An AI clone can take every call, but unlimited access creates value only when the original person has earned attention through years of useful work. The clone scales accumulated judgment; it cannot manufacture the experience behind it.
A founder watches the professor’s avatar answer its hundredth question. The voice remains patient. Every visitor gets time. Nobody waits for office hours, crosses a border, or competes for a place in the room.
Then the founder sees the harder truth. The technology multiplied access in weeks. The professor spent decades becoming worth asking.
What the recording preserved
In 1964, Richard Feynman delivered a series of lectures at Cornell University. He stood before an audience and worked through questions about physical law, uncertainty, symmetry and scientific reasoning.
The BBC recorded the lectures. That decision separated Feynman’s explanations from the limits of the room and the calendar. People who never visited Cornell, including people born years later, could watch him reason through the material.
The recordings did not make Feynman credible. By 1964, he had already contributed to quantum electrodynamics, worked at Los Alamos during the Manhattan Project and developed a distinctive way of explaining difficult physics. He received the Nobel Prize in Physics the following year with Julian Schwinger and Shin’ichirō Tomonaga.
Decades later, Bill Gates acquired the rights to the recordings and made them available online through Microsoft’s Project Tuva. Microsoft Research documents the project and its purpose: widening access to Feynman’s Cornell lectures.
The camera expanded the audience. Feynman’s previous work gave the audience a reason to stay.
That is the useful analogy for AI clones. A convincing voice, a familiar face and instant answers can remove the scheduling constraint around a person. They do not remove the knowledge constraint inside that person.
Access exposes the quality of the source
Founders often approach cloning as a capacity problem.
There are too many sales calls. Too many questions from customers. Too many junior team members asking how to make decisions. Too many people requesting advice after a demo in Accra, a conference in Berlin, or an introduction from someone in New York.
A clone appears to solve the arithmetic. One person can serve a hundred conversations at once.
But the hundredth answer reveals what the first answer could hide. Does this person have a clear position? Can they explain the tradeoff between hiring an engineer and protecting runway? Do they know when an AI demo is evidence of demand and when it is theatre? Have they recorded enough real decisions for the system to distinguish a useful exception from a convenient excuse?
Scale makes weak source material more visible.
I have seen the same pattern in product work. Teams rush to automate a process before agreeing on how the decision should be made. The software then produces inconsistent answers faster. An AI agent can repeat the confusion across every customer conversation before lunch.
That is why the first asset is the decision record. Write down the calls that went well, the ones that failed, the evidence used and the conditions that would have changed the choice. If your specialist sleeps while an agent proposes a code change, the important question is still who has earned the authority to approve it. I explored that boundary in Should I Approve an AI Agent’s Code Change While the Specialist Sleeps?.
Build the professor before the clone
A useful clone needs more than a library of polished content. It needs traces of judgment.
Start with ten decisions that carried a real cost. Use decisions where the answer could have gone either way: accepting an overseas contract, delaying a release, rejecting a feature request, narrowing the customer segment, or choosing three investor calls from 34 introductions. Record the context available at the time, not the tidy explanation produced after the result.
Then test whether another person can use those records to reach a defensible answer. If a capable colleague cannot tell which facts mattered, an AI system will probably imitate the language while missing the reasoning.
This is also where founders should resist the temptation to sound certain. A credible knowledge base contains boundaries. It says when the available evidence is weak, when local market conditions matter and when the question needs a human who can accept responsibility for the outcome.
That restraint increases trust. A clone that answers everything confidently becomes dangerous precisely when access feels most convenient.
Make one answer worth repeating
Feynman’s lectures survived because the underlying explanations could bear repeated attention. Project Tuva widened their reach, but it did not rescue empty material.
The practical order for a founder is similar. First, make one answer useful. Ground it in a shipped product, a customer decision or a constraint that cost you something. Explain what you knew, what remained uncertain and why you chose one path.
Next, test that answer with the people who face the same decision. Watch where they misunderstand it. Revise the reasoning, not merely the phrasing.
Only then should you make it available at all hours.
The founder watching the hundredth conversation should be measuring more than call volume. The better question is whether answer number one hundred still carries the judgment that made answer number one worth hearing.
Comments
No comments yet.