The Same Correction, Every Quarter
By Daniel Koo - Head Of Product
08/07/2026
There is a moment that shows up in almost every team running an AI specialist for more than a couple of quarters. Someone goes to correct something, and realizes they have corrected it before.
The exception that was explained in March gets explained again in June. The context somebody wrote out carefully at the start of a project gets written out again for the next project, in the same words, to a system that treats it as new. The team is putting in real work and none of it is stacking up.
That pattern gets read as a discipline problem. It is usually not. It is a design problem, and the design question underneath it is what the specialist is allowed to keep.
Most of what a specialist knows was handed to it
Think about what a person knows on their first day in a role, before they have done anything.
They know the company handbook. They know their department's way of working, which is usually more specific and occasionally contradicts the handbook. If the team is big enough to have its own norms, they pick those up in the first week. All of that arrives without them earning it. Someone wrote it down, and joining the role gave them access to it.
An AI specialist works the same way, and this part is generally handled well. Organizations are good at authoring context. They write the policies, the standards, the definitions of terms the business uses in its own particular way. A specialist that inherits those layers starts out knowing what the organization knows.
What it does not have on day one is anything from doing the job.
The layer nobody authored
The fourth layer is the one a person builds themselves. It is the notes they keep, the pattern they noticed in the third week that nobody warned them about, the client who always asks for the same thing, the shortcut that turned out not to work.
Nobody wrote that down for them. They earned it by doing the work, and it is usually the layer that separates someone who has been in a role for two years from someone who started last month. Both of them read the same handbook.
For a specialist, this is the layer that decides whether the ongoing work compounds or resets. The three inherited layers are static until someone updates them. The fourth is the only one that grows on its own, from the actual work, in the specific situations the specialist has actually met.
When teams find themselves making the same correction twice, this is almost always what is missing. The correction went in as a one-time fix rather than as something the specialist keeps. The work of running the specialist is real, and it evaporates.
Earned knowledge and authored knowledge are not the same thing
It is tempting to solve this by letting whatever the specialist learns flow back into the shared layers, so everyone benefits. That instinct is right about the goal and wrong about the mechanism.
Authored knowledge has been reviewed. Someone with standing decided it was true, decided it applied broadly, and put their name on it. Earned knowledge has not been through any of that. It is a working conclusion drawn from a limited set of situations, and it might be a genuine insight or it might be a pattern that held three times by coincidence.
Both are valuable. They are not interchangeable, and the difference between an organization that gets smarter and one that quietly accumulates confident errors is whether anything sits between the two.
The workable shape is a proposal. A specialist that notices something durable can put it forward for the shared layer, and a person who owns that layer decides whether it belongs there. Slower than automatic promotion, and it is the difference between a body of knowledge and a pile of assertions.
Reaching sideways is a decision, not a default
One more distinction matters, and it is the one most likely to cause a problem later.
Inheriting downward is automatic. A specialist reads its own line, and everything above it in that line, because that is what its role entitles it to. Reaching sideways into another part of the business is different. It should be granted deliberately, scoped to the work that requires it, and it should end when that work ends.
This sounds like governance overhead. In practice it is the same judgment any organization already makes about people, which is that access follows the work rather than curiosity. The reason to be explicit about it with specialists is that a specialist will not feel the awkwardness a person would feel reading something clearly not meant for them.
What this is really about
The last piece in this series made the case that standing up a specialist is only half the work, and that running it is the other half. This is the part of the machinery that decides whether that other half accumulates into anything.
A specialist with no memory of its own can still be useful. It will just be as new in its sixth month as it was in its first, and every improvement its team makes will have to be made again. That is a lot of effort to spend on something that keeps arriving back where it started.
The question worth asking about any specialist is not only what it knows. It is what it is allowed to keep, and who decides when something it learned becomes something everyone knows.
Head Of Product, Arkane Digital
Daniel leads product strategy and system design for AI platforms, focusing on turning complex capabilities into structured, usable systems that support real business operations.