the evidence has a consistent shape: these systems raise the floor rather more than they raise the ceiling. They help the inexperienced disproportionately. They help on well-bounded tasks far more than on messy ones. And they are most likely to mislead a buyer precisely when the people using them are the ones reporting on how well they work. None of that resolves into a percentage of jobs. It resolves into a question about which tasks in a particular building are well defined enough for the tool to touch.
The Machine Age · Technology
Adoption figures that differ by sixty points, a famous failure rate assembled from 52 interviews, and two randomised trials that disagree by seventy-five points. A short course in taking a statistic apart before it takes your budget.
Every buyer of this technology is handed two numbers before being handed a price. The first says the market is enormous. The second says everyone else is already using it. Both are usually true in some sense and useless in every sense that matters, because they are assembled from populations and questions the buyer never sees. Learning to take a statistic apart has become part of the purchase. It is considerably cheaper than the alternative, which is paying for a story and finding out eighteen months later which parts of it were load-bearing.
Start with adoption, where two credible-sounding figures differ by sixty points. Stanford's AI Index reports that 78 percent of organizations said they used AI in 2024, up from 55 percent the year before. The Census Bureau's AI supplement, covering November 2025 to January 2026, puts the share of US firms using AI in a business function at 18 percent. Neither is wrong. They describe different universes: organizations that chose to answer an industry survey, against a sample of US employer businesses asked a specific question about the previous two weeks.
The second figure's provenance is worth knowing, because it is unusually good. The Business Trends and Outlook Survey draws on roughly 1.2 million businesses, split into six panels of about 200,000, each reporting once every twelve weeks, with collection every two weeks. It is a probability sample of US employer businesses, which is why it can say that fewer than 20 percent of firms with four or fewer employees use AI while 37 percent of firms with 250 or more do, and that use runs at 39.7 percent in the information sector against roughly 14 percent in retail trade. No vendor survey can make those cuts, because no vendor survey knows who it failed to reach.
It is also a lesson in wording. In its October 2023 wave, the survey found 3.8 percent of businesses using artificial intelligence — but the question asked specifically about using AI to produce goods and services. In November 2025 the Bureau revised it to ask whether the business had used AI in any business function. Measured the new way, use runs between 17 and 20 percent. A meaningful part of the apparent surge in American AI adoption is a change in what the question counts. The Bureau says so plainly. Coverage of the jump generally did not.
Now the famous pessimistic number. The claim that 95 percent of corporate AI pilots fail traces to a document titled The GenAI Divide: State of AI in Business 2025, published in July 2025 by MIT's Project NANDA. Its own front matter describes it as preliminary findings, and its methodology as a review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected across four major industry conferences. It is not peer reviewed and it is not a failure rate. It is a readable essay with a memorable figure attached to it.
The regulators have been blunter than the researchers about error in the other direction. In March 2024 the Securities and Exchange Commission fined two investment advisers a combined $400,000 for what its chair called AI washing — Delphia for claiming to put collective data to work to make our artificial intelligence smarter when the Commission found it had no such capability, and Global Predictions for advertising itself as the first regulated AI financial advisor. In September 2024 the Federal Trade Commission brought five cases under the banner Operation AI Comply, including a $193,000 settlement with DoNotPay over its robot lawyer.
Investment advisers should not mislead the public by saying they are using an AI model when they are not.Gary Gensler, then chair of the Securities and Exchange Commission, March 2024
This is the reason to treat vendor-produced evidence as marketing until shown otherwise. It is not an accusation that vendors lie. It is that there exists no mechanism by which a seller's internal study can be checked by the buyer — no sample frame, no preregistration, no independent replication, no published code. Where the only number available comes from the company selling the thing, the honest reading is that no independent number exists yet. That is a fact about the state of the evidence rather than about the vendor, and it belongs in the decision either way.
Good evidence does exist, and it looks different. Two randomised trials on software development illustrate the whole problem. One asked developers to implement an HTTP server in JavaScript as quickly as possible and found the group with an AI assistant finished 55.8 percent faster. Another, run by METR in 2025, gave 16 experienced developers AI tools on 246 real issues in repositories they maintained, and found them 19 percent slower. Both were randomised. The difference is that one measured a scripted exercise with a known answer and the other measured work.
On the question buyers most want settled — does this raise a firm's productivity — the truthful answer is that nobody knows yet. BLS has nonfarm business productivity growing at 2.1 percent annualised since late 2019, the same as its rate since 1947. Daron Acemoglu puts the macroeconomic gain at under 0.66 percent of total factor productivity over a decade, and revises it lower still. Others put it far higher. Credentialed economists with full access to the data disagree sharply. A vendor who does not is selling a certainty that is not currently for sale.
