Semafor Sep 17, 2026 1h 9m 53m saved
With Kai-Fu Lee, CEO of 01.AI and author of AI Superpowers and AI Native
Kai-Fu Lee has sat down one-on-one with roughly a hundred chief executives around the world, by his own count. He said almost all of them want an AI transformation and almost all of them are doing it wrong.
The standard corporate move is to hand the problem to the chief information officer and fund a queue of departmental pilots. Lee's argument is that this produces agents in customer service and legal that never touch the income statement, because the thing being transformed is the company itself, not its software stack.
"So they delegate and defer to the CIO who is an important asset to the company but whose job is to install software. It's not to drive AI transformation."
Lee ran Google China, wrote AI Superpowers in 2018 when almost nobody was framing AI as a US-China contest, and now runs 01.AI from Beijing, where he says the company is preparing a Hong Kong listing next year. His new book, AI Native, is the reason for the conversation.
The full interview is covered here so you can skip it. 69 minutes of audio, 16 minutes of reading.
Here are the 14 predictions that matter.
Key Takeaways
The CIO is the wrong owner of an AI transformation, because the job being done is organizational rather than technical
Middle managers are the "antibodies" — they are measured on headcount, so they want the change to move slowly
The only test Lee accepts is whether it moves the dial on the earnings call
His own agent keeps a "promise ledger" of what every manager said they would do, because the CEO forgets and the CEO's AI does not
An organization that swaps 8,000 human roles for AI workers needs 400 middle managers, not 3,000
US export controls have failed, on his read, and China will be self-sufficient in models and chips within two or three years
Anything purely digital is "dead in the water"; brick-and-mortar has three to five years
The next device is always listening, and he expects the West to refuse it
1. Delegating It to the CIO
Lee's diagnosis starts with a category error. Traditional-company chief executives, he said, treat AI the way they treated the internet or a database migration: a technology they are not expert in, therefore a technology to delegate. The result is a queue of departmental pilots, each with its own budget request, none of which changes the company.
"I've probably spent time with one-on-one with about 100 CEOs around the world and I noticed a lot of common themes. One is that they all want to do AI transformation. Two is they're mostly doing it wrong." — Kai-Fu Lee
"So they delegate and defer to the CIO who is an important asset to the company but whose job is to install software. It's not to drive AI transformation." — Kai-Fu Lee
He gave three reasons the work has to sit with the chief executive: the transformation changes strategy, execution and the organization together; it is not another piece of software; and the lack of technical expertise is not the obstacle leaders think it is, because what has to happen is already clear even when the solution is not.
A host pushed on the gap between activity and results, noting that every chief executive he interviews reports dozens or hundreds of pilots. Lee called the pilots cosmetic.
2. The Antibodies
The second obstacle is internal and human. Lee said companies are full of people who do not want the transformation to happen, and that they sit in the middle of the org chart.
Middle managers are not technology experts, so they fear being replaced by someone who is. They are also, he said, measured on how many people report to them, which makes any change that shrinks headcount a direct threat to their standing.
"Middle managers feel vulnerable because they're not technology experts." — Kai-Fu Lee
"So they want the AI transformation to be as slow as possible" — Kai-Fu Lee
His conclusion is that unless the chief executive visibly drives the effort and is personally seen to understand it, the middle of the company will wait it out.
3. The Earnings Call Test
Asked to describe the standard he proposes in the book, Lee said a leader should refuse to be satisfied by activity. The test is the quarterly report.
"It is about moving the dial on the earnings call. Making the quarterly report better, exceeding expectations in revenue, growth, profits." — Kai-Fu Lee
He added that the chief executive has to be a user, not a sponsor. Personal use produces a bottom-line effect directly, and it also removes the excuse that nobody at the top understands what is being asked.
"And I think what's really important is that the CEO personally needs to be using AI effectively." — Kai-Fu Lee
On whether today's chief executives are the right people for the job, Lee said they can be, but they have to stop managing the risk and stop deferring to direct reports who, in a traditional company, probably do not understand it either.
4. Boss AI's Truth Engine
Asked what at 01.AI passes his own test, Lee described building an agent for himself. That prototype became the company's product, Boss AI, which a host noted is assembled from 19 specialized agents.
The first function is visibility. Every meeting, sales call and customer service request is recorded and loaded into one database, and the model reads what no person could.
