When AI Gets Smarter, What Should Founders Actually Worry About？

Today at Stanford, I listened to Carina Hong — MIT graduate in mathematics and physics, nine published papers as an undergraduate, winner of the Morgan Prize, the highest honor for undergraduate mathematics research, a Rhodes Scholar at Oxford, and now the founder and CEO of Axiom, an AI-for-mathematics company.

What struck me first wasn't her résumé. It was her style. She never rushes to answer a question. There is an optimistic restraint in the way she thinks — it reminded me of something I learned at Stanford GSB: don't rush to answer every question. Learn to hack the question first.

Her talk was about how AI can now organize more than ten thousand agents — dividing labor, collaborating, tackling mathematical problems of staggering complexity, and verifying results through formal methods. Impressive. But as a former CEO, and now as a coach to founders, what stayed with me were the implications for how we run companies.

If the Direction Is Wrong, the More Efficient You Are, the More Dangerous It Becomes

Formal verification can prove that a result follows the rules. It cannot prove that the rules were worth following — that the original problem was defined correctly.

This is the question every company should ask before adopting AI: what problem are we actually solving? And is it worth solving?

A company can have an excellent execution team, mature technology, and the most advanced AI. But if the leader defines success incorrectly at the start, all of that execution simply makes the company faster at pursuing the wrong goal.

Looking back on my years as CEO, I've come to see that many organizational problems — the ones that look like execution, communication, or talent problems on the surface — are really failures to align on something deeper: what our goals are, where the boundaries lie, and who owns what. AI will amplify this misalignment, not fix it.

So here is the question I would ask every founder using AI: do you need AI to execute faster right now — or do you need someone to help you think the problem through again, from the start?

Bill Gates once said that everyone needs a coach. I heard that line echo in the room.

Domain Expertise Isn't Losing Value. It's Worth More — in a Different Way

Many professionals are anxious right now: if AI can write code, do analysis, even prove theorems, is my years of accumulated experience still worth anything?

Here is what I took away: someone who truly understands a domain may only need to set the right constraints, structure the knowledge, and decompose the task — and AI avoids an enormous amount of wasted computation. Deep domain knowledge can improve engineering efficiency by a hundredfold or more.

The competition ahead isn't just about AI capability. It's about return on effort.

In the future, professional value may no longer show up mainly in how much work you personally complete. It will show up in whether you can make the whole system make fewer mistakes and take fewer detours. Experience isn't for racing AI on speed. It's for knowing which races aren't worth running.

More Agents Don't Mean More Productivity

How do you divide the work? How do agents collaborate? How do you control the cost? It sounds exactly like managing a team.

If the organizational design is poor, adding people just creates more internal friction. The same is true for AI.

When AI Produces Results We Can't Understand, Will the CEO Still Dare to Sign Off?

The liveliest philosophical moment of the session: is mathematics more like chess, or more like cancer research? The deeper question underneath — do we care more about understanding the process, or about the final value?

And when AI becomes a black box, who bears the responsibility?

In management, a CEO can never fully understand every detail of how each person does their work. We have to build trust — in professionals, in team processes, in organizational systems. But there is a critical difference: when we delegate to people, we can question their accountability, their motives, their values. When we delegate to AI, we must be far more deliberate about boundaries, verification mechanisms, and accountability.

This isn't about trusting or distrusting AI. It's about deciding what we are willing to hand to machines — and which decisions must ultimately be owned by a human being.

After years as a CEO, I know that the hardest part of running a business was never execution. It was trade-offs.

AI can help us do things right. But what is worth doing, why we are doing it, and who bears the cost — that remains the founder's responsibility.

And that, to me, is the inner capacity leaders most need to build in the AI era: the strength to hold complexity, make clear-eyed trade-offs, and stay human while doing it.