Distinguish the main AI-adjacent roles and where you might fit.
Goal: Distinguish the main AI-adjacent roles and where you might fit.
Bina Velasquez spent eight years teaching high-school biology before she changed careers into AI work. Six months in at Trellix Health — a 40-person startup whose product, Scribe, drafts clinical visit notes and patient-message replies for small clinics — she's in a standup, watching four people argue about one feature.
Marc wants to know whether Scribe should show the doctor how unsure it is about a drug name. Pri says she can wire a confidence score into the draft, but she has no idea what number should count as "too unsure." Bina has spent the week reading hundreds of Scribe's drafts and knows exactly where it gets shaky. And Theo, on the call from a pilot clinic, says the doctors there won't trust any number until someone sits with them and explains it.
One feature. Four completely different jobs. That room is this whole topic.
The AI-adjacent world is full of titles — AI trainer, prompt engineer, AI PM, conversational designer, enablement consultant — and they blur into each other on job boards. But underneath the noise there are really four main shapes of work: deciding what AI should do (AI Product Manager), getting reliable results out of it (prompt-focused work), judging whether its output is any good (AI operations and evaluation), and helping people actually use it (AI consulting and enablement). Learn these four shapes and almost any AI job posting will snap into focus.
The titles are a fog. The four shapes of work underneath are clear.
Marc Devlin owns what Scribe's AI features do and how they behave. He was a regular product manager for years and learned AI on the job. He cannot train a model — and doesn't need to.
An AI Product Manager is a PM who specializes in products built on AI. The classic PM question is "what should we build, and why?" Marc lives there too, but AI adds a second set of questions that ordinary products never raise. When Scribe is unsure, should it guess, hedge, or stay silent? When it's wrong, how does the doctor catch it before a bad note reaches a patient? What does good enough even mean for a clinical draft, and who decides?
Those are product decisions about uncertainty, error, and trust. Marc doesn't write the model; he decides how the model should act in front of a real user, and he defines the bar for quality that everyone else builds toward.
So the AI PM needs the usual product skills — talking to users, prioritizing, saying no — plus enough AI literacy to make these calls without being fooled. Notice what's not on the list: building or training models. That's the engineer's job. Marc's value is judgment about the product, sharpened by understanding what AI can and can't be trusted to do.
If you're drawn to products and decisions, this is your target. (Pair it with the PM track.)
Bina's first real win at Trellix had nothing to do with code. Scribe kept ending patient messages with a cheery "Have a great day!" — fine for a pizza order, jarring under a note about a worrying lab result. Bina rewrote the instructions the model follows, tested a dozen versions against real cases, and found the wording that made Scribe match the gravity of each message. Tone fixed. No engineer required.
That's prompt engineering: crafting, testing, and refining the instructions that get reliable results out of a model. Because how you ask an AI changes what you get, the wording becomes real, valuable work — and it leans on clear writing, logic, and patient experimentation more than on programming. That mix makes it unusually accessible to people coming from non-coding backgrounds, like an ex-teacher who's spent years writing instructions a room of distracted teenagers will actually follow.
Now the honest part, because a paying learner deserves it.
If you go looking for jobs literally titled "Prompt Engineer," you'll find fewer than you'd expect. Standalone postings for that exact title dropped roughly 30% from 2024 to 2026. But that's not the skill dying — it's the skill spreading. Over the same stretch, roles that require prompt skills grew about 3x. The work got absorbed into broader titles: AI Engineer, AI Product Manager, AI Developer, Conversational Designer.
And inside those modern roles, prompting is only part of the job.
In a real AI role today, writing prompts is roughly 30% of the work. The other 70% is building evaluations, running tests, and iterating on data.
Bina lives that ratio. The "Have a great day!" fix took an afternoon of writing; proving the new tone held up across hundreds of messages took the rest of the week. Which leads straight to the next role.
Pri Reddy can build a model, connect it to a clinic's records, and ship a slick draft into the doctor's screen. What Pri cannot do is read that draft and know whether it's clinically right. She's an engineer, not a physician. Someone has to be the human who looks at the output and judges it.
That someone is Bina — and that's where she started.
