We have all seen it happen: someone who has stood out elsewhere enters the wrong system and becomes strangely ordinary. Someone easy to overlook enters another and seems to acquire range, confidence, judgment — almost a different person. We explain the change in the language closest at hand: leadership, culture, opportunity, fit. All of those words are true. None quite locates the mechanism.
Artificial intelligence has made that mechanism harder to ignore.
The same model, dropped into two different systems, behaves like two different machines. Give one the right context, tools it can actually use, a way to inspect what happened after it acted, and boundaries clear enough to move without guessing; place the other behind the same interface with those things missing, and the difference can look like intelligence. The model did not change. Everything around it did.
For a while, the model absorbed most of the attention because it was the visible source of the result. But as AI systems moved from demonstrations into real work, the engineering problem widened to include what the model could know, what it could reach, where it could act, and what happened after it did. What had looked like plumbing around intelligence became part of the machinery required to make intelligence useful. The model still mattered; the boundary of the system had simply moved out around it.
That shift stayed with me because when the intelligence is human, we draw the boundary of the system much closer to the person — close enough that the person is the system, and everything around them is circumstance. A person arrives with context already inside them: memory, experience, instinct, years of absorbing what was never written down. The surrounding system does not have to supply it at the moment of use. So when a system leaves something out, a person tends to close the gap without anyone noticing there was one.
Much of what a human system can leave to interpretation, an AI system has to settle by design: what reaches the agent, what it may touch, when uncertainty requires escalation, what returns after action, and what counts as enough. As those problems become distinct enough to engineer, they acquire names.
And naming changes what can be seen.
A name does not separate a thing in the system. It separates it in thought. A pattern that had been carried implicitly can now be examined on its own, compared with alternatives, questioned, constrained, redesigned. The system remains entangled; what has changed is our ability to reason about it. That is why the vocabulary emerging around AI matters beyond AI itself.
The two systems need not share an architecture. Distinctions made explicit in one can still expose dependencies the other has long been able to leave implicit.
Perhaps capability was never located where it appeared to be.
The capability you can call
An organization can hire judgment it has no way to use.
Put the same engineer inside two different decision systems. In one, a consequential but reversible choice sits close to the work: enough context to make it, enough authority to act, and a quick way to learn whether it was right. In another, the same choice travels upward, the boundary of authority is less clear, and being wrong is harder to recover from. The same person may look decisive in one and cautious in the other. Caution is not a trait here. It is an accurate reading of the second system.
There is also rarely a single system experienced equally, which is part of what makes an organization difficult to read from the people inside it. Two engineers in the same company can receive different amounts of context, trust and opportunity. One is brought into an important decision; the context sharpens their judgment, the judgment earns more trust, and the trust brings them into the next decision. The other is not in that loop. The difference begins to compound, though no one decided that it should. In time, a difference the system helped produce becomes hard to tell from a difference in the people themselves. The system has not merely revealed capability. It has helped decide which capability became visible.
In both cases, what differed was not necessarily the person but what they could reach: the context that arrived, the authority that was clear, the decision they were brought into. AI engineering has been forced to treat capability the same way — as something that has to be made reachable. A tool can exist and still be useless if the model cannot call it. Knowledge can exist and still be absent if it cannot enter the context at the right moment. Authority can exist somewhere in the system, but if the actor has no route to whoever holds it, the work still stalls. The model has not become less intelligent. More of its intelligence has simply become unreachable.
Organizations have the same problem. An understanding of a customer problem can sit outside the decision it should have informed. A team can own an outcome without owning the decision that determines it. People compensate — they find the meeting, ask the person, carry the missing context across the boundary themselves — and the organization keeps working, which makes the architecture look better than it is.
Software offers a useful way to say it. There is a difference between a function that exists and a function the system can call, and organizations hold capability in the same two states. Expertise on the payroll is not the same thing as expertise available at the point of decision.
A system does not get the capability it contains. It gets the capability it can call.
Organizational design does more than arrange people: it determines what they can know, reach, decide and change. The org chart describes where people sit. The system opens and closes the paths their capability can travel.
