What AI is, what it changes, and what it resembles

Nobody held a meeting and decided to adopt artificial intelligence. It turned up anyway — in the video call that now writes its own summary, in the inbox that offers to draft the reply. Four questions are worth answering before the next subscription.

This article is the introduction to this site. It answers, in plain terms, what these tools are, what they change in a working week, what is at stake beyond this year, and what the whole thing should be compared to. The last question turns out to be the most useful of the four, and the usual answer to it — the internet — is half right. The half that is right is not the half people usually mean.

What actually arrived

The arrival was unusual, and it explains a lot of the confusion that followed. Most workplace technology is bought: someone decides, someone signs, someone installs it. This one appeared as a chat box on a website at the end of 2022, and then as a button inside software that was already being paid for. No decision was taken. A button was clicked, once, to see what would happen.

Adoption of that kind spreads faster than evidence about it. The economists Anders Humlum and Emilie Vestergaard linked surveys of around 25,000 Danish workers per round, across 11 occupations exposed to chatbots — accountants, legal professionals, marketers, software developers and others — to the country’s employment records. Adoption was widespread and workers reported real benefits. Two years after ChatGPT’s launch, no effect could be detected on their earnings or their recorded working hours. Most users, 85% of them, said the time saved had gone into other tasks.

That is the shape of the thing to keep in mind: real use, real reported benefit, and no measurable trace yet at the level of a paycheque or a working week. Both halves of that sentence are true at once, and most writing about AI keeps only the half it prefers.

1. What is AI, as far as your desk is concerned?

The textbook definition matters less than what the tools actually do. Almost every AI product sold to professionals today is built on a large language model: software trained on enormous amounts of text to predict what comes next in a piece of writing. At that scale, prediction becomes fluent enough to draft an email, summarise a contract, or answer a question in ordinary language.

A meeting tool usually chains two of these systems. Speech recognition turns the recording into words; a language model turns the words into a summary. Both make mistakes, and the two kinds of mistake behave differently. A transcription error usually looks like an error — a mangled name, a word that makes no sense in the sentence. The language model’s mistakes do not. A wrong answer is written in exactly the same confident tone as a right one. A decision that was never taken looks, on the page, identical to a decision that was.

The industry calls this a hallucination. Invention is the more honest word, and it is the reason inventions are counted separately from omissions in every test published here: a missing decision gets noticed by whoever was in the room, while an invented one gets acted on.

Two more properties are worth knowing before the first subscription. These tools know nothing about your business beyond what is put in front of them in the moment, which is why the quality of what comes out depends so heavily on what goes in. And whatever is put in front of them usually leaves your machine: it is sent to a server belonging to someone else, kept for some length of time, and governed by terms that change. That is a separate subject, with its own article.

2. What does it change, at work and at home?

At work, the honest answer is that the change is mostly a transfer: from producing a first draft to checking one. Summaries, transcripts, reformatted documents and routine replies now take seconds to produce and still take minutes to verify. Whether any time is saved at all depends on that second half — which is precisely the half left out when the saving is estimated.

The best evidence available on that point is uncomfortable. In 2025 the research group METR ran a randomised trial with 16 experienced open-source developers working on 246 real tasks from their own projects, each task randomly assigned to be done with or without AI tools. Before starting, the developers expected AI to make them 24% faster. Measured, they took 19% longer. Asked afterwards, they still estimated that AI had made them 20% faster.

That study does not prove AI slows people down in general, and its authors say so. A follow-up published in February 2026, with newer tools, points towards a speedup instead — with a margin of error that still includes zero, and a warning that the developers who relied most on AI had become reluctant to take part in a study that required working without it. What the first study established is narrower and more useful: in that setting, the people doing the work could not tell, from the inside, whether they had been faster or slower.

One profession is already living the whole of this: translation. The machine drafts, the professional checks, and the practice has a name — post-editing. Checking is paid less than producing, even when it takes the same judgement, because from the outside it looks faster.

At home, the same tools draft the letter to the landlord, plan the trip, and explain the tax form. The rule is the same in both places: they are safe where mistakes are cheap and easy to spot, and risky where mistakes are expensive and invisible. The difference is that at home a mistake costs an evening, and at work it can cost a client.

One habit carries across both. The same account often gets used for the holiday itinerary and for the client file, and that is where confidentiality problems start.

3. Why is it a question for the future, not only for now?

Predictions about which jobs disappear are cheap to make and impossible to check, so none are offered here. Three consequences, though, are already concrete for anyone working for themselves.

Clients will expect the speed the tools advertise. Whether or not the gain is real, the expectation is arriving, and it shows up in what people are willing to pay for a draft. Knowing the actual gain on your own tasks stops being a curiosity and becomes a commercial fact.

