When factories stopped making their own power
Around 1900, a factory that wanted electricity had to make it. Power was a capability you built in-house, with your own generators, your own plant, and your own staff to run it; as late as 1909, 64 percent of the electric motor capacity in US manufacturing still ran on power generated on site.1
Then central stations took over. Samuel Insull’s Commonwealth Edison in Chicago proved that large utilities, running big steam turbines, could make power more cheaply than any factory could make it for itself,23 and by 1919 utilities supplied 57 percent of the electricity used in US manufacturing.1 For almost every factory, power was a necessity rather than something it competed on, so generating it went from a capability to a line item. The capability didn’t disappear, though: it moved to the companies for whom power was the product.
Show data table
| Year | On site | Utilities | Motor capacity (k hp) |
|---|---|---|---|
| 1899 | 63% | 37% | 475 |
| 1904 | 72% | 28% | 1,517 |
| 1909 | 64% | 36% | 4,582 |
| 1914 | 56% | 44% | 8,391 |
| 1919 | 43% | 57% | 15,612 |
| 1925 | 40% | 60% | 25,092 |
| 1927 | 37% | 63% | 29,153 |
| 1929 | 36% | 64% | 33,844 |
| 1939 | 36% | 64% | 44,827 |
Source: Historical Statistics of the United States, Table Dd848-853 (Atack & Bateman), horsepower of electric motors. See also Devine, “From Shafts to Wires”, 1983.
AI is now doing the same to knowledge work—which is why “Will AI replace designers, engineers, writers, lawyers?” is the wrong question. The same profession can be safe in one company and at risk in the next; even two people with the same title, at the same company, can face very different futures. What matters is not the profession but two things about the work itself: what kind of work it is, and what it is for. The rest of this essay takes them in turn.
Strategy and execution are a spectrum
Start with the kind of work. Every task is some mix of two things:
- Execution is turning a known intent into output. This is where AI is strongest, because someone has already specified what they want.
- Strategy is deciding what the intent should be: what’s worth doing, in what order, and why. This is where original insight comes from.
Vaughn Tan has a useful name for the second: meaningmaking, “any decision we make about the subjective value of a thing.” In his view, every important and difficult business decision requires it, and “only humans can do meaningmaking work, while machines cannot do meaningmaking work at all (for now).”4
Throughout this essay, “strategy” means meaningmaking in Tan’s sense—a function, not a job title. It happens well outside the boardroom: a designer defining a brand’s visual language and an engineer choosing what a system should do are both doing strategy, because the core of their work is deciding what’s worth making rather than making it.
That matches my own experience. AI is remarkably good at translating between domains—English into code, text into images, long text into short—and at generating output once the intent is clear. What it isn’t good at is creating meaning: judging what matters most, what is worth doing at all, or how best to get from where you are to where you want to be. In short, execution is translation, and strategy is meaningmaking.
Still, the two aren’t a binary. Every piece of work sits somewhere on a spectrum, depending on how much of the intent is still open when it reaches you. In software, that spectrum runs from typing out a function someone else specified, through implementing a feature from a spec and designing the system behind it, to choosing which features to build and which customers to build them for. Even the first step involves small judgment calls, but at each step the intent is less settled, so there is less to translate and more meaning to make.
- Choosing which customers to build for
- Choosing which features to build
- Designing the system behind a feature
- Implementing a feature from a spec
- Typing out a function someone else specified
Shares are illustrative, not measured.
Very few jobs sit at a single point on this spectrum; most people do a mix. And the mix is what matters, because AI takes over the spectrum from the execution end.
Necessities converge on the standard answer
The second question is what the work is for. Here, too, there are two poles:
- Necessities are functions a company needs but doesn’t compete on. They have a point of diminishing returns: past “good enough,” the next dollar is better spent elsewhere.
- Advantages are functions a company competes on. Because an advantage has to be defensible, it must come from some original insight,5 and it can’t come from the model, for the same reason economies of scale are not exclusive: anything available to everyone is an advantage to no one. AI lowers costs for everyone, so on its own it can’t widen anyone’s value-cost gap.
As with strategy and execution, few functions are pure examples of either; most contribute something to how a company competes, just more or less of it. The split has exceptions, too. A function that’s usually a necessity can briefly become decisive during a replatforming, a crisis, or a regulatory change—regulation can create opportunity for whoever handles it best—and a function that’s an advantage overall still has plenty of corners that are pure upkeep.
This is where the two spectrums meet, because where a function sits on this one largely decides where its work sits on the other. Necessity work still involves judgment—what counts as good enough, which trade-offs to accept, what this particular contract needs—but it’s judgment with a standard answer. Since the company doesn’t compete on the function, it has no reason to differ from everyone else, so the meaningmaking converges on the same store, the same process, the same contract everyone else has. An intent settled by convention is still a known intent, and turning a known intent into output is execution, even when it feels like judgment to the person doing it. The standard answer, in turn, is precisely what an LLM produces, because it has absorbed every example of it.
The two spectrums, then, are distinct but move together: the less a function contributes to competitive advantage, the sooner it hits diminishing returns, the more of its meaningmaking the market has already done, and the further its work slides toward execution.
