AI is everywhere and nowhere at once. So how long until it pays off?
Barely 1,200 days after ChatGPT, the boosters say AI should have transformed business already; the critics say it's a fad about to fizzle. The truth is messier than either. AI is genuinely making gains — and genuinely stuck against a wall of jagged capabilities, institutional inertia and human resistance. History suggests transformative technologies take decades, not quarters. Here are the five things to hold in your head at once.
The five things to hold at once
01Everywhere and nowhere at once
Step inside a large company today and AI is everywhere and nowhere at the same time. Employees use it to summarize meetings, draft emails and generate first cuts of presentations. But those gains haven't yet translated into a clear, economywide productivity surge — or fundamental changes to how people do their jobs.
So how long will it take AI to reach its potential? Answering it means sorting through a tangle of obstacles facing the business world: institutional inertia, human resistance to change, limited and often just plain messy data, privacy and security concerns, and the imaginative leap required to redesign how organizations actually function.
It is, as the essay puts it, an out-of-focus picture. But even though it isn't as sharp as we'd like, it can still tell us a lot about where we stand — and how far we have to go. The trap is reading it as either pure hype or pure failure; the reality sits stubbornly in between.
Adoption is high. Measured payoff isn't.
The most extreme AI boosters say the technology should have transformed business already — or will any day now. The loudest critics say AI is the latest tech fad, doomed to fizzle before it achieves anything — also any day now. Both camps have an "any day now." The honest answer is that the timeline is longer and lumpier than either side's story.
02Making gains: it's wrong to say adoption is stalling
For all the grumbling and negative news, AI is making real strides in the business world. Surveys of CIOs and CEOs consistently show companies planning to spend more on AI this year and next. A January Deloitte report and a separate Wharton study both show large companies moving beyond experimentation and integrating AI into essential operations — and three-quarters of 801 executives surveyed reported positive returns.
Where the gains are showing up
Given all that, says Wharton professor Ethan Mollick — who studies how companies adopt AI — it is flat-out wrong to claim adoption is stagnating. The "stuck in pilot mode" narrative, he argues, is an outdated idea that the evidence no longer supports.
The momentum is real: budgets are rising, deployments are leaving the lab, and a clear majority of executives say the spend is paying off. That is the half of the picture the doom narrative leaves out.
03The jagged frontier: great at some things, badly wrong at others
But the revolution is up against many obstacles. There's basic skepticism about the hype — boards and investors keep pushing for clearer evidence the spend is paying off — and AI hasn't yet shown it's versatile enough to transform whole industries at scale. Researchers have a term for its uneven capabilities: the "jagged frontier." Models are great at some things and surprisingly bad at others, and it's rarely obvious which is which until a company has already committed.
AI excels at tasks with clear structures — coding, legal-document review, financial analysis. But ask it to navigate the more contextual work that fills most of a workday, and the jaggedness shows: it gives wrong answers with great confidence, and can't draw on the human factors — judgment calls, unwritten rules, hard-won instincts — that never make it into training data. That, says independent analyst Benedict Evans, who tracks enterprise adoption, is a hard ceiling on what present-day AI can do.
Acemoglu believes current AI tools will affect only a fraction of jobs. And to do anything useful, AI needs a lot of "wrapping": the right data, the right permissions, the right guardrails, and defined roles for the humans who oversee it. Because every company's systems and workflows differ, that surrounding architecture usually has to be built from scratch — and that's a lot harder than it looks.
The frontier, drawn jagged
Structured vs. contextual
| AI excels | The jaggedness shows |
|---|---|
| Coding | Judgment calls |
| Legal-document review | Unwritten rules |
| Financial analysis | Hard-won instincts |
The trouble isn't only capability — it's deployment. Useful AI needs the right data, permissions, guardrails and human oversight roles wrapped around it, and because every company's stack is bespoke, that scaffolding is rebuilt each time. The "jagged frontier" plus custom wrapping is why so much spend produces dazzling demos and disappointing roll-outs.
04The human factor: the tech may be the easy part
As obstacles go, the technological issues may be far easier to overcome than the human ones. Simply put, a lot of people need to be convinced before the AI revolution can happen in earnest — and their reluctance is not irrational.
Automate the old process — or reimagine it
Who needs convincing
Executives
Five-year planning cycles, depreciation schedules on systems bought three years ago, and boards demanding returns. Risk aversion here isn't irrational.
Workers
People who believe they're training their own replacements aren't going to be enthusiastic partners in making it work.
Executives face real constraints; workers face an existential one. And management and employees alike can be hesitant to genuinely integrate AI into operations rather than use it for drudgework. The instinct is to bolt AI onto existing processes instead of rethinking them — which leaves most of the upside on the table.
Consider an insurer handling a fender-bender claim. The typical move is to use AI to speed the paperwork while keeping the same review and approval layers. The real opportunity lies in redesigning the process entirely. That kind of reimagining is difficult — and it threatens established hierarchies and routines, which is exactly why it so rarely happens.
05The big picture: transformations always take longer
Finally, it's worth remembering that transformative technologies have always taken longer to bring about the deep changes their champions promised. Electricity rewired civilization but took four decades to show up meaningfully in productivity data. The internet reshaped business, work and global competition — but needed 10 to 15 years to seep into the bones of the economy.
The internet's early years looked, from the inside, a lot like AI does now: spectacular promise, uneven results, and an industry with every incentive to tell you the revolution had already arrived. Stanford's James Landay — co-director of the Stanford Institute for Human-Centered Artificial Intelligence, who has spent years watching businesses struggle to absorb new technology — puts the realistic horizon in years, not quarters.
The pattern is consistent enough to plan around: the more profound the technology, the longer the lag between its arrival and its measurable impact. AI looks unlikely to be the exception.
How long until it showed up in the data
What makes the timing so easy to misjudge is that the early innings of a real transformation and a pure bubble look identical from the inside: big promises, patchy results, and incumbents insisting the future is already here. The internet was both overhyped in the short run and underestimated in the long run. AI is shaping up the same way.
What the piece is — and isn't — saying
This isn't a verdict for the boosters or the critics. AI is genuinely creating value and genuinely constrained; it is advancing fast and arriving slowly. The useful conclusion is a both/and: believe the destination, discount the timeline, and plan for a transition measured in years.
The gains are real
Spend is rising, deployments are leaving pilot mode, and three-quarters of surveyed executives report positive returns. "Stuck in pilot mode" is an outdated idea.
The ceiling is real too
A jagged frontier, custom "wrapping" for every company, and human inertia and fear are harder to clear than the technology itself.
The clock runs in years
Electricity took 40 years, the internet 10–15. AI looks faster but still slow — about five to 10 years to truly reshape how organizations work.
The boosters are directionally right about where this is heading. The skeptics are probably right about how long it takes.
AI will almost certainly prove as consequential as the internet, and probably will take about as long to reshape the economy. That single sentence dissolves most of the argument: there's no contradiction between "real and important" and "slower than promised." Both are true at the same time.
The picture is out of focus, but its shape is clear enough to act on: invest for the long arc, ignore the "any day now" on either side, and design the transition — for organizations and for workers — before the impact, not after.