"We call that the truth engine because we basically collect all the data both the ones existing in the company but also new data such as every single meeting is collected recorded and entered into a giant database" — Kai-Fu Lee
His reasoning is that meetings are where strategy is argued and decisions are made, and that a chief executive normally learns about them through middle managers who filter for their own department and their own position.
"But if the CEO knows what's going on, we actually achieve what Ray Dalio calls radical transparency." — Kai-Fu Lee
Asked whether staff found it threatening, Lee said not at his own company, where everyone is an AI person, but that he expects resistance elsewhere. He framed the choice as Darwinian: a company that accepts the transparency will out-execute one that does not.
5. The Promise Ledger
A host raised an example from the book about people quietly letting tasks slide, and asked whether the technology goes past accountability into helping people finish more. Lee said it does both, and that the same tool works for every manager, with the benefit rising with company size.
The feature he named is a record of spoken commitments.
"So we keep a ledger of every middle manager's commitment." — Kai-Fu Lee
"And the CEO does forget, but the CEO's AI does not." — Kai-Fu Lee
He said formal targets now have to be augmented with every verbal undertaking, and acknowledged the obvious objection: the idea sounds threatening to anyone who is not the boss.
6. 2,000 Humans, 8,000 Agents
A host put the arithmetic to him directly: if the accountability is built in and the software does more work, why does that not mean far fewer people? Lee separated two cases.
A company that keeps output flat can cut, and will. A company that expands does not have to. His worked example takes a firm of 10,000 people whose work can now be done by 2,000 humans and 8,000 AI workers, and argues the ambitious version keeps all 10,000 people and adds 40,000 AI workers instead.
"But if you have the ambition and the foresight to embrace AI early, maybe your company can actually increase itself five times so that you'll have 10,000 human workers and 40,000 AI workers and grow your company for five times." — Kai-Fu Lee
The management layer is where his arithmetic bites. The 3,000 middle managers existed to run a chain of command over 10,000 people; 2,000 people might need 400. The AI workers need none, because each one receives the full information rather than having it passed down.
What replaces the layer is a smaller group he calls directly responsible individuals, a term he credited to Steve Jobs and to a recent paper by Jack Dorsey. He described a unit of one such person, three employees and twenty agents as moving like a startup.
"So, they're basically execution machines who are not accountable." — Kai-Fu Lee
That, he said, is why every cluster of agents needs a human partner: an agent cannot be fired, cannot have its pay cut and carries no fear, so the accountability has to sit with a person. He said the skill set is trainable, requires no coding, and that people who learn it should be paid three to five times more, because the role outranks some vice presidents.
7. What AI Superpowers Saw
A host asked what Lee saw in 2018 that others missed. His answer was that he was reading trajectories rather than making a call.
He had noticed that Chinese super-apps were already using recommendation engines well, and that AI needed volumes of data that a country comfortable with contributing it would supply faster. ByteDance was his central example, profiled before most readers knew the name and before TikTok was a hit. He noted Toutiao came first, and pointed to Meituan, which he described as running Groupon's and Yelp's playbooks through a recommendation engine, and to Didi.
"So basically I saw that AI required a lot of data. So whatever company know saw that early and whatever country was more willing to contribute the data which I called data is the new oil" — Kai-Fu Lee
8. Why Asia Adopts First
Asked what is different about running an AI company from Beijing, Lee said the product he sells lands better in Asia. Boss AI depends on a chief executive with real authority and on employees who accept meeting recording and message capture, and he said his firm meets no resistance to either across Asian markets.
He grounded it in a cultural contrast: the West weights individualism and privacy more heavily, while Asian firms are more collectivist and concentrate power at the top.
He then complicated his own point. American technology companies concentrate power too, and he named Elon Musk as the extreme case, saying Musk is more Asian in that respect than Asian chief executives are. He put Steve Jobs in the same category and said Mark Zuckerberg has the trait as well, adding that if the metaverse had been the right call Meta would be dominating today.
Asked whether this amounts to a shock the West does not see coming, Lee said no. Western firms can adopt it; they will need to see a competitor's results first, because leaders are pragmatic and a book will not move them.
9. Distilling and Open Weights
A host raised the American rebuttal to DeepSeek and Moonshot: that the open-weight Chinese models are distilled copies. Lee gave ground on talent depth and none on capability.
