Read Kai-Fu Lee's new book "The Future of AI Is Here": Why do 90% of companies spend money to introduce AI, but only 40% actually make money back?
By the way, I am in the same WeChat group as the famous entrepreneur Kai-fu Lee. But most of the time, like everyone else, I learned about his current situation from the media. Occasionally, I would see him surface in the group and leave one or two comments about the industry. But this time is different. He has just released a new book, “The Future of AI is Here: Transformation Strategies for CEOs, Enterprises, and Individuals in the Wave of Change (opens in a new tab)”. The set of numbers quoted at the beginning of the book just hit the spot that I have seen when accompanying companies in their AI transformation in the past few years.
A set of numbers that debunks the myths about AI transformation
In his book, Kai-Fu Lee quoted survey data from McKinsey (opens in a new tab)’s annual report “The State of AI in 2025”: nearly 90% of companies are already using AI, but only 39% said that AI has brought substantial financial returns. In other words, nine out of ten companies have already boarded the AI boat, but less than half have actually rowed to the shore.
To be honest, this gap is not because the model is not smart enough. Kai-Fu Lee cited the 2025 report of the MIT NANDA Project, calling it the “learning gap”: AI projects fail often because the system is not truly connected to the enterprise’s workflow, context, and feedback loops. A deeper reason is the misplacement of scenarios: Scenarios that can really improve financial reporting, such as pricing decisions and supply chain optimization, have the highest implementation threshold and involve the most departments; scenarios that are easy to deploy, such as meeting minutes and first drafts of copywriting, have almost no impact on the profit and loss statement. Most companies’ AI investments end up focusing on the latter.
Kai-Fu Lee used to be Google’s global vice president and president of Greater China. He is now the founder and CEO of Zero One Thing and chairman of Sinovation Ventures. The recommended list of this book is not just made up casually. It contains a lot of gold: Nobel Prize winner Jeffrey. Hinton, Michael. Spencer, Microsoft CEO Satya. Nadella and AMD Chairman Su Zifeng both wrote recommendations. I read Spencer’s words twice: He said that the most profound impact of AI on enterprises is to eliminate the information asymmetry that “executors know but decision makers cannot see”. This sentence, combined with the previous 39% information, is the core issue that the entire book will deal with: AI is already smart enough, but companies have not yet learned how to use it on the edge?
Mistake 1: The CIO should not shoulder this responsibility
There is a dialogue in the book that seemed particularly familiar to me. Kaifu Lee asked the CEO, “Are you serious about promoting AI transformation?” The other person replied, “The chief information officer has been given full responsibility.” When asked again “What is the KPI”, the answer usually stops at “Deploy a few AI agents and increase usage”; when asked “What about ROI”, the conversation gets stuck.
He quickly dissected three-layered problems: The CIO’s bounden duty is to seek stability, but AI transformation is a risky change, which is a mismatch of functions; AI scenarios that can truly improve financial reporting often require cross-department mobilization of the supply chain, finance, and pricing, which is beyond the scope of the CIO’s authority, which is an overreach of power; building a new generation of AI technology architecture from scratch is also beyond the capabilities of the general internal IT team, which is a lack of technical capabilities. If the three layers are stacked together, the transformation will naturally be unstoppable. This is also the same thing I always wanted to make clear when I wrote the article Enterprises import AI, bosses should not just watch a few days ago.
Mistake 2: Going from bottom to top will be eaten by “pocket veto”
The second mistake is more subtle. The company first lays out the model structure, and then collects requirements from various departments: the marketing department needs copywriting tools, the customer service department needs automatic response robots, and the human resources department needs a resume screening system. This mechanism looks pragmatic, but in fact it cannot produce truly valuable scenarios, because those scenarios are inherently cross-departmental and will not appear on any department’s wish list.
Even more troubling is the implicit resistance of mid-level managers. Kai-Fu Lee used a term called “Pocket Veto”: the supervisor nods superficially, stuffs the task into his pocket after the meeting, and lets it disappear naturally in daily affairs, leaving no trace of accountability. A report from the MIT NANDA project provides numerical evidence: the successful deployment rate of self-built AI projects within the enterprise is about 33%, which is about half of the external strategic cooperation route. In my practical experience accompanying small and medium-sized enterprises to introduce AI Agent, I have seen the same resistance pattern again and again.
Mistake 3: A bunch of AI agents cannot form a brain
The third mistake is being dazzled by the SaaS vendor’s AI agent. Over the past year, I have installed more than a dozen meeting recording tools, contract generation tools, and report summary tools. Each one looks good on its own, but none of them can simultaneously connect finance, supply chain, sales, and customer service data. Kai-Fu Lee’s judgment is straightforward: software vendors are rushing to add AI agent functions to every product. They are essentially defending their old business model of selling seats, and are not really helping companies solve problems. When ten pieces of information are stacked together, they are still ten pieces.
These three mistakes, I have encountered at least one in almost every workshop during the past few years when I have been accompanying companies to introduce AI. The only difference is that Kai-Fu Lee systematized it, gave it a name, and gave a diagnostic question: Can the KPI you set for Information Chang be reflected in the next quarter’s financial report? I plan to quote this sentence directly and borrow it to ask every subsequent customer.
(opens in a new tab)
What a boss should do is not to understand technology, but to shoulder responsibility
In the second half of the book, Kai-fu Lee begins to talk about solutions: building an enterprise AI decision-making center, equipping the CEO with a “boss AI”, and using DRI (Directly Responsible Individual) to reshape the organizational structure. This part reads more abstract than the first half, but the logic is consistent: AI transformation is not a project of the information department. It is the responsibility of the CEO, and the responsibility must fall on specific people. It cannot stop at slogan-like claims such as “We are promoting AI transformation.” This is in the same direction as what I talked about in Mid-to-senior managers upgraded to Agent Commander-in-Chief: Instead of waiting for others to import AI for you, it is better to learn to command it yourself.
Who should read this book by themselves?
If you are a business owner or senior executive, this book is worth taking your time to read, rather than leaving it to the CIO to read it and report the key points to you. The reason is simple: the first mistake diagnosed in the book is that you handed over the matter to someone who should not be responsible. If you are interested, you can go directly to Books.com.tw (opens in a new tab) to buy this book and read it yourself.
After reading this, you can at least ask yourself one question: Can the KPIs I set for AI transformation be seen in the next quarter’s financial report?
Next step
Bring this topic to your team
Vista Cheng speaks and runs workshops on AI workflows, writing and content systems for companies, schools and public agencies.
