Back to Blog

The Values Moat

AILeadershipStrategyKnowledge Work

I used Claude (Anthropic) to help research and write this piece. The analysis and perspective are mine, but Claude did the heavy lifting on synthesizing industry data and making my draft readable.

The French philosopher Luc Ferry has a line about artificial intelligence in medicine that contains the whole argument: software can beat the best doctors at diagnosis, but it will never have the capacity to hold a patient’s hand.

There are two kinds of judgment hiding in that sentence. The diagnosis is a judgment of fact: what is true here, what the evidence shows, what the right answer is. The hand is a judgment of value: what matters, whose interest is served, what is owed to the person in the room. One of these is being automated faster than almost anyone forecast. The other is not. The widening gap between them is about to become the central question in how a knowledge business defends itself.

The machine beats the team

Start with the uncomfortable evidence, because it is stronger than most leaders realize.

In a 2024 randomized trial run out of Stanford and published in JAMA Network Open, fifty physicians worked through written clinical vignettes drawn from real patients, graded on the quality of their diagnostic reasoning. One group had GPT-4 alongside their usual resources; the other had the usual resources alone. The physicians with AI scored a median 76 percent. The physicians without it scored 74 percent. The tool barely moved them. Then the researchers ran GPT-4 on its own, with no physician involved. It scored 92 percent.

Read that ordering again, because it is not the one we were promised. The machine alone did not merely beat the unaided doctor. It beat the doctor-plus-machine team. The human in the loop was a drag on the result. The study’s authors are careful about why: the likely cause was that the physicians had not yet learned to wield the tool, not that their judgment was worthless. Grant the caveat. The direction still holds. The unaided expert was no longer the ceiling.

This was not a one-off. When a model called EchoNext was tested in 2025 on reading ECGs for structural heart disease, it scored 77 percent against a panel of cardiologists at 64 percent. The detail worth pausing on: when those same cardiologists were handed the model’s own assessment to work from, they climbed only to 69 percent. Given the answer, they still pulled it down.

The ranking is not universal. On open-ended, messy, real-world reasoning the gap narrows and humans still add value. But the direction of travel is unmistakable, and it does not stop at medicine. Diagnostic reasoning is a stand-in for a whole category of expert work. A great deal of investment research is pattern recognition over data; much of compliance review is classification against rules; credit analysis is probability under uncertainty. None of these reduces cleanly to its factual core, but that core is a large part of the job, and the factual core is exactly what the machine is becoming superhuman at.

What the machine cannot do

So is anything structurally human, or only tasks the machine has not reached yet? Two old ideas suggest the difference is real, not temporary.

The first is John Searle’s Chinese Room, from 1980. Imagine someone who speaks no Chinese sealed in a room with a vast rulebook. Chinese characters come in; following the rules, the person sends correct Chinese characters back out. From outside, the room appears to understand Chinese. Inside there is no understanding, only symbol manipulation. The point lands on today’s models: producing the right output is not the same as grasping what it means, and the system has no stake in what it produces. It returns the diagnosis without wanting the patient to live.

The second idea is sharper, and it carries the real weight. In the eighteenth century David Hume observed that you cannot derive an “ought” from an “is.” No accumulation of facts, however vast, contains a value. The facts can tell you what the case is. They cannot tell you what to do about it, because the doing depends on what you care about, and caring is not in the data.

This is why the machine’s growing factual mastery changes less than it appears to. A model does have values, in a sense, but it does not choose them. As Ferry puts it, their values are chosen by the programmers, who could as easily have installed different ones. The values are fitted from outside. By Hume, they cannot be read off the facts. By Searle, the system has no stake in them. The authorship of values stays human, not because the machine is weak, but because that authorship is a different kind of thing from the facts the machine commands.

The values moat

Now bring this into a firm.

The classic moat in a knowledge business is proprietary expertise: what our people know that the competition does not. That moat is draining. When everyone can rent the same frontier intelligence, the knowing-the-answer layer levels out. What your analysts know, the market can increasingly know too, at the price of a subscription.

The moat does not disappear. It relocates. The team still matters, but decreasingly for what it knows, because the machine knows it as well. It matters for what it values, and for how reliably those values show up in what the client receives. Why values? Because a model cannot author its own: Hume says they cannot come from the facts, and Ferry says a model’s values are chosen by its makers. A firm is the maker of its own values. Its culture is a chosen, human-authored value system, sitting in precisely the layer the machine cannot occupy.

There is a harder edge to this than culture. The firm is also the party that can be held to account: licensed, examined, liable, able to be sued when the judgment goes wrong. The machine can be none of those things, because it has no stake in the outcome. Accountability is what turns a value from a sentiment into something enforceable, and it is the plainest answer to the question the rented model raises, which is why the client keeps paying the firm at all. Someone has to own the consequence, and only a human institution can.

Imagine a firm built on two such values: service and learning. Watch how cleanly they map onto the two halves of the case.

Learning is the meta-competence that compounds when facts are free. Not knowing the answer, which is now cheap, but the institutional habit of getting better at asking, and at wielding the tools faster than they change. It is what keeps a firm ahead on the factual layer even as that layer commoditizes. It is learning to learn, raised to the scale of an institution.

Service is the judgment of value itself. Whose interest the analysis is bent toward. The fiduciary form of holding the hand: what is good for the client, beyond what the model was optimizing for. This is the layer the machine cannot enter, because entering it means caring about an outcome, and the machine does not care.

The work that remains

Across this series the pattern has held: AI accelerates execution and, in doing so, exposes the quality of everything underneath it. This is where that arc ends. The machine will keep getting better at facts. That much is settled. What it will not do is decide what the facts are for.

Three things follow for anyone watching the analytical layer commoditize. What your people know is no longer the moat; assume the competition can rent it. The habit of learning faster than the tools change is the edge that renews itself. And the values that decide whose interest the intelligence serves, backed by an institution willing to be accountable for the outcome, are the part that cannot be copied, rented, or automated, so long as they are real enough to cost you something.

The machine can hold the answer. It cannot hold the hand. That remains the work, and the moat.

References

Share this article