There is a role in every AI-enabled process that the org chart has never bothered to name.
You can find it in the analyst who edits the model’s variance narrative before it reaches the CFO. In the safety coordinator who accepts the incident summary with two small changes. In the recruiter who reads the ranked candidate list and decides which three names to bring forward. In the customer-service lead who picks up the escalation from the chatbot and repositions the case for the manager. In the operations manager who takes the dashboard’s green indicator and describes what it means to the board.
None of these people wrote the system’s first sentence. All of them wrote the sentence the organization ended up using. They are translators — the humans who stand between machine output and organizational language. Their edits are small. Their titles are ordinary. Their power, in the shape of the decision that reaches the outside, is disproportionate to both.
Why the role exists at all
The reason there is always a translator is that the outputs AI produces do not, in most cases, arrive in a form the organization can act on directly. A model’s ranked list is not a shortlist until someone decides which cutoff is defensible for this role. A summary is not the incident report until someone edits the framing that carries risk. A generated variance explanation is not the finance narrative until someone decides which of its emphases the CFO will stand behind. A chatbot’s answer is not the company’s answer until someone accepts it as such.
The translator is the person who makes that transition. It is often invisible work — a re-ordering, a softening, a promotion of one fact and a demotion of another, a rewrite that reads like an edit. From the outside it looks like clerical polish. From the inside, if you slow the tape down, it is where authorship of the company’s language actually settles.
This is why the role tends to accumulate power quietly. The translator does not have decision rights on paper. They have framing rights in practice.
The two edits
There are two kinds of translation edits, and they matter differently.
The first kind changes fidelity. The system produced language that was accurate but not true enough to decide from. The translator corrects it. The summary emphasized a variance that was cosmetic; the translator restores the underlying operational fact. The ranked list treated an unconventional path as a demerit; the translator marks it as strength. The chatbot’s answer was technically correct and materially misleading; the translator issues the response the organization actually stands behind.
This kind of translation is the reason the role exists. It is where human judgment enters the pipeline. When it is working, it is the most important sentence-level control the organization has.
The second kind changes framing without changing facts. The translator does not correct the system. They shift the tone so a difficult answer sits comfortably on the manager’s desk. They compress a nuanced summary into a bullet that will not slow the meeting. They accept a green indicator and describe it as if the color were the answer, not the input.
This kind of translation is where the role becomes dangerous, because it looks identical to the first kind from a distance. Both are described in the workflow as “review.” Both leave a clean record. Only the translator knows which of the two just happened.
Why the translator is often the most powerful role in the sequence
Three properties compound to make this role structurally consequential.
Position. The translator is the last human before the sentence leaves the company. Whatever they hand upward or outward becomes the company’s answer.
Framing rights. They control the language, not the underlying data. Language is the interface through which the rest of the organization — and everyone outside it — will experience the decision. A well-framed weak analysis will beat a poorly-framed strong one almost every time inside a busy leadership team.
Deniability on both sides. If the decision goes badly, the manager can point downward: “I acted on what I was given.” The system’s owner can point upward: “I only produced the input; the human made the call.” Between the two, the translator’s authorship is difficult to reconstruct after the fact. Nothing in the record marks the moment the framing shifted.
The result is a role that carries a great deal of organizational language on a set of small edits, and is nowhere on the responsibility map.
Why the translator is often the least visible role in the sequence
The org chart does not describe them because they are not doing a new job. The recruiter is still a recruiter. The analyst is still an analyst. The safety coordinator is still a safety coordinator. What changed is not their title. It is that a system now sits in front of them, doing the work they used to do first, and the work that used to be their whole role has quietly compressed into a series of edits on top of the system’s output.
From a headcount perspective, nothing looks different. From a work perspective, the center of gravity has moved. The person’s day is no longer mostly composing. It is mostly translating. The organization does not usually notice, because the output volume goes up and the timing improves.
But the shift changes what governance would need to inspect. When the role was composing, oversight looked at the person’s judgment. When the role is translating, oversight has to look at two things at once: the input the person received, and the change they made to it before it left. Almost no organization is set up to see both.
What a mature organization does about it
The translator problem is not solved by removing the role. The role is doing real work, and much of the time it is the reason the system’s output becomes trustworthy. What can be done is to make the work visible.
Name the translation step. In the class of decisions where framing matters, mark on the workflow that a translation is happening — that a human is reframing system output into organizational language. Names change what people can see. Once the step exists on the workflow, questions about it become legitimate.
Preserve both artifacts. Keep the system’s output and the version that left the company. Not to litigate every edit — most edits are corrections and should stay off the audit path — but so that when a decision goes wrong, the framing shift is a recoverable fact, not a story assembled later from memory.
Give the translator authority the org chart does not currently show. If the analyst’s edit is going to become the CFO’s narrative, the analyst needs the authority to refuse the system’s framing, escalate a disagreement, and be protected when they do. Otherwise the role’s power will keep operating without the accountability that would make it healthy.
Ask, at the design stage, whether the translation should exist at all. Some categories of decision — safety findings, denial letters at consequence, communications with regulators — should not have a translation step layered on system output. The system should not be positioned to produce a first draft. The role should be composing, not translating.
The role the org chart has to add
Somewhere in most AI-enabled organizations right now, a person is looking at a screen and deciding which of two accurate sentences will become the sentence the company owns. That person is not senior. They are not on any list of AI decision-makers. If the sentence goes wrong, they are not the one who will be asked to defend it.
They are also the human who authored the language before it traveled.
The translator problem is not that the role exists. It is that the org chart pretends it does not.
This essay draws on Interlude I, Chapters 4 and 5, and the Final Interlude of AI in the Org Chart.