The Speed of AI Adoption

Make a prediction before you read on. Take three technologies that changed office work: the personal computer, the internet and generative AI. Date each from the launch of its first mass-market product: the IBM PC in August 1981, the opening of the internet to commercial traffic in April 1995, and ChatGPT in November 2022. A couple of years after each launch, what share of American adults was using it?

In August and November 2024 three economists, Alexander Bick, Adam Blandin and David Deming, put that question to nationally representative samples of Americans aged 18 to 64. They compared the answers with long-run data on computer and internet use that had been gathered with very similar questions. Three years after the IBM PC, about one American in five used a computer. Two years after the internet opened, about one in five used it. Two years after ChatGPT, 39 percent used generative AI¹.

Whatever you guessed, the number matters less as a trophy than as a clue. It shows where the speed came from, and that in turn shows who now sets the pace of AI inside an organization. It is not the program office.

Three clocks

When a new AI capability arrives, three clocks start running, and they no longer run at the same speed.

The technology clock marks when a capability appears. It moves with every release from every vendor, and no customer controls it. Why it has moved so fast in recent years is the subject of Why AI, Why Now?.

The employee clock marks when people start using the capability on real work. The opening comparison is a reading of this clock, and it is running faster than it did for the PC or the internet.

The enterprise clock marks when the organization can see that use, set bounds on it, measure it and redesign work around what succeeds. It moves only as fast as the organization learns.

For most of the history of corporate technology, the second and third clocks were tied together. Nobody could use a mainframe, an enterprise resource planning system or the corporate network until the organization had bought it, installed it and trained people on it. The enterprise clock ran first and the employee clock waited. Generative AI broke that link. The employee clock now runs ahead, and the distance between the two is where leadership either turns speed into learning or lets it turn into noise.

Where the speed came from

The opening comparison is unusually trustworthy because it holds the measure constant: closely matched survey questions, and time counted from the first mass-market product of each technology. It still rests on choices. Had the authors dated the PC from the Apple II and its 1977 rivals instead of the IBM PC, the computer would look slower still¹. Treat the result as one careful measurement in one country, not as a ranking of every invention in history.

Its most useful finding appears when the authors split use at work from use elsewhere. They can do this for the PC as well as for generative AI, though not for the internet.

At work, generative AI has moved at almost exactly the pace of the PC: 27 percent of workers used it for their job two years in, against 25 percent for the PC three years in. Outside work the two are not close, at 34 percent against 5 percent¹. The two rows measure different groups: the work figures are shares of employed people, and the outside-work figures are shares of all adults aged 18 to 64. The extra speed came from people trying the technology on their own time, and then carrying the habit into the office.

Two cautions keep the number honest. First, adoption curves bend. PC use climbed from 20 percent in year three to 70 percent only by year twenty-two, and internet use rose quickly to 60 percent by year seven before creeping toward 90 percent over the next two decades¹. Nothing in this evidence promises that the curve stays steep. Second, use is not intensity. The same researchers estimate that only between 1 and 5 percent of all US work hours were assisted by generative AI at the time¹. Broad but light use is the pattern The AI Adoption Curve described, where spread runs well ahead of depth.

A first try costs almost nothing

Why did people take up generative AI on their own so much faster than they took up the PC? The authors’ answer is the cost of adoption¹.

To try a personal computer in the early 1980s, you bought expensive hardware that did not travel. To try the internet in the mid-1990s, you bought a modem and signed a contract with a service provider. To try generative AI, you opened a tab on a device you already owned. Many tools were free or cheap, and none needed technical expertise¹. The interface is ordinary language, so almost anyone who writes can begin on the first afternoon. Why that interface changes so much is the subject of Generative AI Changes the Game.

When a first try costs almost nothing, three things follow. More people try. They try more often and on more kinds of work, because a failed attempt costs a minute rather than a budget. And they need nobody’s permission, because nothing has to be bought, installed or approved. The usual delays of buying, installing and training have not merely shrunk. For a first try, they have disappeared.

The authors add a caution that executives should keep in view: similar adoption rates do not imply similar benefits. A technology that is cheap to try will be tried even where the payoff is small¹. Fast adoption is evidence of easy access. It is not yet evidence of value, the distinction AI Is Changing Everything drew between using AI and profiting from it.

The employee clock runs ahead

The same paper offers a rough reading of the enterprise clock. In February 2024 the US Census Bureau’s survey of businesses found that 5.4 percent of firms reported using AI, up from 3.7 percent two months earlier. The authors call that a rapid rise but still far below their estimates of individual use¹.

The two numbers count different things, workers in one case and firms in the other, so the distance between them is not a precise measurement. The direction is what matters. The authors suggest that low adoption costs and the consumer focus of the products explain why individual adoption has outpaced official adoption by firms¹. Firm adoption has risen a great deal since then, as The AI Adoption Curve showed, but the order has held: people first, organizations after.

The on-ramp has reversed

Traditional enterprise technology runs from the top down. Leaders approve, IT implements, then staff adopt. Generative AI often runs the other way. Someone finds a use, the team copies it, and leaders respond afterwards.

The pattern is not entirely new. Bick and his co-authors quote a technology chief executive who calls it “almost a one-to-one parallel” with the consumerization of IT in the early 2010s, when employees frustrated with corporate document tools simply signed up for consumer cloud services¹. What is new is the reach. A consumer cloud folder held files. A consumer AI assistant can draft, summarize and analyze almost anything an employee pastes into it.

Call the result bottom-up adoption pressure. The organization is no longer the only on-ramp, so the leadership question changes. It is no longer only how to introduce AI. It is how to see, bound and learn from what people are already doing. A large international survey shows what that looks like in practice. Between November 2024 and January 2025, the University of Melbourne and KPMG surveyed more than 48,000 people in 47 countries².

