If you had to describe the result of your company’s AI program on your next earnings call—one sentence, of the kind a shareholder would care about—what would that sentence be?

Most chief executives I meet cannot name it. They can produce activity: pilots run, licenses purchased, employees trained, a partnership announced. None of that is a result. Survey after survey finds that most corporate AI programs deliver disappointing measurable returns.

In my view, this is not because of the models, the data, or the talent. It is because of where companies are pointing the AI itself. Today’s systems can reason across an entire business, and we have assigned them to handling tickets and summarizing meetings. This is like hiring a business operations genius and putting him to work as an assistant. The place where AI can move the most money is not the call center. It is the corner office.

In fairness to CEOs, the cards are often stacked against them. A company is an information system, and the quality of information degrades at every level it passes through. Going up the chain, uncomfortable findings and early warning signs get discarded; what often arrives is a clean, but incomplete, summary. Going downwards, management directives are often diluted, re-interpreted, re-purposed, or quietly “pocket-vetoed”—nodded through in the meeting and left to die. The result? A company of 10,000 ends up knowing less than a company of 10.

A good chief of staff can help compensate for these challenges. But none can read every document the company produces, sit in every meeting, and hear every customer call.

Now, AI can expose and accelerate radical visibility in organizations at a depth and speed that far surpass any human capacity and any business intelligence system. Every company keeps two sets of records. One is official: quarterly reports, project updates, performance reviews, polished at each level until it is clean and coherent. The other is what actually happened: meeting transcripts, chat threads, edit trails, sales notes, code check-ins. Managers have always worked from the first, because the second was too numerous and scattered to read.

No longer. Every company writes down the most critical issues, the hottest debates, and the truth about itself every day. Almost none of them read it. A CEO-level agentic system can wake up that “dormant data.”So at my company I asked my team to build management agents for me: to sit in meetings, track commitments, and read the data underneath what I was being told. The more I used them, the more we added. People called it “Kai-Fu’s Boss AI,” and the name stuck. We eventually turned it into a product.

What did my boss AI find in its first weeks? A priority project reported on track every week, with no code committed in five. A customer marked highly likely to renew, while the account manager had not called in two months. A draft contract nobody could have made profitable, as reading relevant sales notes, the signed terms, and product requirement documents together showed free customization promised in a room but never written down.

Those are not minor administrative matters. They are a delivery slip caught early, a renewal saved, a margin leak stopped before it had to be honored. That is the connection to the bottom line every AI pilot has missed: not a productivity percentage, but a number that moved because AI caught it in time.

AI’s second contribution is to surface hidden good news that might have been under the radar. Chief among these is talent and identifying star performers.

Last year I ran our performance review differently: no resumes, no KPI decks, no department heads recommending their own. We agreed first on what “engineering excellence” means when a machine can now write most of the code, then scored a year of repository history against it. Several of the strongest names were people I could not have picked out of a meeting. One vice president wrote close to 300,000 lines with AI in a month, against his own ceiling of 30,000 20 years ago, when he was the most prolific coder I knew.

He is not typing faster; he is directing AI. And the reverse: once the machine does the status-collecting and information-passing, the employees whose only output was status-collecting and information-passing are obsolete.

This is what an AI-native chief executive is: not someone who just signs off on an AI budget, but someone who puts his own meetings, calendar, and blind spots under the machine’s scrutiny first. I regularly ask my agent what I got wrong in a meeting, whether my time went where I said my priorities were, and what people are afraid to tell me. Subordinates answer those questions politely. The machine can tell me the blunt truth.

This is how a company can be truly transformed by AI. The chief executive asks sharper questions, and everyone in the room tends to follow. Often, middle managers resist AI because they fear it will put their teams at risk. In fact, I would argue that employees resist AI not because they doubt it works but because they suspect it does.

So: no more cosmetics. We should use AI to address businesses’ hardest problems in the building: the ones on the CEO’s desk. Then, the words the CEO proudly proclaims on the next earnings call will have three parts: a business number that moved, the person who moved it, and the compute that did the work.