AI operations and evaluation is the human-in-the-loop work that keeps AI accurate and safe: evaluating output quality, labeling and curating data, ranking which of two responses is better, and red-teaming the system by deliberately trying to make it fail. The titles vary wildly — AI Quality Analyst (Bina's first one), AI Trainer, Data Annotator, Model Evaluator — but the core is the same. You are the judgment the model lacks.
This is very often the most accessible entry point into AI work, because it rewards exactly what a teacher, nurse, paralegal, or editor already has: careful judgment, attention to detail, and real domain knowledge — not coding. Bina's old classroom instinct for spotting a confident-but-wrong answer is the single most useful thing she brought to Trellix. When Dr. Soledad Marchetti, a physician at a pilot clinic, flags that a Scribe draft is subtly wrong — plausible-sounding, but unsafe — Bina is the one who turns that into structured feedback the model can learn from.
A note on pay, because it's the most variable role. Entry-level annotation and trainer work tends to be contract, around $15–30/hr. But specialized domain experts — people who can evaluate medical, legal, or financial output — command roughly $40–80/hr, or about $80–120K full-time. The judgment is cheap to start and gets expensive fast once a subject sits behind it.
Theo Brandt-Okonkwo never touches a prompt or a label. His job is the clinic on the other end of the phone. When a new practice signs up for Scribe, Theo figures out what that clinic actually needs, shows them how AI fits their workflow, trains the staff, and stays close while they adopt it. He's the bridge between "what AI can do" and "what this clinic will actually use."
That's AI solutions consulting and enablement: helping a business or team apply AI to its real problems. It blends consulting, communication, and AI literacy, and it's a strong fit for people-oriented types who'd rather sit with a worried doctor than stare at a spreadsheet of model outputs.
Now step back and look at all four — Marc, Bina, Pri's evaluators, Theo. They share four threads:
Because the four roles share so much, the right entry point is a question of fit, not a wall you can't climb back over. Bina started in evaluation and is already drifting toward prompt work and product calls; the threads carry over.
A rough guide:
| If you love… | Look at… | Why |
|---|---|---|
| products and decisions | AI PM | you'll shape what gets built |
| writing, logic, tinkering | prompt-focused work | the craft is words and tests |
| detail, judgment + a subject | AI operations / evaluation | usually the easiest entry |
| people and business problems | consulting / enablement | you bridge AI and humans |
Don't over-index on the exact title — the field is young, labels blur, and they're renamed constantly. Read each job description for what it actually involves, the way you'd read past a product title to the responsibilities.
A Scribe draft goes out in testing with a patient's medication listed at double the real dose. Caught internally, thank goodness. Watch the four jobs light up.
Bina (evaluation) catches it first. She's reading drafts against the source records and flags that Scribe is mis-reading dosages whenever the chart abbreviates the unit. She files it with three real examples and a note: this is a safety bug, not a tone bug.
Pri (engineering) confirms the model is stumbling on those abbreviations in the records — that's the technical root cause, hers to fix.
Bina again, now in prompt-focused mode, drafts a sharper instruction telling Scribe to never infer a dose it can't read cleanly and to flag it for human review instead. She tests it against her three cases plus forty more she pulls overnight.
Marc (AI PM) makes the product call: until this is solid, every dose in a Scribe draft ships with a "confirm this" flag the doctor must clear. He's decided how the AI should behave when it's unsure — accept a little friction to protect a patient.
Theo (enablement) tells the pilot clinics what changed and why, so Dr. Marchetti sees the new confirm-step as Scribe being careful, not broken — and keeps trusting it.
One bug. Four roles, four kinds of "done." Bina touched three of them in a single day, which is the whole point: the lines are real, but they're doors, not walls.
Pull up two real AI-adjacent job postings online — search any of "AI trainer," "AI product manager," "AI evaluation," or "AI enablement." Ignore the titles. For each one, read the responsibilities and sort them into the four shapes from this topic: deciding what AI does, getting results out of it, judging the output, helping people use it. Most postings mix two or three. Write down which shapes dominate each — that's the move that tells you whether a job actually fits your strengths, no matter what it's called.
Preparing your quiz…