Among employees who use AI at work, about 70 percent rely on free public tools, compared with 42 percent who use tools their employer provides. Nearly half have uploaded sensitive company or customer information into public tools, and 44 percent admit to using AI in ways that go against their organization’s policies². None of this requires bad intent. It is what the employee clock looks like when the enterprise clock has not caught up: people solving real problems with the nearest tool, on terms nobody in the organization has read.

A ban stops the wrong clock

Faced with that picture, many organizations reach for the brake. In a Cisco survey of 2,600 privacy and security professionals in 12 countries, published in January 2024, more than one in four said their organization had banned generative AI, at least temporarily³. A widely reported case, told more fully in What Is AI Governance?, came in May 2023, when Samsung told staff in one of its largest divisions to stop using generative AI tools after employees uploaded sensitive internal source code to a public chatbot⁴.

As an emergency measure after a leak, a pause is reasonable. As a standing answer, it fails on the clock arithmetic. A ban stops the enterprise clock: the organization stops evaluating, stops measuring and stops learning. It does not stop the employee clock, because the tools remain a browser tab away on personal phones and home laptops. The survey finding that 44 percent of users already work around policy suggests how much a rule alone holds back. The same survey found that uploading sensitive company information to public tools was reported most often by employees whose organization had banned generative AI: 67 percent, against 33 percent where there was no policy². What a ban reliably does is move the use out of sight.

The opposite reflex fails too. Buying licenses for everyone and announcing an AI-first organization speeds up the employee clock, which was already fast. It does not move the enterprise clock at all. Usage climbs while the checks, the measures and the work itself stay as they were.

The response that works sits between the two, and it is designed to travel at the speed of the use. Its rules are short enough to remember while typing: a few lines on which data may go where, who checks what before it reaches a customer, and when to disclose that AI was involved. It offers an approved tool good enough that people choose it over the free one. And it makes a habit of asking teams, without blame, to show what they are trying, so that each experiment can be measured and then kept or stopped. Running that decision well is the subject of AI Leaders vs AI Followers; turning it into governance is the work of Module 7. The point here is narrower. When adoption moves at employee speed, a rule that takes a year to write governs nothing.

Tools also spread far faster than the redesign of work around them. A license can reach every desk in a week while every process stays as it was, the modern version of the electric motor bolted onto the old drive shaft that AI Is Changing Everything described. The employee clock measures access. Only the enterprise clock measures change.

Story: the explainers that outran the checks

In November 2022, the month ChatGPT launched, the personal-finance desk of a large US technology news site began a quiet test. An AI engine built inside the company drafted short explainers on basic financial topics. Editors wrote the outlines, then expanded and edited the drafts before publication. Over the next two months the desk published 77 such stories, about 1 percent of the site’s output in that period⁵. They ran under a generic staff byline, and readers could learn how they had been made only by clicking on it⁶.

In January 2023 another publication spotted the stories, and one was rightly cited for factual errors. Put yourself in the editor-in-chief’s chair that week. Readers trust the site on money. Many of your own staff are learning about the test from outside reporting. You have three options.

Option A is tempting: every story had been edited, and they were a tiny share of output. Option B is tempting too: it ends the embarrassment in one sentence. Decide which you would take before reading on.

The site took a version of option C. It paused the AI-drafted stories, and editors audited all 77. The audit led to corrections on 41 of them⁷. A small number needed substantial correction; others had incomplete company names, transposed numbers or vague wording, and in several cases the plagiarism checks that should have flagged closely copied sentences had not been used properly⁵. Leadership promised clear labels, previews of the technology for staff, and an AI working group at the parent company⁶. In June 2023 the site published a policy stating that no story would be produced entirely by an AI tool and that people would do the hands-on reviewing and testing⁸.

Read the case through the three clocks. The technology clock and the desk’s own employee clock had moved in weeks. The enterprise clock had not: no checks matched the new way of drafting, no disclosure was easy for readers to see, and many staff did not know what was being tested. So outsiders found the problems before the organization did. Option A would have kept the speed and the blind spot. Option B would have ended the learning with nothing written down. Option C cost a public correction, but it produced what the test had lacked from the start: an audit, a rule and a record of what failed.

Notice, too, that this was a sanctioned experiment, run by editors who knew their trade. If official use can outrun the enterprise clock, unofficial use, which leadership sees even less, can do so more easily. Learning in public is an expensive way to learn.

What this means for leaders

The speed of adoption is not a forecast to admire or fear. It is a fact about the employee clock, and it changes the job. Leaders used to set the pace by deciding what to buy. Today the pace is set by what people find, so the leadership work moves to seeing that use, setting bounds that people will actually follow, and learning from it faster than competitors do.

That work starts with an honest reading of your own clocks. Most organizations know roughly how fast a capability reaches employees; few know how long it then takes for the organization to evaluate it, bound it and change a process because of it. The second number is the one competitors are racing on. Being early to a tool is easy and temporary. Being quick to learn compounds.

Check yourself

1. Measured from launch, generative AI reached a larger share of Americans than the PC or the internet had at the same point.MythFact
2. At work, generative AI has spread far faster than the PC did.MythFact
3. Fast adoption shows that a technology is creating value.MythFact
4. In early 2024, far more US workers used generative AI than US firms reported using AI officially.MythFact
5. A ban stops AI use until the policy is ready.MythFact
6. Most employees who use AI at work do so through tools their employer provides.MythFact

Reflection: find your slow clock

What comes next

AI is spreading quickly, and part of that spread is easy to see: the assistant someone opened this morning. A harder question follows. Is AI only something people are trying, or is it already part of how the business runs, including in places leadership does not look? The next chapter, AI Is Already Inside Your Organization, takes up that question.