Confidentiality obligations have not moved; the ease of breaking them has. A client’s file is now one paste away from a third party’s server. The rules that govern lawyers, accountants and therapists were written before that was possible, and they still apply.

The ground moves every few months. Tools change, plans change, terms change. Any result — including every result published on this site — is a statement about a date, and is worth exactly as much as the date attached to it.

And one question worth watching rather than answering: the work these tools do best is the work junior people used to learn on. What replaces that apprenticeship is not yet clear, and anyone claiming otherwise is guessing.

4. What should it be compared to?

The internet is the comparison everyone reaches for. It is a good one in three places and a poor one in a fourth, and knowing which is which is what makes it useful.

The resemblance, point by point

The enthusiasm. Then: a technology that arrived everywhere at once, a promise that everything was about to change, and the feeling that staying out was the risky choice. Now: the same tone, the same everywhere-at-once, the same feeling. Anyone who worked through the late 1990s recognises it, including the parts that are embarrassing in hindsight.

The adoption curve. Then: 18% of American households had internet access in 1997, 41.5% by 2000, half by 2003. Now, in figures that count firms and workers rather than households: between 17% and 20% of American businesses reported using AI between December 2025 and May 2026 — 37% among those with 250 employees or more, under 20% among the smallest — while roughly 41% of workers said in November 2025 that they use generative AI for work. Different unit, same shape: a curve climbing fast while most of the market is still watching from the side.

The money. Then: 476 companies went public in the United States in 1999 and gained 71.2% on average on their first day of trading, 86.7% for the 370 technology offerings among them; the Nasdaq Composite then lost roughly three-quarters of its value between March 2000 and the autumn of 2002, and the web kept working the whole way down. Now: in the first half of 2026, $355.9 billion of the $412.7 billion invested by American venture capital went to AI companies — 86 cents in every dollar. In both cases the money is a statement about expectations, not a measurement of what the technology does on a Tuesday morning, and the two get confused constantly.

The promise of a new way of working. Then: kept. The internet changed the working week permanently — the client on another continent, the file sent instead of couriered, the practice run from a spare room, whole categories of independent work that exist only because of it. Now: the same size of promise, and so far a change that sits inside the task rather than around it. In the one profession that has already been through it end to end, translation, what changed was the order of operations — the machine drafts, the professional checks — and the rate paid for the checking version, about 20% lower.

The calendar. Then: American productivity growth ran at about 1.5% a year from the early 1970s to 1995, more than doubled between 1995 and 2003, and fell back to roughly its old pace after the mid-2000s. Now: nothing detectable yet, as question 2 set out — real reported benefits, no measurable trace in earnings or hours. If AI follows the same calendar, the numbers arrive years after the enthusiasm, and they arrive for the people who reorganised the work rather than for everyone who subscribed.

Where the resemblance stops

Technically, the two are opposites. The internet is infrastructure: standards and cables whose value grows with every person who joins, which is why building it took a decade and enormous amounts of capital, decided far above the level of any one professional. An AI tool works the other way round. It is bought one seat at a time, for a few tens of euros a month, and it acts on a task that is already yours. Nothing about it improves because a colleague subscribes too.

And the output behaves differently. A web page either loads or it does not; the failure is visible. An AI answer arrives complete, fluent and confident whether it is right or wrong, which creates a cost the internet never created: someone has to check. That cost lands on the person who used the tool, it is invisible from the outside, and it is the single most under-estimated number in the whole subject.

For work, the difference is where the change lands. The internet changed who you can work with and from where — it rearranged the space around the task. An AI tool changes which part of the task you still do yourself. That is a change inside the work, felt one job at a time rather than one industry at a time.

Which is why, for deciding whether a particular subscription is worth it on a particular Tuesday, a smaller and older comparison is more concrete.

The concrete one: the spreadsheet

VisiCalc shipped for the Apple II in the autumn of 1979. Computer stores sold it for $200, and it was bought one copy at a time by professionals who wanted to run their own numbers, often without asking the company’s computer department — which is exactly how AI tools are bought today.

It automated one precise, expensive task: financial modelling done by hand, on paper, with an eraser. Not "office work" in general — one task, done slowly by people who were expensive. And it did not eliminate the profession, it eliminated a category of manual work inside it. The shift is still visible in the official projections: the US Bureau of Labor Statistics expects employment of bookkeeping, accounting and auditing clerks to decline 6% between 2025 and 2035, while accountants and auditors grow 5% over the same period. The clerical category shrinks; the judgement category does not.

The mechanism: the electric motor

The third comparison is less flattering, and it explains why a subscription on its own changes nothing.

