Chapter 14: The Role Landscape
You can do the job. The last chapter asked where the job is going, and the honest answer is that it is going to a place the market has not finished building, where some of the work has a price on it and the most important part of it does not. This chapter is about what you are worth in that place, and how anyone is supposed to see it. Not the labor market explained. Your record, held up to the light, and what it buys you.
Start with two postings.
Two postings, same week, same company
They went up on a Tuesday. I want you to read them side by side, because the gap between them is the whole chapter.
The first is for an Evaluation Lead. It has a salary band, a tight one, printed right there in the listing. It names the work precisely. Own the golden dataset. Calibrate the judge. Hold the regression suite across model updates. Report a number the company can stand behind. It reads like a posting written by someone who knows exactly what they are buying and what it costs. You could check yourself against it in an afternoon.
The second is for a Senior Product Manager, AI Products. It has no band, or a wide one that means the same thing. The description is a paragraph of nouns. Strategy, stakeholders, roadmap, cross-functional leadership, agentic workflows, outcomes. Buried in the middle, stated as if it were one item among many, is a sentence that quietly assumes an entire unwritten job description. The person in this seat will decide which agent behaviors are acceptable. They will own the gate where the green checkmark gets signed. They will write the brief the agent executes against. They will hold the boundary in the room when the schedule wants to ship past it, and notice, before anyone else does, when the thing has drifted. None of that is named as a skill. It is folded into the word “leadership,” and there is no band on it because the company does not yet know what it costs.
That is the split, and it is the whole of it. The work that builds the agent has a price. The work that supervises the agent does not. The first posting pays for building because the market has always known how to pay for building. Product manager compensation at the frontier labs has become its own headline, medians well into the high six figures and beyond, because someone decided the judgment of what to build with AI is scarce. Look for the band on the person who watches whether the thing is still safe, and it is not there. The supervision is real work that real teams are starting to do, and the company, at the level of budgets and titles, has not caught up.
[FACT-CHECK: Book 3 states the PM-comp contrast qualitatively (“high six figures and beyond,” Levels.fyi medians, not AI-PM-specific) and flags the contrast, not the precise number, as the load-bearing claim. If a sourced figure is added later, attribute it. Do not invent a number for the supervision roles; the point is precisely that none exists.]
A handful of those supervision slices now have names that did not exist eighteen months ago. The eval owner who keeps the checkmark honest. The context owner who writes down what the agent can actually see against what the job requires it to see. The operations lead whose entire week is the running agent and its drift. The one-page inheritance at the front of this book maps that grid; this chapter does not re-walk it. The only fact you need from it is the one the two postings already showed you. The build slices are getting named and priced first. The supervision is not. And the unpriced part is where most of the irreplaceable judgment lives.
An interview that turns on a record
Now sit in on a conversation that the priced-and-unpriced split runs straight through.
A hiring manager is interviewing for the second posting, the one with no band. She has done this nine times this quarter and the conversations blur. Every candidate says the same words. Strong on agentic systems. Deep experience with LLM products. Owned the AI roadmap. She has stopped hearing them, because the words are free and everyone has them.
The tenth candidate is twenty minutes in when something changes. She has asked her standard question, the one designed to separate the talkers from the rest. Tell me about a time the model was confident and wrong and you caught it. The first nine answered with a story, well told, unfalsifiable. This one answers with a story and then says, I can show you. He turns his laptop around.
On the screen is not a slide. It is a ledger. Dated entries, quarter by quarter. A column for the calls where he overrode the model and a column for whether ground truth later said he was right to. A catch rate that held steady while, in the next column, the share of his work the model was doing climbed. One territory with a falling line and a note written beside it, in his own words, about what he stopped trusting himself on and what he did about it.
The conversation is no longer an interview. The manager is asking about the falling line. She wants to know about the territory he marked for repair, and he tells her, and the answer is one no confident candidate could have given, because the confident candidate has no record of ever having been wrong on purpose, in the open, to find out where the edge was. She is not taking his word that his judgment is live. She is reading the trend.
What just happened in that room is the entire argument of this chapter, compressed. The market has not learned to price supervision. It has no band for it and no certificate for it. But a record of maintained judgment, dated, with a trend and an honest down quarter, is the one thing in the candidate’s possession that demonstrates the capability instead of asserting it. It cannot be produced in a weekend. In a field where the claim is free and the proof is scarce, the proof is the whole asset. The candidate who walks in with it is selling the thing the market most needs and least knows how to ask for, and he has made it legible without waiting for the market to write the band.
Hold that record where it is for now. The next chapter is about why no one will ever certify it. The point here is narrower. It is the difference between the ninth candidate and the tenth, and it is not talent. It is evidence.
The posting that hunts a zebra
Not every posting is honest about what it wants, and one kind of dishonesty poisons the whole market. You need to see it from both sides of the table.
There is a saying every clinician learns early. When you hear hoofbeats, think horses, not zebras. Look for the common animal before the exotic one. The zebra exists. It is just rare, and a search built to find one passes over every horse standing in plain view.
Picture the manager who wrote one of these. He needed a person to own an agent in a regulated domain, and when he sat down to write the posting he reached for everything that would feel safe to a skeptical executive reading over his shoulder. He asked for fourteen years of experience at the intersection of AI and digital health, for a role defined by tools three years old. He asked for deep prior expertise in a stack of regulatory frameworks, as a thing you must already carry through the door, when a capable practitioner can now get a fast structured first pass on any of those frameworks in an afternoon and surface the questions that need a lawyer. He was not trying to be unreasonable. He was trying to be careful, and careful, written down, came out as a description of an animal that does not exist.