At the advantage end, the opposite holds. The intent stays open, because differing from the standard answer is the whole point: all good strategies are unique, and best practices are a floor, not a ceiling. Here, AI raises the baseline without giving anyone an edge, since every competitor has the same models; the edge still has to come from the insight.
You might object that cheaper execution will create more demand for necessities. It will—companies will review more contracts, produce more content, and answer more support tickets—but that extra demand is for the standard answer, so it’s AI that meets it, not the people who used to.
The same profession, a different company
Put the two ideas together and an important consequence follows: where a profession lands depends on the company, not the profession. Few professions are an advantage everywhere or a necessity everywhere; some simply sit closer to one end in more companies than others.
Take software engineering. At a software company, it’s the product itself, and as the cost of building falls, the cost of not building rises, so the company will keep building more. At a consumer brand, the same work is a necessity: the brand builds a good-enough store and stops. Design, likewise, is the advantage for a fashion brand that competes on taste, and a necessity for a business-to-business distributor that just needs a usable catalog. Even R&D—scientists at a pharma company or chefs at a restaurant group, whose job is by definition to find the insight the company competes on—turns into necessity work once the same scientist moves to quality control, or the same chef to a corporate cafeteria.
None of this is entirely new. Geoffrey Moore distinguished a company’s core, which differentiates it, from its context, which it should outsource or automate.6 Nicholas Carr made the electricity argument about IT, and his critics answered with Walmart, Amazon, and Google—both sides were right, because it depended on whether IT was core.7 And Wardley mapping charts how capabilities drift from genesis to commodity.8 What AI changes is the pace: it accelerates that drift for a whole class of knowledge work, and makes it starker than ever that the same profession can be core in one company and context in the next.
Two moves to stay relevant
If AI absorbs execution, and necessity work converges into execution, then staying relevant means doing work whose intent is still open. Putting the two spectrums together gives a simple map:
| Execution | Strategy | |
|---|---|---|
| Advantage | Pressured. Execution still gets cheap, which squeezes headcount and price. | Safest. This is where original insight about how the company competes happens. |
| Necessity | Most exposed. AI replaces it outright, as with a consumer brand’s storefront, routine contract review, or first-line support. | Exposed. Demand is thin, because the intent converges on the standard answer and leaves little to decide. |
The rows decide whether the market for your work grows or shrinks; the columns decide whether you capture value in it. The natural path is therefore up and to the right—in that order. In a necessity, demand for strategy is thin by definition, so moving right without moving up just means getting better at something nobody is asking for.
Move up. Work in an area that is, or will become, a source of competitive advantage for your company, since that’s the only place strategy is in demand. There are two ways to get there. One is to change the function, moving into adjacent work that’s hard for AI to replicate. Design is the obvious candidate: AI still has very limited taste, and taste increasingly sets brands apart. But this only helps if design is an advantage for that company; to one that doesn’t compete on taste, it’s just another necessity. The other is to change the company, doing the same work somewhere that competes on it—the engineer at a consumer brand moves to a software company, the designer at a distributor to a fashion house. Either way, remember that “AI can’t do it well yet” describes today’s models, not a moat: if you move up but stay on the execution side, you’ve found a better market, but you aren’t safe.
Move right. Within that area, do more strategic work and less execution, because being in the right context protects the function, not everyone in it. That doesn’t necessarily mean fewer people. As execution gets cheaper, the cost of not executing rises, so companies that compete on a function will do far more of it: software companies will build more software than ever and, since AI moves validation downstream, test more ideas by shipping them. All that execution, though, demands more coordination and more judgment—about what to build, for whom, in what order, and whether it’s any good—and that work still needs humans. Demand for people may well grow; what changes is what they do. The catch is that the supply of people chasing that work will grow faster. AI will create new roles, but everyone pushed out of execution and out of necessities is heading for the same corner of the map, and there won’t be room for all of them.
This second move is what people mean when they say “AI won’t take your job; somebody using AI will.”9 The slogan undersells it, though. It suggests you keep the same job and pick up a new tool, when in reality you may have to change what your job is.
Software engineering is a good example. We’re not far from a world where AI can execute complex software projects end to end, with little or no human supervision. In that world, the traditional software engineer—who spends most of the day turning specs into code—may disappear from most companies, even those where software is the advantage. The work shifts upstream, toward deciding what to build, for whom, and why, and toward orchestrating far more execution than any team could do by hand. There may be as many people doing that as there were engineers, or more; but that isn’t the same profession adopting AI—it’s a different profession.10
Not everyone will like that, and not everyone who excelled at the old job will excel at the new one: technological change doesn’t have to destroy jobs in aggregate to destroy the specific job you were good at. Nor can anyone opt out—the models won’t get worse because we’d prefer they did.
Nor is this a move you make once. As models get more capable and cheaper, the line between what AI can and can’t do keeps moving up and to the right with you: work that took judgment last year becomes execution this year, and execution that was too expensive to automate becomes cheap enough. Staying ahead means moving continuously.
Source: METR, Time Horizon 1.1, updated May 2026. State-of-the-art models only. Trend fitted to models since 2023 below METR’s 16-hour reliability ceiling: doubling every ~131 days.