"I say that on the one hand the top American companies do have a tremendous talent advantage that if you look at the top 1,000 people in AI in the US they are significantly better than the top 10,000 people in China." — Kai-Fu Lee
At the top 100,000, he said, the two countries are comparable. Chinese labs mostly do use distillation but do not need it, several of the recent best products claim not to, and some American chief executives have conceded that distillation alone cannot account for the results.
"So I think it would be a terrible mistake to underestimate the capability of Chinese companies." — Kai-Fu Lee
His framing for the market is Apple against Android. Closed models are the premium product and will make more money from enterprises that get real value; open-weight models are cheaper on price-performance and will take much larger share. He called the arrangement healthy, because the open models cap what the closed ones can charge.
Pressed on whether a six-month lag is enough to justify a premium indefinitely, Lee agreed it is not — unless the talent density in Silicon Valley produces a breakthrough the open models cannot close quickly. On four years of evidence, he said, the pattern favors the challengers, and the jury is out.
10. Export Controls Backfired
Lee said he initially thought the chip restrictions might work, because training needs volume. Two assumptions have since broken: that Chinese engineers could not get a good enough result on far less compute, and that American semiconductor dominance is permanent.
"It's becoming clearer that the export controls do not work and it's interesting to see people acknowledge it hasn't worked therefore we need more of it." — Kai-Fu Lee
On inference he called the outcome settled. Chinese chipmakers do not have to match Nvidia, only to be good enough, and they are used to operating on thin margins against what he called Nvidia's very high ones.
Asked whether Chinese model developers move to Huawei's Ascend line, he said they already are, and that a Huawei chip may take twenty times the engineering effort to train an equivalent model because it lacks Nvidia's software libraries. He said the restrictions forced the practice that will close that gap.
"if you look at a horizon of two or three years China will be completely self sufficient self-reliant on its own model and chip ecosystems." — Kai-Fu Lee
11. Too Close to Call
On the capital gap between trillion-dollar American listings and Chinese firms, Lee's first answer was that a rising tide lifts all boats, and that if Anthropic and OpenAI are overvalued the Chinese companies are not.
Asked whether he was making that argument, he declined to call it either way. The valuations hold if the leaders keep pulling ahead and fail if open models keep closing in six months.
"We see that Anthropic's ARR is slowing down" — Kai-Fu Lee
He did concede the other side of it: a multi-trillion-dollar valuation funds a war chest for talent and chips, which is itself a route to the breakthrough that would vindicate the price. Deciding requires knowing how much OpenAI and Anthropic have in reserve, which he said he does not.
On Musk's group, he said the models look second tier now but that nobody should write Musk off, and repeated Musk's own line that he is always right but sometimes late. He expects SpaceX and Tesla to merge and thinks the combination could be the strongest empire in the sector.
"Well, I think one thing we've learned in from history is never count out Elon Musk, right?" — Kai-Fu Lee
"But the models, it does appear he's now second tier." — Kai-Fu Lee
12. Ontology, Not Just Models
A host set two competing theories against each other: Allianz's Oliver Bäte, who told him the largest companies will pull away because they can afford frontier models, and Netflix co-founder Reed Hastings, who told him the technology levels the field and that smaller firms move faster.
Lee rejected the first as stated. A general model, however good, cannot know a particular company's workflow, decision rights or structure, so a layer has to sit above it that captures and represents what the firm knows about itself. He called that ontology, credited Alex Karp with the argument, and said 01.AI is following Palantir's path.
The practical consequence is that the choice between open and closed models matters less than most boards think, and that the ontology layer also suits companies unwilling to put their data on someone else's cloud.
13. Dead in the Water
On large against small, Lee sided with Hastings for any business that is fundamentally digital.
"The writing is on the wall for anything digital." — Kai-Fu Lee
Entertainment, content, gaming and internet media are the exposed categories, and he said accumulated audience insight will not save them, because models can now write, cast and cut. He pointed to the short vertical dramas growing fast in China, where the scripts, the casting and the episode breaks are AI-made and tuned for retention.
A host offered the Hollywood counter-argument: studios use the technology to cut costs, lean on their intellectual property and earn more. Lee's reply was that this is the argument Blockbuster made against Netflix and that taxi fleets made against Uber. A studio dismissing short-form drama as artless, he said, is defending a shrinking market.
He put financial services in a middle category, protected for a while by regulation and balance sheets but not indefinitely. Physical businesses have three to five years, he said, naming a coal mine, a farm and a manufacturer, and they compete with their existing rivals rather than with a one-person startup. That, he said, is the audience his book is written for.