From the 1890s, factories bought electric motors, and for roughly three decades the productivity figures barely moved. The reason, as the economist Paul David set out in 1990, is that the factories kept the building they had: a layout designed around a single steam engine driving one central shaft, with machines placed by how close they had to sit to that shaft. An electric motor bolted into that layout is a quieter steam engine. The gains arrived in the 1920s, when the floor was finally rearranged around machines that each had their own motor — placed by the logic of the work rather than the logic of the shaft.

Translated to a desk: buying the tool is the cheap part of the decision. Changing how the week is organised around it is the part that does something, and the part nobody sells.

The caution that belongs with all three

AI has been announced as transformative before, and twice the money left the room. Sir James Lighthill’s survey for the UK Science Research Council, written in July 1972 and published the following year, recorded that "most workers in AI research and in related fields confess to a pronounced feeling of disappointment in what has been achieved in the past twenty-five years"; British funding was dismantled shortly afterwards. A second collapse followed in the late 1980s, when the market for specialised AI hardware disappeared and the companies built on it went with it.

Those episodes do not prove that this time is the same. They are a reason to write down, in advance, what would have to be observed for the sceptics to be right.

What would show the sceptics right

Two years is a reasonable horizon. Four things would be worth treating as evidence.

Measured gains stay inside the noise. Randomised, task-level studies keep producing effects whose confidence intervals include zero, while the reported gains stay large.
Renewals fall quietly. Seats bought in the first enthusiasm are not renewed, and vendors stop publishing user counts.
The checking burden does not fall. The share of output that has to be verified before it can be sent stays where it is, tool generation after tool generation.
Prices rise without capability moving. Plans get more expensive, limits tighten, and the same tasks come back the same quality.

The opposite pattern — narrowing error rates on tasks with a known answer, verification getting cheaper, gains showing up in hours and earnings — would say the comparison to the spreadsheet was the right one.

Either way, the way to find out is to measure, on tasks that matter, against an answer written down in advance.

What this site does about it

Kalytia tests AI tools on real work and publishes the numbers. One realistic task, the same input for every tool, a scoring grid fixed before the first test, and every figure traceable to a dated run. The first category under test is meeting tools: the task is a written meeting with a reference answer prepared by hand, so that what a correct summary should contain is known before any tool hears the recording.

Three articles carry the rest of the introduction:

→ The method, in full: How to tell whether an AI tool actually saves you time

→ For professions bound by confidentiality: What never to paste into an AI tool

→ What the vendors’ own terms say: What happens to what you type

Sources — consulted 12 September 2026

Anders Humlum and Emilie Vestergaard, "Large Language Models, Small Labor Market Effects", NBER Working Paper 33777, 2025. nber.org/papers/w33777
METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", 10 July 2025. metr.org
METR, "We are Changing our Developer Productivity Experiment Design", 24 February 2026. metr.org
European Language Industry Survey 2026, published March 2026. elis-survey.org
US Census Bureau, "Large Firms With at Least 20 Employees Biggest AI Users", 26 May 2026. census.gov
Jeffrey S. Allen, "Monitoring AI Adoption in the U.S. Economy", FEDS Notes, Federal Reserve Board, 3 April 2026. federalreserve.gov
PitchBook-NVCA Venture Monitor, Q2 2026, published July 2026. nvca.org
US Census Bureau, "Computer and Internet Use in the United States: 2003", P23-208. census.gov
Jay R. Ritter, "Initial Public Offerings: Mean First-day Returns and Money Left on the Table, 1980-2025", University of Florida, version of 16 March 2026. warrington.ufl.edu
Nasdaq Composite index, daily series, Nasdaq Inc. via FRED. fred.stlouisfed.org
John Fernald and Bing Wang, "The Recent Rise and Fall of Rapid Productivity Growth", Federal Reserve Bank of San Francisco Economic Letter, 9 February 2015. frbsf.org
Dan Fylstra, "Personal Account: The Creation and Destruction of VisiCalc", Computer History Museum, 2004. computerhistory.org
US Bureau of Labor Statistics, Occupational Outlook Handbook, "Bookkeeping, Accounting, and Auditing Clerks" and "Accountants and Auditors", page last modified 27 August 2026. bls.gov
Paul A. David, "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox", American Economic Review, 80(2), 1990, pp. 355–361.
Sir James Lighthill, "Artificial Intelligence: A General Survey", Science Research Council, July 1972 (published 1973). chilton-computing.org.uk
Daniel Crevier, AI: The Tumultuous History of the Search for Artificial Intelligence, Basic Books, 1993 — for the late-1980s collapse.