Which leaves the only measurement genuinely available to a small or mid-sized firm: its own. Choose one task. Record how long it takes and how often it has to be redone, for two weeks, before anything is installed. Then run it again afterwards. The METR result exists only because somebody insisted on measuring rather than asking, and found the opposite of what the participants sincerely believed. That experiment costs a business almost nothing, and it is the only study whose population is guaranteed to resemble the buyer's own.
The Machine Age · Technology
Courts have now recorded 2,262 cases of someone relying on fabricated AI output, and the sanctions fell on the humans. A procurement checklist for a product category that is still writing its own rules.
A contract is where a sales narrative becomes an obligation, and it is the last point at which a buyer holds any leverage. The questions worth asking are not technical ones. They are the ordinary questions of commercial life — who is responsible when this is wrong, what happens to the data, what happens when the product changes, what happens when the buyer wants to leave — put to a product category whose vendors are frequently younger than the contracts they are asking to sign. None of it requires an engineer in the room.
Begin with responsibility for output, because the record is now large enough to count. A public database maintained by the researcher Damien Charlotin tracks legal decisions in which a court or tribunal explicitly found that a party relied on hallucinated content or material. As of 2 November 2026 it lists 2,262 such cases, 1,533 of them in the United States, with the record beginning only in the second quarter of 2023. Canada accounts for 240, Australia 112, the United Kingdom 89. The curve is steep, and none of these were close calls — they are findings of fact.
2,262 Court and tribunal decisions recording reliance on AI-fabricated material, worldwide, as of 2 November 2026 — 1,533 of them in the United States — AI Hallucination Cases database, Damien Charlotin
What makes that collection instructive is who paid. These are overwhelmingly findings against the professional who filed the document, not against the software that produced it. The system generated plausible text; a person put their name on it and sent it to a court. Every indemnity clause a buyer is offered should be read against that pattern, and so should every workflow built on the tool. If the output reaches a client, a regulator or a court, somebody in the building has to be accountable for checking it — and that person needs the time to actually do so.
The system produced the error. The record shows the sanction landing on whoever signed underneath it.On the 1,533 US cases in the AI Hallucination Cases database
Some of this is already law, and it binds the buyer rather than the seller. New York City's Local Law 144 of 2021 has been enforced since 5 July 2023. Any employer using an automated employment decision tool in the city must ensure the tool has been subject to a bias audit within one year of use, must make information about that audit publicly available, and must give candidates notice. The obligation sits with the employer. A vendor that cannot supply audit documentation has not sold the buyer a product; it has sold them a compliance problem with a login.
More is arriving. Colorado's Senate Bill 24-205, signed in May 2024, divides duties between developers and deployers of high-risk AI systems, with implementation set for 1 February 2026. Deployers must maintain a risk-management policy and programme, conduct impact assessments, review deployments annually, notify consumers when an AI system makes a consequential decision about them, and offer an opportunity to correct the underlying data and appeal to human review where feasible. Discovery of algorithmic discrimination must be reported to the state attorney general within 90 days.
Firms with European customers inherit a further timetable. The EU AI Act entered into force on 1 August 2024. Its prohibitions applied from 2 February 2025, obligations on general-purpose AI models from 2 August 2025, and requirements for the high-risk categories listed in its Annex III from 2 August 2027. None of this is cause for an American business to panic. It is cause to ask a vendor, in writing, which of these regimes it intends to comply with and by what date — and then to notice how specific, or how vague, the answer turns out to be.
For structuring the questions, the most useful document available happens to be free. The National Institute of Standards and Technology released its AI Risk Management Framework, version 1.0, on 26 January 2023. It is explicitly intended for voluntary use and organised around four functions: govern, map, measure and manage. Its value in a negotiation is not compliance but vocabulary. It gives a buyer a defensible structure for asking what the system is for, what could go wrong, how that would be detected and who acts when it is. A serious vendor will recognise the questions.
The commercial terms are where the technical novelty bites hardest. The model behind the product will be replaced, probably more than once, inside a normal contract term, and a replacement can change outputs that the business has come to depend on. Conventional software does not behave this way, so conventional contracts do not address it. The agreement should say what notice precedes a model change, whether the prior version stays available, and what the remedy is if measured quality falls after one. Very few standard forms say anything at all.
Pricing deserves equal scepticism. Consumption-based pricing shifts the cost of inefficiency onto the customer, and first-year volumes are rarely predictable even in good faith. A buyer should insist on a cap, a trial period priced separately from any multi-year commitment, and a defined exit that returns data in a usable format on a stated timetable. The Federal Trade Commission has brought cases, and the Securities and Exchange Commission has fined advisers, over claims made in precisely the register in which this software is sold. That record is free leverage.
None of this demands technical expertise. It demands treating the purchase as a procurement decision rather than a technology decision — the same instinct a reader would already bring to a payroll provider, a commercial lease or an insurance renewal. The questions are old ones, and the person asking them has usually asked them before in another context. The only genuinely new thing is how many sellers in this particular market are quietly hoping they will not be asked at all.