14. The Always-On Device
Asked what he sees coming that Silicon Valley does not, Lee named hardware. His claim is that the phone is the wrong form factor because speech is the preferred interface and an app is too slow when the user talks to an agent hundreds of times a day.
"I think the phone is the wrong device for AI because speech is unquestionably the preferred mode of interface" — Kai-Fu Lee
A device built for speech is always on, which makes it always listening, which gives it something close to unlimited memory. He is investing in the category, in both consumer and enterprise forms, and said Boss AI is a limited version of the idea because it listens only in meetings. He described a founder he backs who records everything and gets useful marriage counseling out of it, not because the model is good at counseling but because it has heard every conversation.
The constraint he sees is social rather than technical.
"I would argue the western world will not let that happen and that will cause it not to have the huge amounts of data" — Kai-Fu Lee
Asked directly about the always-listening device OpenAI is building with Jony Ive, he was blunt.
"Yeah, it will not work because people won't fully accept it." — Kai-Fu Lee
He expects Chinese manufacturers to get there first on cost, supply chain and turnaround speed, and expects Chinese consumers to accept the trade eventually.
"I'm not saying Chinese people don't care about privacy. They do. But it's one of the priorities, not an unassailable requirement, which it is for the Western world." — Kai-Fu Lee
Bonus Insights
The two hosts debriefed after the interview and pushed back on several points.
One doubted that AI-generated short video will disrupt the major studios on the same timeline in the US, and wondered whether the appetite for it is a China-specific taste. He allowed that the hopeful reading is independent filmmakers challenging the studios
The hosts named a tension they did not think Lee resolved: he argues legacy companies with a digital face will be disrupted fast, while also arguing those same companies hold rich internal knowledge that ontology can unlock
A second unresolved point: Lee says the chief executive must personally drive the transformation, and also that the company should bring in an outside vendor to do it
One host noted that Bridgewater, Ray Dalio's firm, is the standing example of radical transparency and remains an outlier after decades, which suggests the obstacle is cultural rather than technical
The hosts said no American chief executive they interview talks about surveillance of work in the terms Lee used, even if some think it, and that this is part of what has fed anti-AI sentiment
One argued there is a pro-employee version of the pitch nobody has made well: use the technology to cover what a person is weak at so they can spend more time on what they are good at
They said the parts of Lee's argument they do hear from Western and European executives are the collapse of functional silos and the wish to capture the expertise that walks out of the building at retirement, and that both run into resistance
They closed on risk: centralizing every piece of company data for a model to read creates security and intellectual-property exposure, so being at the frontier is itself a risk position, and the bold may be the ones rewarded
Lee's bottom line is that AI transformation is a change to the org chart and the reporting culture rather than a software purchase, and that the companies which accept the loss of privacy it requires will take share from the ones that do not.
Products, Companies & Tools Mentioned
01.AI (Lee's company, which sells Boss AI, is building an ontology layer and is eyeing a Hong Kong listing next year)
Palantir (The model he says 01.AI is following, and Alex Karp's argument that a great model alone does not solve an enterprise's problems)
DeepSeek and Moonshot AI (The open-weight Chinese labs American firms accuse of distillation; Lee says they mostly use it but do not need it)
Nvidia and Huawei (Every Chinese training company he knows is moving to Huawei's Ascend line or considering it, despite roughly twenty times the engineering effort)
OpenAI and Anthropic (The valuations he will not call either way; he cites Anthropic's slowing ARR as the thing to watch)
ByteDance (His 2018 example of data compounding into advantage, profiled before TikTok broke out; Toutiao came first)
Meituan and Didi (The other two Chinese recommendation-engine businesses he pointed to)
Tesla and SpaceX (He expects a merger and thinks the combined group could be the strongest empire in the sector)
Netflix (Twice over — Reed Hastings' leveling argument, and the Blockbuster comparison he uses against studios)
Meta (His example of concentrated founder power producing the wrong call on the metaverse)
Books & Resources Mentioned
AI Native – Kai-Fu Lee (The new book, written for brick-and-mortar companies with three to five years of runway)
AI Superpowers – Kai-Fu Lee (The 2018 book that framed AI as a US-China contest and argued data is the new oil)
Jack Dorsey's recent paper on directly responsible individuals (Where he says the term Steve Jobs coined was popularized again)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: