Comparison
Imprinted AI vs an AI Agent
This is the one comparison on the site where the two things are not competing at all. They sit on different axes. Reading them as alternatives is what produces the most expensive mistake in the list.
An AI agent is capability. It plans, uses tools, takes multi-step action and keeps going without you standing over it. The question it answers is what the system can do unsupervised.
An Imprint is authority. It holds what you have ruled, so the question it answers is what the system is entitled to conclude in your name.
It does not supply the thing it multiplies.
An agent inherits a decision procedure. It does not have one.
Every autonomous run makes decisions. It has to. Something has to choose which option to pursue, when the output is good enough, what to do about the case nobody anticipated.
The agent does not generate that procedure. It applies whichever one is available to it.
If your captured judgment is available, it applies your standard, and more agency gives you more of your standard operating in more places.
If nothing is available, it applies the model's defaults, which are a reasonable consensus assembled from an enormous amount of other people's work. Competent. Bland. Not yours. And more agency then gives you more of that, further from your supervision, with your name on the result.
So the capability upgrade does not close a judgment gap. It widens the consequences of one. This is the counterintuitive part and it is why the problem gets worse as the tooling gets better.
The two axes
| Low agency: you are in the loop | High agency: it runs alone | |
|---|---|---|
| Judgment guessed or absent | Workable. You are supplying the judgment yourself, in real time, and you barely notice you are doing it. | Where the damage happens. Confident output at volume, built on a stranger's defaults, discovered late. |
| Judgment captured | Useful but under-used. The system holds your standard and you are still doing the deciding anyway. | The point. Your standard applied to work you are not present for, with the ruling traceable to where you set it. |
Notice which cell most teams are trying to move into, and which direction they approach it from. Almost everyone travels along the top row, buying agency, and the top-right cell is the one that costs money.
The eighty percent problem is misdiagnosed
The common experience: an agent gets most of the way through a real piece of work and stalls. The remaining part comes back to you.
The usual reading is that the agent was not capable enough, so the response is to buy a better one. Then the same thing happens at a slightly different point.
Look at what the residual actually consists of. Which of these two defensible options. Whether this clears the bar. Whether the exception applies here. Whether to raise the objection or let it go.
None of that is capability. It is a call, and it comes back to you because you are the only place the call exists. A more capable agent arrives at the same wall with more momentum.
The thing that moves that wall is not a better agent. It is having ruled on cases like this before, in a form the system can apply.
The sequence, and the public example
Order matters here more than in any other comparison in this set.
Automate a process before capturing the judgment that made it good, and two things happen at once. Throughput goes up. Quality goes down. Both are visible in the numbers immediately, and the second one keeps getting attributed to model quality.
Klarna ran this in public. It automated a large share of its customer service operation, quality fell, and it reversed course and brought people back. The model was not the failure. The order of operations was. The judgment that made the service good had never been captured, so what got automated was the shape of the work with the substance left out.
The repair costs more than the automation saved, every time, because you are now undoing at volume.
The quiet version: decisions you never authorised
The dramatic failure is easy to spot. The expensive one is not.
A long autonomous run makes dozens of small calls nobody asked about. Each is defensible in isolation. Read the final output and it is coherent and confident, and nothing in it marks which parts came from you and which were decided on your behalf because something had to be.
You end up responsible for conclusions you never reached, and you cannot find them, because they do not look different from the parts you did reach.
Provenance is what makes them findable. When every applied ruling knows where it came from, the calls with no ruling behind them can be surfaced instead of blending in.
When an agent alone is the right answer
When the decision procedure is already settled and the outcome is checkable.
Deterministic multi-step tasks. Work with an external verifier, where the result is right or wrong on inspection rather than a matter of taste. Reversible actions. Anything where the standard is a specification rather than a judgment call.
For that class, agency is the whole win and there is no judgment layer needed, because there is no judgment involved.
The test is simple: could two capable people who both follow the spec produce materially different work, and would you prefer one of them? If yes, judgment is involved, and agency without it will pick for you.
Frequently asked
What is the difference between an AI agent and an Imprinted AI?
They answer different questions. An agent answers what this system can do without you, which is capability. An Imprint answers what it is allowed to conclude in your name, which is authority. They are not alternatives and they do not substitute for each other. An agent with no captured judgment still acts, it just acts on the model's defaults instead of yours.
Why do my AI agents get most of the way there and then stall?
Because the part that is left is usually not capability-shaped. The remaining work is a call: which of two defensible options, whether this clears your bar, whether the exception applies here. Buying a more capable agent does not supply that, which is why teams can upgrade repeatedly and land in the same place with a faster version of the same stall.
Does a better agent reduce the need for captured judgment?
It increases it. Agency is a multiplier on whatever decision procedure is underneath. A more capable agent applies that procedure to more work, further from your supervision, with less chance of interception. If the procedure is your captured judgment, you get more of your standard. If it is the model's defaults, you get more of a stranger's taste with your name attached.
Should I automate first or capture judgment first?
Capture first, on the work you are about to automate. Automating a process before the judgment that made it good has been captured produces output that is faster and worse at the same time, and the repair generally costs more than the automation saved. There is a well-known public instance: Klarna automated a large part of its customer service, saw quality fall, and reversed the decision. The model did not fail there. The order of operations did.
What are decisions I never authorised?
Small calls an autonomous run makes on your behalf because something had to be decided and nobody was there. Each one is individually defensible. The problem is the compound. Across a long run, dozens of them stack into an outcome that is coherent, confident and not the one you would have arrived at, and nothing in the output marks which parts were yours.
Where this sits
Read the category definition at What is Imprinted AI. The provenance mechanism referenced above is public: Imprint preserves the origin of every captured ruling, which is what lets an unauthorised call be distinguished from an authorised one after the fact.
Related comparisons
- Imprinted AI vs a Custom GPT. Configuration as a way to set an agent's standard, and where it runs out.
- Imprint vs an SOP. This gap in the era before anyone said agent.
- Imprinted AI vs AI memory. Why continuity alone is not enough to run unsupervised.
The full argument
This page is one distinction pulled out of a longer case. The whole thing runs about 23,000 words, it is free, and there is no email gate on it.
Read The A.I. Business Manifesto
Nothing on this page is for sale. If the distinction is useful, the useful thing to do with it is pass it on.
Last updated: 28 July 2026