What is an LLM manager?

There is a role appearing in most organisations that nobody has written down. The people doing it do not have a title for it, HR has no job description for it, and the accountability arrived before either.

In an interview on 23 September 2026 I described them as “LLM managers”: people who no longer do the production, they oversee the platforms and put their name to the output.

What the day actually looks like#

Four activities, none of which is the work they were hired for.

• Running the platforms. Knowing which system is used for what, where the context lives, what has been tried before.
• Testing one model against another. Putting the same problem to two systems and reading the difference, which is a research skill nobody has been trained in.
• Adding the nuance. Supplying what the machine has no way of knowing: the history of the account, the thing the client said on the phone, the reason the obvious answer was rejected last year.
• Putting their name to it. The part that looks like the smallest and is the whole job.

It is not the same as managing people#

The word manager carries a set of assumptions that do not transfer. A person you supervise can explain their reasoning when asked, gets better when corrected, and tells you when they are out of their depth, usually. None of those hold. The system is fluent when it is wrong, does not improve because you told it off, and has no way of signalling the difference between an answer it has good grounds for and one it does not.

So the skills that make somebody a good manager of people are not the skills that make them a good manager of models, and an organisation promoting on the first will be disappointed by the second. Who supervises work they cannot do is the sharper version of the same problem.

The qualification problem underneath it#

The people who are good at this are almost always people who did the underlying work for long enough to recognise a wrong answer at a glance. That is not a preference, it is what the role requires: the value added is the correction, and you cannot correct what you could not have produced.

Which creates the difficulty nobody has answered. They learned it by doing the work that is now being given to the machine. If the entry-level version of the task is the first thing automated, the supply of people qualified to oversee it is being cut off at the same time as the demand for them rises. The missing rungs.

Where it goes wrong#

Two ways, and they are opposite.

The first is that approving becomes indistinguishable from deciding. Enough volume, enough speed, enough output that is nearly always fine, and the check becomes a rhythm rather than a judgement. What the organisation has recorded as an oversight mechanism is a signature. Proving you did the work is what a person in that position eventually gets asked for and often cannot produce.

The second is quieter. The work is genuinely being done, the output is genuinely better than it was, and the person doing it feels they have stopped doing anything. Reviewing all day is not the same experience as making something, even when the result is better, and the discounting people apply to their own verification work is a real and measured effect. The verifier’s discount.

The stamp is the job#

If the role is going to be written down, this is the sentence to write down first. The machine did not sit in the rooms. It has not read what you read the way you read it, and it does not carry the consequences. When a person puts their name to something, they are putting a stamp of accountability on it, and in a job where the production has moved to the machine, the stamp is what is left. It carries the whole of the value.

A job description that says so is a better protection for the person holding it than any policy about approved tools. It should say what they may approve, what they must escalate, roughly how long an approval is expected to take, and what evidence exists that they can still tell a plausible wrong answer from a right one.

What this does not establish#

How many people are in this position, whether the role is net positive for the organisations creating it, or whether it is a transitional stage on the way to something else. None of that is measured here, and this page is an argument about a shape that keeps appearing in rooms rather than a finding. The evidence behind the parts that are measured, the supervision problem and the discounting of verification work, is on the pages linked above with what each study does not show stated alongside.