Benedict Evans on AI, jobs, and why it’s probably going to be okay

Why does it matter? Because the companies with the most capital and hype are probably the ones who'll profit least

Benedict Evans has mapped more technology platform shifts than almost anyone working today — and his read on AI arrives with the calm authority of someone who has seen this pattern before, twice. The transformation is real. So is the mismatch between where most people expect the value to land and where it will.

• Foundation model companies — OpenAI, Anthropic, the rest — are probably building the telecoms infrastructure of the AI era: enormous capex, no durable pricing power, converging toward commodity margins
• The analytical framework everyone uses to predict job losses (decomposing professions into automatable tasks) shares the same flaw as 1970s expert systems — it mistakes the task for the job
• Even teenagers show only 15–20% daily active use of AI tools; most of the applications that will define the next decade haven't been built
• Software engineering — the profession that looked most immune to automation three years ago — turned out to be the most transformed; Evans's point is that direction-of-disruption is almost never predictable

OpenAI and Anthropic are probably building tomorrow's low-margin commodity utilities — just like the telcos thought they'd own the mobile internet

Three to six companies selling essentially the same thing, burning hundreds of billions in capex, with no structural mechanism for pricing power — that's Evans's read on the foundation model layer, and it's not what the current consensus has priced in.

"The models don't seem to have network effects," he said. "So there doesn't seem to be a winner-takes-all effect where one of these will run away ahead of the others." Without differentiation and without network effects, competition persists indefinitely — and persistent competition means margin compression. Sam Altman's vision of selling intelligence "like water or electricity" doesn't unsettle Evans; it confirms the bear case. Water utilities are not great businesses.

The more instructive analogy is telecoms. Global mobile industry: roughly a trillion dollars in revenue, $200 billion in annual capex, data consumption on an exponential curve since 2010. "The stocks have gone nowhere in 25 years because it's an X-growth low-margin commodity utility — all the cool stuff is made by you." The telcos believed they'd control everything running over their pipes. The value traveled up the stack.

When models commoditize, distribution becomes the decisive moat. Meta — widely dismissed inside tech circles — ranks alongside ChatGPT and Gemini in actual usage surveys because they "sprayed it on every service surface and it wasn't that bad. It was fine." Evans's framing: "Distribution of an adequate product when the field is basically commodity — distribution and brand become a big deal."

Scoring a profession's AI exposure by decomposing its tasks is "the most ridiculous bunch of deluded horseshit"

Evans's most useful analytical move is the distinction between task and job: what AI can automate versus what you were actually hired to accomplish. Almost every published AI job-impact study conflates the two.

The O*NET-style decomposition exercises — "17% of a senior law partner's tasks could be automated" — share the fatal flaw of 1970s expert systems. "You can't look at a senior partner at a law firm and say, well, 17% of their work could be automated," Evans said. "This is horseshit." Building up logical steps to describe a profession, like building edge detectors to recognize a cat, eventually produces hundreds of steps that don't work.

McKinsey doesn't get paid for the 75-slide deck. "What you actually pay Bain to do is to go and walk all over your enterprise and work out — yes, but why is it that you didn't do that? And how do the politics of this work?" Claude Code can generate a passable deck today. It cannot diagnose why an organization failed to act on the last three decks.

The same split runs through software: "Claude Code can write you the code, but what code do you want? What features do you want? Who's your customer?" Amazon gets you the SKU; knowing which SKU is a separate job entirely. The elevator operator was the exception — the task literally was the job, so automation killed it. Almost nothing else is that simple.

We're in 1997 — and picking today's AI winners is as futile as choosing between Excite and Yahoo

Even among 13-to-18-year-olds, surveys show roughly 15–20% are daily active users of AI tools, with about 60% of that demographic saying they don't use it at all. The people who've "taken the pill" systematically mistake their own depth of adoption for the median.

"If you're going to make the internet comparison, it's like we're in 1997," Evans said. "It's very exciting. Most stuff kind of doesn't work yet. Most of the stuff that people are going to do hasn't been built yet." Picking now between OpenAI and Anthropic is as futile as the Excite-vs-Yahoo debate — "and the answer was no, generally." The winner was a has-been PC company from Cupertino. Nobody would have predicted it.

The direction of disruption is almost always the surprise. In 1997, newspapers looked safe — their printing bills would go down. Taxi drivers looked immune — the internet had nothing to do with moving people. Then Uber. Personal trainers looked like obvious survivors; Evans now describes balancing his iPhone against gym equipment and asking AI to build his routine and correct his form in real time. "You can't predict which things are going to be exposed."