Machines leave less room for implication. A model can infer a great deal about the unwritten organization — who really decides, which rule bends, where the reliable knowledge lives, when an exception is legitimate — but only from what was recorded somewhere and put within its reach. People absorb the same things by being present: we observe, infer and adapt until the operating system nobody wrote down becomes familiar enough to navigate. That adaptability has long allowed — and still allows — much of an organization’s operating logic to remain distributed across tools, rituals, hierarchy, culture and memory, without ever needing to be described as an architecture.
It has left us with a curious asymmetry. We are becoming precise about the architecture that makes artificial intelligence useful, and still describe the architecture around human intelligence with words like process, culture and management. Those words are not wrong; they carry real meaning. But they name the territory, not the mechanism.
What we never had to say out loud
A rule can look complete until a machine has to follow it. A team escalates a change it considers high risk. Between people, much of what high risk means can remain implicit in judgment. Ask an AI agent to act on that rule and the judgment behind it has to become available in a form the system can use: what counts as high risk, which exceptions matter, who decides when nobody is sure. The machine did not create the ambiguity. It exposed how much of the rule’s meaning had been living in interpretation.
That interpretation does not necessarily live in one place. It can remain stable enough across a team to feel like the rule itself. Then someone leaves. Or the team doubles. Or a new manager reads high risk differently, and reasonably. Nothing changes on paper, so the change may leave no formal trace. The rule still exists. Its operating meaning has moved.
People are not merely navigating the organization. Part of it runs inside them.
The same hidden execution runs through ordinary organizational language. Communication is a good example. A company can communicate constantly and still leave a decision without the context it needs: the document exists, the meeting happened, the ticket is current, and the person doing the work still cannot see enough of the problem to make the right choices. Sometimes the needed information was never produced. Sometimes it was produced and never traveled. Sometimes it arrived after the decision. Sometimes it arrived in time but without the authority to act on it — which is no longer a communication problem, though it may still be counted as one. One symptom, several mechanisms.
One part of that execution shows up the moment an agent is involved: the information people recover for themselves. A person can notice what a ticket leaves out, remember a conversation, ask someone nearby, or infer what was never written down. An agent can reason too, but only from what the surrounding system makes reachable. Anthropic describes context engineering as bringing the most useful information into reach while it can still shape the result. If what matters never reaches the system, rewriting the prompt cannot recover it.
Information is only one case. Once these dependencies have to be engineered, vagueness starts to cost. In AI system design, concerns such as context, harness, loop and governance are increasingly named and built separately, often represented as layers. The distinctions belong more to the design model than to the system itself; relationships, constraints and feedback cut across them. They help isolate design problems without pretending the system itself is divided along the same lines.
What transfers is not the taxonomy, but the separation it makes possible. Distinctions made necessary in one system give us a higher-resolution view of another. AI engineering gives us a vocabulary precise enough to hold apart, in thought, what organizational systems carry together.
The first thing a finer view shows is how much it does not cover. Much of an organization’s intelligence still lives in what people do without any formal channel: reading an exception, sensing that the stated problem is not the real one, negotiating what a word means, carrying trust across a boundary nobody drew. AI engineering is teaching us how much scaffolding intelligence can require. The harder question is when scaffolding stops supporting judgment and starts replacing it.
The next question, then, is not what else can be made explicit. It is what should remain in the person.
Where the analysis runs out
Early in my career, during my first week at a startup, I was called into the CEO’s office with the CTO and my director. The meeting lasted perhaps fifteen minutes. Build the reporting section for the new video-advertising platform. Make it look like Google Analytics. Aim for queries under ten seconds. I had spent the previous year on a project alongside Google Analytics engineers, so the reference point was familiar. The domain was not.
The brief was broad, but for a startup, hardly unusual. Product work often begins before the answers arrive. Advertisers, publishers and content owners could look at much of the same data and want different things from it. Which information deserved attention, what should be left out, how much a single view could carry before it stopped being readable — those were questions the work itself had to answer. I formed a point of view, tested it as I learned the domain, and changed it when the evidence changed. As confidence in the work grew, so did my influence over what belonged on the roadmap and when the product was ready to ship.
Something else was happening alongside it. Knowledge that already existed inside the company did not always find its way to the project. Data definitions, domain meaning, even who understood particular parts of reporting had to be pieced together through conversation. Months in, I met someone elsewhere in the organization with deep reporting expertise. He had been there the whole time.
