AI Sherpa Talk to us →

← All thought leaders

Insights · AI Thought Leaders, Distilled

Kai-Fu Lee on becoming AI Native

Podcast · BCG Thinkers & Ideas · 13 October 2026 · 33 minutes

AI Native with Kai-Fu Lee. Kai-Fu Lee in conversation with Nikolaus Lang, global leader of the BCG Institute. Listen on bcg.com →

About Kai-Fu Lee

Kai-Fu Lee is a computer scientist, executive, investor and author who has worked on AI for close to four decades. At Carnegie Mellon, where he took his PhD, he built the first large-vocabulary, speaker-independent continuous speech recognition system, and an Othello program that beat a world champion in 1988. He went on to lead speech and language work at Apple, held senior roles at SGI and Microsoft, where he set up Microsoft Research China (now Microsoft Research Asia), and was president of Google China. In 2009 he started the venture firm Sinovation Ventures, and in 2023 the AI company 01.AI. He is a Fellow of the IEEE, was named to the TIME 100 in 2013, and has co-chaired an AI council of the World Economic Forum. His books include AI Superpowers and, with Chen Qiufan, AI 2041.

Why we see him as a visionary

Few people have seen AI from as many sides: as a researcher who built early systems, as an executive who brought AI into products at large companies, as an investor who funds it, and as a founder who builds it today, in both the United States and China. His books explained, for a general audience, where AI was heading years before it reached most organisations. That range of vantage points is why we listen when he describes what AI will ask of companies next.

What they discussed

In this episode of BCG’s Thinkers & Ideas, Nikolaus Lang talks with Kai-Fu Lee about his new book, AI Native: The Mandate to Transform Your Company. Lee’s starting point is that AI is commoditising intelligence itself, which in his view makes it a larger shift than earlier general-purpose technologies (from 1:58). Six themes run through the conversation.

A mandate for the leadership, not a project for IT

Lee describes four flaws he sees in how companies approach AI (from 3:53). One is treating AI as another piece of software for the IT department to pilot, budget and scale. Another is treating it mainly as a way to cut costs, when he sees its larger value in growth, new markets, faster products and spotting risks that were missed (at 4:51). His short pitch to a chief executive (at 27:14) starts with using AI yourself and setting the example.

A general model does not know your company

Lang offers the picture of a bright graduate dropped into a department without context (at 6:31). Lee builds on it (at 6:52): a company has governance, workflows, decision processes, an organisation chart and goals, and a general model given a handful of documents cannot weigh what matters across all of them. He sees general models as useful for many contained tasks, but not for the hardest problems a company faces, which in his view need a map of the company that AI can work from (at 9:30 to 10:11).

Your own data, and where it goes

Lee argues that a company should keep its private data under its own control rather than hand it to an outside model provider (from 8:30). He points to what he calls dormant data, the unstructured record of meetings and of customer and sales calls, as a large resource that AI can now make sense of (at 11:37). And he sees the data a company holds, and what it learns by running AI on it, as what will set it apart from competitors (at 27:29).

A named person answers for what AI does

The unit of Lee’s AI-native company is a directly responsible individual (DRI) who leads a small team of people and many AI agents on a defined goal (from 14:02), with fewer layers of management between the work and the chief executive. He stresses that for important tasks AI cannot be held accountable, so the DRI checks the results, approves the process and takes responsibility (at 17:14 to 17:53).

Matching the model to the task

On cost, Lee suggests that companies route work by its stakes: the strongest models for the few decisions that matter most, and capable, cheaper models for routine work such as document comparison and customer service (from 20:18). He adds that a model that is cheap to use as a service can still be costly to run on your own hardware (at 22:17). Lang likens the habit of using top models for everything to taking a racing car to the shops (at 22:58).

What stays human

Lee expects AI to take on much of the analytical and execution work (from 24:28). What stays with people, in his account, is judgement where there is no past data to learn from, together with the qualities Lang draws from the book: taste, foresight, conviction and courage (at 23:51). For everyone, he adds, the people skills of trust, communication and teamwork (at 26:21).

The episode closes with a discussion of AI safety (from 28:57), where Lee calls for strong guardrails built into AI systems rather than left to the person writing the prompt.

AI Sherpa’s analysis: what needs care in Singapore and APAC

Lee’s themes fit what safe AI adoption needs. Here is where each needs care for organisations in Singapore and APAC.

  • Leadership (theme 1). We agree that AI is a decision for the top. For a financial institution regulated by MAS, it is also an expectation: MAS’s Guidelines on AI Risk Management (final, 7 October 2026; in force from 7 October 2027) expect the board to oversee AI risk and to address it explicitly in the firm’s risk appetite (para 3.4).
  • Context (theme 2). Knowing where AI already runs in your organisation, and on what data, comes before deciding what it should do next.
  • Recordings and calls (theme 3). Under Singapore’s PDPA, using the personal data in meeting and call recordings for a new purpose, such as AI analysis, generally needs consent for that purpose, unless one of the Act’s exceptions applies. Check your notices, consents and the exceptions before pointing AI at the archive.
  • Data leaving Singapore (theme 3). Many AI services process data outside Singapore, and the PDPA’s transfer limitation obligation (section 26) applies when personal data goes overseas. Whether a provider may train on your data is set by its terms; read them rather than assume either way.
  • Accountability (theme 4). For a financial institution regulated by MAS, the same Guidelines ask for human oversight that fits the risk (para 5.9) and keep the firm accountable for AI it buys from others (para 5.11); a named, accountable person fits that. For AI agents that act on your behalf, the DRI also needs the means to check and to stop them.
  • Company size (themes 2 to 4). Lee’s examples are large companies. In an SME the DRI may be the owner or one manager, and a company-wide AI platform may be out of reach. The lessons still hold at small scale: own AI at the top, know where your data goes, and name who answers for each AI tool.
  • Cost (theme 5). Choosing models by the stakes of the task is a governance decision too: it should be recorded, reviewed as models and prices change, and matched to what data each model may see.
  • People (theme 6). Fewer people doing routine work means more weight on the people who judge, check and decide. Plan the training before the restructuring.

Our analysis of MAS’s Guidelines on AI Risk Management sets out the detail for financial institutions.

How AI Sherpa can help with these themes

If these themes are live in your organisation, these can help.

  • Board & Executive AI Briefing. Your organisation’s own AI position and the decisions in front of the board, with a written follow-up. For: setting AI as a leadership decision
  • AI Readiness Diagnostic. In one to two weeks: where you stand on AI, the main risks, a shortlist of use cases and a roadmap. For: mapping where AI runs and what it needs to know
  • AI Risk Appetite Workshop. Scenario pre-work, then a half-day session. You leave with a risk appetite statement, calibrated policies and a board paper. For: deciding which data may go where
  • Governing Agentic AI. A practitioner workshop on how to govern AI agents against Singapore and APAC rules and guidance (IMDA, MAS, PDPA). For: naming who answers for each AI agent
  • Fractional AI Advisor. A monthly retainer for governance input and decisions as your use of AI grows. For: governing model choice as prices and models change
  • Team Enablement Workshop. Half a day on how to use AI tools well in real work, with agreed norms for responsible use. For: building the skills that stay human
Talk to us → See Advisory and Academy →

Sources