Who, Not What
By Jeff Chung - Chief Strategy Officer
09/11/2026
Three people in the same company opened the same AI assistant one afternoon. One got a cautious editor. One got a bold strategy partner. One got something that agreed with everything she said. The same tool, in the same hour, produced three different workers, and nobody found it strange, because nobody was expecting a worker. They were expecting a tool.
That is the product working as designed. A general assistant is built to become whatever the prompt in front of it requires, and across an enormous range of tasks, that is exactly the right design. The trouble starts when an organization asks that shape-shifting to hold a role.
Fine for a task, fatal for a role
A role is not a stack of tasks. It is a standing assignment: this part of the business belongs to you, handled the way we handle it, with some things raised to a person and the rest left to your judgment. A person holding a role can be counted on to be roughly the same person on Thursday that they were on Monday. That continuity is so basic to how organizations run that nobody thinks to write it down as a requirement. No role charter says "please be the same person tomorrow." It is assumed, and everything else about working together gets built on top of the assumption.
A general assistant cannot offer that continuity, and not for lack of capability. It is whoever the last prompt asked it to be. The cautious editor and the agreeable one from that afternoon were both real, and neither is guaranteed to show up tomorrow. That does not make the general tool bad; it is genuinely good at what it is for, and what it is for is tasks. A role asks for something it was never built to hold: a who that persists.
A correction has to stick to someone
Two earlier articles looked under the hood of an AI specialist from two directions. One argued that a specialist earns knowledge by doing the work, the way a person two years into a role knows things no handbook covers. The other argued that capability should be authored as a unit, written once and handed to any worker whose role calls for it.
Both arguments share a premise worth saying out loud: that there is a someone for those things to belong to. Earned by whom, and handed to whom? A correction only compounds if the worker you corrected is the worker who shows up next quarter. A skill is only worth reviewing and granting if there is a defined recipient with a mandate to use it. Continuity is the precondition for accumulation.
That sounds abstract until a team tries to build on a worker that resets. The review that went well last month tells you nothing about this month. The authority a team was ready to extend has no one to extend it to. Both of the earlier arguments need the stable who: without it there is no one whose memory it is, no recipient for the capability. Corrections land on whoever happens to be there, which is to say they land on nobody.
Trust needs somewhere to accrue, and a worker that is someone new every session gives it nowhere.
One role's worth of who, written down
So what is an identity, for an AI worker? The word gets used loosely, so it helps to be concrete. Not the name and the avatar, though it will probably have both. The identity is the written answer to the questions a role charter answers for a person. What the worker covers, and where its coverage ends. Which outcome it is accountable for. The way it leans when a call is close, toward caution or toward speed. And the two entries organizations tend to forget: what it escalates without being asked, and what it declines because that work belongs to a different role.
Notice the size of that definition. It is one role's worth of who, not a function's worth of everything. A worker defined to be useful to everyone for anything is a generalist with a nicer nameplate. Organizations already know this discipline; nobody writes a charter that says "everything, as needed." The who gets the same treatment: one role deep, with edges.
Once written, the who has to hold everywhere the work happens. The worker a team lead consults in the morning has to be the same worker that turns up in the afternoon's reporting flow, with the same scope and the same defaults. That reads as obvious. It is also the part that quietly breaks first, because every new surface is a new chance to configure the worker a little differently, and a dozen small divergences add up to a stranger.
Authored, and honest about it
This is the part of the argument that invites misreading, so it is worth saying plainly. Giving an AI worker an identity is not pretending it is a person, and it is not a personality costume stretched over a tool to make it likable. The identity is authored. The organization writes it and owns it, and can change its mind. Nothing about that claims personhood. What it claims is provenance: someone decided this worker's shape, and someone answers for it. That is a sturdier foundation for trust than any amount of simulated warmth, because it gives people something they can check.
I would go further: the authored who is what makes accountability available at all. Work gets reviewed against a stated mandate, and authority gets granted within defined edges. Trust needs somewhere to accrue, and a worker that is someone new every session gives it nowhere.
None of this is the finish line. A written who settles who is in the seat. It does not settle how that worker behaves when the work gets contentious, or where its judgment ends and a person's begins. Those are conduct questions rather than identity questions, and they only become askable once there is a someone to ask them about.
Chief Strategy Officer, Arkane Digital
Jeff advises organizations on AI transformation, focusing on connecting business strategy to practical implementation without fragmented or reactive adoption.