At the time, those two experiences did not feel different. They both felt like the job. Whether an answer still had to be discovered or already existed somewhere else, the immediate demand was the same: keep moving. That is what makes the distinction so easy to miss. A capable person absorbs both.
Judgment is the obvious answer to what should remain in the person. But judgment names the response, not the problem. Deciding what a report should show was the first kind: nobody had that answer, because it did not exist yet. Finding out who understood the data was the second: that answer already existed, somewhere out of my reach. Both felt like figuring something out. Only one had to.
Some questions become clearer the harder they are analyzed. Others depend on what has not happened yet: how customers respond, whether a new way of working takes hold, what a team learns after putting something into the world. The Cynefin framework gives one language to that second class of problem: in a complex domain, cause and effect can be seen only in retrospect. The evidence is partly created by acting. More analysis can improve the decision; it cannot make the future arrive early.
A system becomes brittle when it encodes certainty the work itself does not contain.
The opposite failure breaks nothing. A capable person compensates — finds the right colleague, reconstructs the missing context, works out which source can be trusted — and the work keeps moving, which is exactly why the design problem stays invisible. Being connected earlier to the reporting knowledge already inside that company would not have told me what the report should show. It would simply have meant less of my time spent retrieving answers that already existed, and more of it on the ones that did not.
For a long time I read that second experience as an oversight. The company had not lost that knowledge, and it was not unaware of it. It was keeping it in a person — which is a particular way of keeping something. A document can be found by someone who does not know it exists; what a person knows can be found only by someone who already suspects they know it. So the search falls to whoever needs the answer — in my case, the person with the least idea where to look.
Our vocabulary finishes the job. The same knowledge, written down, is documentation; held in a person, it is experience. Only one of those words sounds like something a company owns.
Where knowledge is kept does more than decide who appears to hold it. It also decides how far it can travel. The ability to work an analysis to its edge, recognize what cannot yet be known and make a sound call is portable; it leaves with the person who has it. The accumulated map of one organization — who knows what, which rule bends, where the unwritten answers live — does not. The ability to build such a map travels; the map itself expires at the door. And the organization does not keep it either. It will spend the same years teaching the next person what it never wrote down.
Some seniority is expertise. Some is compressed organizational history.
Both are valuable, and neither is the problem. The problem is an organization that cannot tell them apart — one that requires exceptional people to keep routine work moving, and calls the requirement a strong bench. Good design absorbs more of the ambiguity it creates, so that judgment can go to the ambiguity the work contains.
The allocation does not stay neutral. Ask people repeatedly to reconstruct missing context and they become skilled at reconstruction. Leave consequential judgment close to the work and people get repeated practice making tradeoffs with live context, seeing what follows, and correcting their read. A different capability develops.
Whatever line we draw does not only divide the work. Given enough time, it begins shaping the people on either side of it.
And then it promotes them
From inside, none of that feels like being shaped. It feels like getting better at the job, and the evidence is real: what used to take days takes an afternoon, and the reasons are legible — a person knows where things are kept, who to ask, which failure this one resembles. That is what expertise feels like, and most of the time it is what expertise is. What stays out of view is that something chose which capability would compound.
An organization has no standing instrument for observing how it allocates ambiguity. What it can see is outcomes — who delivers, who is unblocked, whose projects do not stall — and by that measure the person who reliably gets things done inside a system full of gaps is its strongest engineer. Which is true. They are the strongest, at that. So they are given more, and then more again. The case written in support of it will say something like operates well under ambiguity — accurate, and silent on where the ambiguity came from.
Something similar happens to what people are able to say. We will only know by trying is an accurate description of a whole class of problem, and it is heard differently depending on who says it. From a principal engineer it reads as judgment. From someone four years in it reads as not having done the work. So people learn, quickly and without being told, how sure they are expected to sound — and the person closest to a problem that has no answer yet is often the one who can least afford to sound uncertain. The organization loses its best information about what it does not know, and loses it politely.
And someone who has spent years becoming excellent at working around a system’s gaps does not experience them as gaps. There is nothing to work around; there is only how things are here. The context that never arrives, the colleague you have to know to ask for, the decision that takes three weeks to find the person who can make it — after long enough these stop registering as design at all. They stop being decisions and become the place itself. Weather, not something built.