Is strategy safe?
Everything so far assumes that strategy is where AI stops. We can’t know that for sure, but we can reason about where original insight comes from.
Raw intelligence is close to a commodity: companies hire it on the open market, and many people could fill any given role. What’s unique is context—and context isn’t something the company owns in isolation. It comes from the relationship between the company and its people. The company brings its history, customers, capabilities, and competitive environment; the person brings their past experience, the other companies and industries they’ve seen, and the particular way they relate to this one. Original insight happens where the two meet.
Most of that context is also illegible. It rarely gets written down, living instead in people’s heads, in hallway conversations, and in a feel for what customers will tolerate and what the organization will get behind. David Autor calls this Polanyi’s paradox—“we know more than we can tell”—and argues it has historically been the main barrier to automating work.11 A person absorbs this tacit knowledge by being immersed in it, whereas an LLM only has what someone feeds it, so it reasons from a thin, legible slice of the context and lands closer to the standard answer.
The early research fits this picture. On paper, LLMs can generate and evaluate business strategies about as well as entrepreneurs and investors.12 But when experts spent more than a hundred hours executing research ideas, the LLM ideas—which reviewers had rated more novel than the human ones—lost much of their score, and the ranking flipped.13 One plausible reason is that AI reasons from data, which makes it backward-looking and imitative, while humans can reason from theory and hold beliefs the data doesn’t support yet.14
Source: Si, Hashimoto & Yang, “The Ideation–Execution Gap”, 2025, Tables 4–5. 19 human and 24 AI ideas. Before execution, AI ideas scored higher on every metric. After, the differences were not significant.
None of this is a guarantee, though. Context is becoming more legible as companies record, transcribe, and log more of their work, and models may get better at inferring what isn’t said. There’s no principled reason a model, immersed deeply enough in a company’s context, couldn’t reach some of the same insights a person would—this is Tan’s “(for now)” again.
So the claim here isn’t that strategy is safe forever; it’s a claim about order. AI absorbs work from the execution end first, and necessities before advantages, because that’s where the standard answer is good enough. The work that depends most on unique context goes last—and whether “last” means never or fifteen years from now, that’s the end of the map to move toward.
Even if AI does get there, the relational nature of context still matters. A model can be given everything a company has written down, but not the career you brought with you, or how that career interacts with this particular company. Once raw intelligence is something anyone can rent, advantage comes from context—and part of that context is you.
The last work to go
Go back to the factories. Most of them stopped making their own power, and nobody missed it: the capability moved to the companies for whom power was the product, and the factories got to focus on what they competed on.
AI is doing the same to knowledge work. It won’t take over whole professions; it takes over execution, starting wherever the standard answer is good enough—and in a necessity, the standard answer is all anyone wants.
So instead of asking “Will AI replace my profession?” ask two better questions. Does my work matter to how my company competes, or is it just necessary? And am I executing, or deciding what gets executed? If the answers are “just necessary” and “executing,” you know which way to move. The profession you trained for may not survive the change, but deciding what’s worth doing will be the last work to go.
Footnotes
-
The U.S. Economy in the 1920s (opens in new tab), EH.net. ↩ ↩2
-
Brian Potter, “The Birth of the Grid,” (opens in new tab) Construction Physics. ↩
-
Nicholas Carr, The Big Switch (opens in new tab), 2008. ↩
-
Vaughn Tan, “The meaningmaking lens on AI,” (opens in new tab) 2026. ↩
-
This holds even for advantages that don’t look like insight. Network effects, distribution, proprietary data, or a regulatory moat can all be sources of advantage, but someone first has to see that they matter here and how to build them—and that is the original insight, which nobody can yet rely on AI to have. ↩
-
Geoffrey Moore, Dealing with Darwin, 2005. ↩
-
Nicholas Carr, “IT Doesn’t Matter,” (opens in new tab) Harvard Business Review, 2003. ↩
-
Simon Wardley, Wardley Maps (opens in new tab). ↩
-
Usually attributed to economist Richard Baldwin, who said it at the World Economic Forum’s Growth Summit in 2023. See “This phrase about AI not taking your job sounds smart—but is it true?” (opens in new tab) ↩
-
This raises a problem. If execution is where people have traditionally learned the judgment that strategy requires, what happens to the apprenticeship path once AI absorbs execution? Where do the next strategists come from? ↩
-
David Autor, “Polanyi’s Paradox and the Shape of Employment Growth,” (opens in new tab) NBER, 2014. ↩
-
Felipe A. Csaszar, Harsh Ketkar, and Hyunjin Kim, “Artificial Intelligence and Strategic Decision-Making: Evidence from Entrepreneurs and Investors,” (opens in new tab) Strategy Science, 2024. ↩
-
Chenglei Si, Tatsunori Hashimoto, and Diyi Yang, “The Ideation–Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas,” (opens in new tab) 2025. ↩
-
Teppo Felin and Matthias Holweg, “Theory Is All You Need: AI, Human Cognition, and Causal Reasoning,” (opens in new tab) Strategy Science, 2024. ↩