Category definition

What Is Imprinted AI?

If a system can write in your voice but cannot make the call you would have made, it is not an Imprinted AI. Voice is the easy half. Voice is recoverable from anything you ever published. The call is not.

Imprinted AI compared to what it gets confused with

Model memory is not the enemy here. It is necessary. You cannot build an Imprinted AI without it. The point is that almost everyone stopped there, and memory alone cannot produce judgment, because knowing what happened is not the same as knowing how you would have ruled on it.

Six more approaches sit close to this category and each one misses it for a different reason. Those are worked through one at a time: vs an AI clone · vs AI memory · vs fine-tuning · vs RAG · vs a Custom GPT · vs a digital twin · vs an AI agent · an Imprint vs an SOP. The comparison index says which axis each one turns on.

Several terms on this page carry a precise meaning that is worth having separately: Imprint, judgment layer, captured, not guessed, the Five Levels of Knowing, founder judgment, articulation density. All fifteen are in the glossary.

If you would rather have the argument short than long, the thirty-five lines are this whole case compressed to the sentences that carry it, and any of them may be quoted with attribution. The term was coined by Rich Schefren in 2026, and the record behind that is on a second domain, dated with its sources named.

Why "captured, not guessed" is the whole distinction

There are only two ways a machine can come to hold a picture of how you think.

Guessed

The system reads your finished work and reasons backwards to the person who made it. This is what a clone does, and it is what memory does when it tries to infer preference from behavior.

It fails for a structural reason. You cannot recover a decision from its output. The artifact shows you what was chosen. It does not show what was rejected, what was nearly chosen, what rule the choice was serving, or which of ten plausible reasons was the real one. A machine reading the artifact has to invent that part. What it invents sounds exactly like the part it got right, because both are written in your voice.

That is the failure mode people miss. Guesses do not sit quietly in a corner. They mix in. Later inferences get built on earlier inferences. Eventually the picture is a blend of things you decided and things the machine decided you would decide, and nothing on the surface tells them apart. The layer has to be gutted and rebuilt.

Captured

You make a real call. The call itself is what goes in. A verdict, a correction, a rejection, a contrast between two options with the reason attached. Nothing enters that you did not decide, so there is never a false memory throwing off everything downstream.

Captured is slower. Captured is also the only version that is still true in a year.

How judgment actually gets captured

Not by writing a document about yourself. People are famously bad at that. The knowing that matters is the part you cannot put into words, which is exactly the part a questionnaire cannot reach. It gets captured the way it always moved: one real decision at a time.

1. Do the work you were already doing. The capture attaches to real output with real stakes. Work invented to train the system produces judgment about invented work.
2. Render a verdict on what comes back. Not "make it better." A ruling. This is wrong, this is close, this one ships.
3. Attach the reason to the verdict. The reason is the asset. The verdict alone is a data point. The verdict plus why is a rule that generalizes.
4. Capture the rejections, not only the approvals. What you refuse defines the standard more sharply than what you accept. Most systems throw the rejections away.
5. Keep the record outside the conversation. A correction that lives in a chat window dies with the chat window. This is why the same fix gets made forever.
6. Propagate a changed belief. When one position moves, everything that rested on it moves. If those downstream positions are not updated, the picture goes stale while still looking current.
7. Let it compound. The value is not in any single capture. It is in the accumulation, and in the fact that nothing in the pile is invented.

What Imprinted AI is not

It is not a digital clone of you. A clone is built to answer in your place. An Imprinted AI is built to work at your standard. Different job, different construction, different asset.

It is not "AI with a long memory." Memory holds the record. The Imprint holds the ruling on the record. A system can remember every meeting you ever had and still have no idea which of the three options on the table you would have killed.

It is not a personality profile, a tone-of-voice document, or a custom instruction block. Those describe the surface. They are useful. They are not judgment.

It is not automation. Automation moves work you have already figured out. An Imprinted AI carries the figuring out.

It is not something you buy filled in. It ships empty. It has to be filled by you, from your decisions. That is not a limitation of the category. It is the definition of it. The moment it can be bought pre-filled, whatever is in it is not yours, and the advantage is gone.

Why the category exists now

For as long as there has been business, the deepest knowledge a person held could travel only one way: another human stood next to them for years and slowly absorbed it. It could not be written down completely, because the person could not say it completely. Michael Polanyi named the problem in The Tacit Dimension in 1966, in a sentence that has not been improved on: "We can know more than we can tell."

That constraint set the ceiling on every business built around one person's judgment. The judgment could not leave the person. So it either moved at the speed of an apprenticeship or it died with them.

AI is the first technology that can take the transfer without the decade. Not by reading what the person wrote. By standing beside them while they decide, for more hours than any human apprentice ever could, and keeping every ruling.

The reason to do it now rather than later is unglamorous. A competitor can install next year's model next year. They cannot go back and capture the judgment you did not preserve this year. The model is rentable. The year is not.

The sequence also matters, and the market has already produced the counterexample. When a company automates the work before it captures the judgment that made the work good, the output gets faster and worse at the same time, and the fix costs more than the automation saved. Klarna ran that experiment in public with its customer service operation and had to reverse it. The model did not fail. The order of operations did.

The mechanism is open source. You can read it.

Two of the pieces this argument rests on are public repositories, not claims.

Atlas — belief revision

Open-source, local-first cognitive memory with AGM-compliant belief revision, all 49 postulates. When a fact changes, the beliefs downstream of it are automatically re-evaluated rather than merely flagged. That is the propagation problem named above, implemented and given away. Public since April 2026.

Imprint — judgment capture

Local-first, provenance-preserving judgment capture and retrieval. Provenance is the entire point: every entry carries where it came from, which is what makes captured judgment distinguishable from inferred judgment forever after.

Note what is open and what is not. The tooling is public. The contents are not. Anyone can run the same capture mechanism; nobody can obtain what a particular person's decisions put inside it. That split is not a business model bolted onto the idea. It is the idea: the system ships empty, and the only thing that gives it value is judgment that was actually exercised by the person who owns it.

Where the term comes from

Rich Schefren has been building and operating an Imprinted AI on his own business, daily, for over a year, before the category had a name. The capture system that produces it, the Imprint Growth System, runs in production and has been packaged and handed to clients.

The naming is not a branding exercise. In 2006 he published the Internet Business Manifesto, which crossed five million downloads and named the distinction between working on a business and working in one for a generation of online operators. The 2026 argument is the sequel to that one, and it needed a name for the same reason the first one did: there was nothing to point at.

Two positions already occupy the conversation around it. One says judgment is the human moat and stops there, with nothing built. The other says clone yourself, and builds a replica of your persona on someone else's platform. Neither occupies the space between them, which is the only space where the thing is useful: your judgment, captured, running inside the systems you actually operate your business with.

That space is what Imprinted AI names.

Frequently asked

There is nothing to buy on this page

The full argument runs about 23,000 words and it is free. No email, no signup, no gate.

Read The A.I. Business Manifesto

If it is right, the useful thing you can do is pass it on.

Rich Schefren is the founder of Strategic Profits. He published the Internet Business Manifesto in 2006. Clients he has worked with have been credited with more than $15 billion in revenue.

Last updated: 28 July 2026