Which is why the next version tends to look like the last one — not from carelessness, but from the opposite. Someone who has succeeded inside a system carries an accurate model of what capable people can absorb, and builds to it in good faith. Their own record is the evidence, and it is not wrong. So the people best placed to repair a system are often the ones it rewarded for compensating, and the obstacle is not self-interest. It is that the thing looks fine from where they are standing.
None of this requires an exotic mechanism. An approval step is added because a change went wrong. The step adds delay, so changes are batched to be worth the wait. Batched changes carry more risk, which is a good argument for keeping the approval step. Nobody in that loop behaves unreasonably, and it tightens anyway.
A competence does not arrive with a label saying where it was acquired. Some of it you can tell: the language you write in travels; so does the way you analyze a situation and make the case for how to resolve it. Some of it you cannot, and that is the part that feels least like something acquired.
All of that assumes the line between what a system holds and what a person holds stays where it is. It has been moving for as long as people have written things down. Every time recording something becomes cheaper or easier, the line shifts — and what a person’s accumulated knowledge is worth shifts with it, usually without anyone naming what moved.
It is moving now. Work an agent will do has to become legible to it, because what people pick up by being around must reach it another way. So the things people knew without being told are being written down, one at a time: which customers get an exception and which do not, which of two conflicting reports people actually act on, who has to have seen something before it counts as decided. Not the hardest of them. But steadily.
Each thing written down is a small answer to a question nobody used to need: how much of this was the person, and how much was the place.
Something else changes once agents are doing part of the work. The unit that produces a result becomes harder to draw around a single actor: a person frames a problem, an agent does a first pass, something checks the result, and a person decides whether it is worth keeping. Capability is present at every one of those points, and whether any of it amounts to anything depends on what can pass between them — what context arrives, where action is permitted, what comes back in time to matter.
So the change may not be that capability is moving from people into machines. More of it now lives between them, where it belongs to no one in particular.
Capability spreads more easily than accountability. When output rises somewhere in an organization, authority follows — toward whoever, or whatever, is now producing more. It moves by practice: decisions start routing around one person and toward another, and nothing has to be announced for it to have happened. Accountability moves by paperwork — a role, a title, an on-call rotation, a name on a document — and paperwork is revised on a slower clock. So the two drift apart, and someone ends up answerable for outcomes they can no longer meaningfully shape. Nobody decides this. It is what happens when one part of a system speeds up and the rest keeps its old timing. People keep working hard inside systems they have stopped believing they can influence.
A system does not only get the people it hires. It gets the people it makes.
Very little of that line — what the system holds and what a person is left to work out — was set by the nature of the work. It was set by a long accumulation of small choices about what to write down and what to leave unstated. It is being redrawn the same way now, quickly, by people who would not say they are designing an organization.
Longer than the place that made it
The person we began with — the one who entered the wrong place and became strangely ordinary — did not lose anything on the way in. Whatever they carried is still there. What changed is what they could spend it on, and what the people around them agreed to call the spending.
Which leaves all of us in an odd position. The account you hold of your own ability — what you are good at, what you are not, how far you might go — was assembled partly out of a system you never saw and could not have named. It cuts both ways. Some of what looks like a limit was never yours; some of what looks like a strength was not either. And the verdict arrives with no record of how it was reached.
That case is closed. The system you are building for the people around you is not. It is assembled out of small acts: what gets written down and what is left for someone to work out, who is connected to what they need, which choices stay close to the work. Where capability lives, for them, is partly your decision — and it only looks like a decision afterwards.
The terms do not hold still. The interval between one arrangement and the next has been getting shorter, and there is no particular reason to expect it to lengthen.
A view formed of someone early survives the reorganizations that follow — recorded in a review, passed along in conversation, inherited by someone who was not there when it formed. But so does everything else the place built: the range a person was given room to develop, the habits they formed working around what was missing, what they came to believe they are for.
Somewhere in a place you have some part in shaping, there is a person who looks less capable than they are. They cannot tell you. They may not know. And looking harder at them may not help, because what you are seeing is already the output of the arrangement.
The person does not have to be fully legible. The arrangement is legible enough: who gets brought into a decision early, what arrives after it was needed, which answers only one person has. None of it is hidden. It stops being noticed, which is not the same thing.
What we stop noticing, we start calling circumstance.