The next size up from the zebra is the unicorn. That is the posting wanting one person who is at once a deep eval engineer, a fluent agent operator, a domain expert, a strategist, and a stakeholder whisperer, senior in all of them, for one salary. That person mostly does not exist, and the teams that actually ship do not look for them. They hire complementary. A strong eval owner next to a strong domain expert next to someone who can hold the space between them, because the work is spread too far across the grid for one head to hold it all at depth.
The trap has two edges and both cut you. As a candidate you read the unicorn posting, conclude you are unqualified for a job no real person is qualified for, and pass on work you could do. As the person who will someday write the posting, you empty your own candidate pool and wait months for an animal that is not coming. Same error, both directions. Stop hunting zebras. Hire, and become, the horse.
The choosing
So you are standing where the manager who wrote the careful posting cannot help you and the unicorn does not exist. You have a record, or you are building one. The question is where you point it, and to make it real, watch one person decide.
She has two offers. The first is a named seat, an eval owner role at a company that has learned to pay for it. The band is printed. The work is finite and she could be excellent at it. The second is the seat with no band, the one that owns the gate and the brief and the boundary and the drift, the seat the second posting could not describe. She has a week to choose, and the choice is the chapter’s three honest options worn by one person, each with its cost.
She can specialize. Take the eval seat, grow deep in the one slice the market has named, and let the printed band do the work of proving her worth. The cost is the narrowing. She trades the breadth that let her hold a whole product for depth in a single cell, and she takes on the specialist’s standing exposure, that the slice can be redefined, automated further, or folded back into another seat when the org chart turns again. She becomes very valuable for one thing, and very dependent on that thing staying a job.
She can orchestrate. Take the broad seat and become the person who directs a team of junior practitioners and agents across the full cycle, whose value is not producing the analysis but looking at an analysis generated in twenty minutes and seeing at once what is missing, what misleads, and what actually matters for the decision. This is the highest-leverage version of the old job and it does not narrow her. The cost is that the market has no clean band for it, so she has to make it legible herself, with exactly the kind of record the tenth candidate turned around on the table.
Or she can hold the seam. Take the broad seat for a different reason: to stand where the eval number meets the domain reality meets the boundary call and notice when they disagree. Every specialist sees one cell. The eval owner sees the checkmark, the context owner sees the blind spot, the operations lead sees the drift, and the failures live in the spaces between them, the handoffs that harden into borders. The seam is scarce and grows more valuable as teams grow. The cost is that it is invisible from inside any one seat, easy to mistake for having no specialty, and impossible to put a band on. She would be most needed exactly where she is hardest to price.
These are not a ranking. They are three relationships to the same split, and the right one depends on her, and on the size of the room she is choosing inside of. At a five-person startup the cells have not separated and should not; the one separation that cannot wait is that someone other than the agent’s builder owns whether it can be trusted, and that someone is her, so she orchestrates with a seam-holder’s eye and specializing would be premature. At a scale-up the cells have become roles and specializing is finally a real option, because the eval work and the context work have each grown past what anyone can do between other duties; the live question becomes whether she wants to be one of the emerging specialists or the person who holds them together. At an enterprise the cells are teams and the seams are borders, intent leaking across a ticket between departments that used to share a hallway; the seam is most valuable and most invisible there, and the danger is the seat that looks full and has no budgeted hours behind the name.
I will not analyze this from a safe distance, because the book has spent its length arguing that the honest move is to claim a position and stand behind it. So here is mine. I would hold the seam, and I have, more than once, chosen it on purpose over the better-priced alternative. My whole working life sits on a seam already, between medicine and product, and what I have learned there is that the expensive failures almost never live inside a discipline; they live in the handoff between two disciplines that each did their part correctly and neither owned the gap. The eval owner seat is real and I respect it, and I would decline it, not because I could not do the work but because taking it would trade the vantage I am most useful from for depth in a single cell, and the cell can be redrawn under me. The cost of my choice is exactly the one the chapter names: the seam is the hardest thing to put a band on, easy to mistake from the outside for having no specialty, and I have had to make it legible myself every time, with a record rather than a title. I would still choose it, because being the person who sees the disagreement between the number, the domain, and the boundary is the one position the tools are furthest from taking, and it is the one I would least want to be without when something breaks.
What survives whichever she picks is worth being exact about. What does not cross over is artifact production. The competitive summary, the first-draft PRD, the market analysis, the work that used to fill a senior week, is now an afternoon for a junior with good tooling. If her value was in producing those, the value has moved. What crosses over is the judgment underneath. Pattern recognition deep enough to say, early enough to matter, I have seen this break before, here. Domain depth, which is not knowledge the tools compress but the thing that makes their output trustworthy instead of merely plausible. Differentiated conviction, the one place the tools keep falling short. The read that resolves the stakeholder who agreed and then blocked. None of that is in any posting’s requirements list, and all of it is the actual asset. The crossing is a demotion of the part of her the market is done paying for and a promotion of the part it cannot yet price.
The choice, and what it rests on
She chooses. You will choose. Specialize into a seat with a printed band, orchestrate across a cycle the market cannot yet price, or hold the seam none of the specialists can reach, each with its honest cost, the choice yours and your stage’s to make.
But every one of those paths rests on the same footing, and the interview scene already showed you what it is. The ninth candidate asserted the capability and the tenth evidenced it, and only the second conversation went anywhere. Whichever posture you pick, you will eventually be across a table from someone who has to decide whether your judgment is live, and the posture is not what settles it. The record is. So the career question turns out to depend on a prior one the next chapter has to answer first: in a field this immature, how does anyone establish a capability like this at all, when the whole apparatus the market reaches for to settle such questions was built for a different kind of skill.