Four years ago, software engineering was the last profession anyone expected AI to reshape. Now it's the most transformed role in any industry. The most valuable AI companies of 2030 probably haven't been founded yet — and what they'll do is currently indescribable, which is exactly how it's always worked.

The consultants AI was supposed to eliminate are being hired more aggressively than ever — by the AI labs themselves

Deploying AI across an enterprise is a project. Projects require people. Companies don't have spare people.

"Companies do not have lots of people sitting around waiting to do a big new piece of analysis or work out how they're going to redesign their internal workflows," Evans observed. Mapping automation opportunities, integrating vertical systems into horizontal platforms, retraining thousands of employees — each step is its own engagement. So who runs them? Bain. BCG. McKinsey. Accenture.

The irony Evans finds sharpest: Anthropic's and OpenAI's "forward-deployed engineers" — the on-site implementation specialists both companies are aggressively building out — are functionally Accenture consultants in San Francisco zip codes. "A forward-deployed engineer is like an Accenture outsourced software developer who lives in San Francisco." The brand changed. The function didn't.

"What's really funny about this trend is you would think AI is going to mean consultants were going to be gone. Instead, the most cutting-edge AI labs are the ones most investing in these folks." For anyone building AI tools for enterprises, the integration and change management layer is likely a larger opportunity than the software itself. Organizational capacity — not model quality — is the binding constraint in AI deployment.

Accountant headcount rose through adding machines, mainframes, ERP, and spreadsheets — and it's rising again now

Evans shows two charts tracking accountant employment across the 20th century. The line goes up through every wave of supposed automation: adding machines, punch cards, mainframes, databases, ERP, cloud, spreadsheets. It went up again after 2000.

The mechanism is price elasticity. Make something cheaper and people do more of it, not less. "Young people won't believe this, but before Excel, junior investment bankers worked really long hours — and now, thanks to Excel, Goldman's associates all work at lunchtime on Fridays." Except that's not what happened. The productivity gain generated more analysis, more clients, more complexity, more revenue — not fewer bankers.

Evans grounds this in 200 years of consistent evidence: "You go back to 1800, like 90% of us were peasants. Ever since then we've been automating jobs and creating new jobs — and you can always see the job that's going to go away, and you don't know the new job because it doesn't exist yet." Railway engineer. Telephone operator. Machine learning engineer. The new jobs sound implausible until they're the most obvious thing in the world.

The transition is genuinely painful — towns hollow out, entry-level hiring freezes, career ladders reshape. Evans doesn't deny that. The macro trajectory over 200 years, however, runs one direction only.

The AI risk that actually matters isn't server cooling bills — it's that mass personal harm now costs nothing to execute

US data center water consumption: 0.017% of US water consumption, per a Livermore Lab study Evans went and read. "The water stuff is just nonsense." Data centers use about 5% of US energy and might grow by one percentage point per year. Real costs, real tradeoffs — but planning problems, not civilizational ones.

The genuine risks follow a pattern Evans traces explicitly to social media. The internet let the only gay kid in a rural village find their community — and also let the only pedophile in a village find theirs. "We connected everybody and unfortunately that meant we connected all the bad people and all of our own worst instincts and every problem in society. And so that will happen again with AI."

The specific capability Evans flags isn't hypothetical: "A 15-year-old kid couldn't use Photoshop to make hardcore pornographic nudes of every girl in their high school and send them to the whole school in one afternoon. And now they can." That's today's capability, accessible to anyone with a phone.

The backlash to AI mirrors the backlash to social media — some concerns genuinely real, some sort of real, some immune to evidence (the persistent myth that Facebook sells user data). Conflating manufactured fears with genuine ones makes both harder to address. The policy energy that goes toward server cooling bills should be going toward what AI actually amplifies at scale: harassment, manipulation, non-consensual imagery.

Most of the world hasn't reached the 1997 moment yet — and that gap is the map of where the next decade's opportunities live

Most of the world hasn't started. The applications, habits, and business models that will define the AI era are mostly unbuilt, which means the gap between early adopters and everyone else isn't a problem to manage; it's the territory where the decade's largest opportunities concentrate. Evans's advice to individuals follows the same logic: performing moral superiority about AI on Bluesky is a real choice, he noted — it just won't help. Diving in will. The accountants who learned Lotus 1-2-3 didn't get replaced by spreadsheets. They became the people everyone needed next.

It'll probably be okay. Not certainly. Probably.

---

Topics: artificial intelligence, future of work, job automation, technology cycles, foundation models, AI sentiment, distribution strategy, professional services, platform shifts, tech investing