Understand the category of roles that work with AI without building it.
Goal: Understand the category of roles that work with AI without building it.
Bina Velasquez spent eight years teaching high-school biology. When she started applying for "AI jobs," every posting seemed to demand a computer-science degree and years of Python, and she nearly closed the laptop for good. Then she landed at Trellix Health, a 40-person startup, as an AI quality analyst. Her job on day one was not to build a model. It was to read what the model wrote and decide whether a doctor could trust it.
Trellix makes a product called Scribe, an AI assistant that drafts clinical visit notes and patient-message replies for small medical clinics. The drafts are good. Sometimes they are confidently, dangerously wrong, and somebody has to catch that before it reaches a patient. That somebody is Bina, and the category of work she stepped into has a name.
An AI-adjacent role is a job that works with AI rather than building it from scratch. People in these roles apply AI to real problems, shape how it gets used, guide how it behaves, and judge whether its output is any good. The toolkit is human: clear thinking, clear writing, domain knowledge, and judgment. Not machine-learning math.
Most newcomers assume an "AI career" means becoming the engineer who builds the model. That is one job, rarely the realistic one for a career-changer. The work Bina does sits right next to the engineering, which is where the name comes from.
When cars were new, you did not have to build engines to have a great automotive career. You could drive, repair, manage a fleet, or teach others to drive. The engine builders mattered enormously, and there were never very many of them. Around that small core grew a whole economy of people who used cars, kept them running, and helped others get value from them.
AI works the same way. A small number of ML engineers build the engine. Everyone else applies, guides, and judges it.
AI-adjacent roles are AI's drivers, mechanics, and instructors, not its engine builders.
Hold that picture, because it dissolves the fear that keeps people like Bina out. At Trellix, Priyanka Reddy is the ML engineer. She builds the models, wires up the data pipeline that feeds Scribe patient history, and connects everything through APIs. Pri is brilliant at all of it. But ask whether a drafted clinical note reads correctly for a family physician, and she'll be the first to say she has no idea. She built the engine; she can't judge the drive. That judgment is a different job, and it's the one this whole course is about.
Here is a fair question: if the AI is so impressive, why does it need all these people around it?
Because modern AI is powerful but imperfect, and non-deterministic. The large language models behind tools like ChatGPT and Claude predict likely text rather than look up verified facts, so the same prompt can produce a different answer twice, and a fluent, well-formatted reply can still be flat wrong. That single property creates the entire job family.
A company racing to use AI discovers it needs humans to:
Theo Brandt-Okonkwo, the consultant who onboards new clinics onto Scribe, puts it plainly. A clinic doesn't need to be told what AI can do; they've read the headlines. They need someone to bridge "what AI can do" and "what this clinic actually needs on a Tuesday afternoon." That bridge is built by people, not models.
It would be easy to dismiss all this as hype. The numbers say otherwise, and they're specific enough to check.
By early 2026, AI skill requirements showed up in roughly 7 in 10 US tech job postings — and that share was still climbing sharply year over year, according to Dice's 2026 Tech Jobs Report. And Gartner projected that by 2026 more than 80% of enterprises would have used generative-AI APIs or deployed GenAI-enabled applications, up from under 5% in 2023. That target year is now here, which means nearly every company is reaching for AI.
Put those two facts side by side and a gap appears. Almost everyone wants to use AI; very few have enough people who use it well. That gap is where the jobs are.
There is a second, quieter advantage hiding in the dates. This field is only a few years old. There is no entrenched twenty-year career ladder, no gatekeeping credential everyone already has and you don't. Compare that to law or medicine, where the people ahead of you spent a decade earning their place. Here, almost everyone is a few years in at most. A motivated newcomer in 2026 is not hopelessly behind; the starting line is unusually close.
So which skills actually win in these roles? Four keep showing up, and a career-changer often already has them.
Clear writing and thinking. Prompting an AI well is, at bottom, clear communication: saying exactly what you want, with the right context, in the right order. Bina's years of writing lab instructions a sixteen-year-old could follow are direct training for this.
Domain expertise. This is the big one, so it gets its own lesson next. To judge whether an answer is correct, you have to know the subject. A nurse can judge medical text; a teacher can judge educational content.
Critical thinking and attention to detail. Spotting the one subtly wrong sentence in a fluent, professional-looking draft is a human judgment skill. Bina's teacher instinct, her nose for a confident-but-wrong answer, was honed grading a thousand essays that sounded right.
Curiosity and adaptability. The tools change monthly. A willingness to experiment and keep learning beats any credential that's already going stale.
That list leaves out programming, calculus, and model architecture. A writer, teacher, nurse, or paralegal can be better at many AI-adjacent tasks than a pure coder. Pri can build a system that retrieves the right patient record. She cannot tell you whether the resulting note would make Dr. Marchetti wince. You might be able to, and that is worth being paid for.
One Tuesday, Scribe drafts a visit note for a pediatric patient. It's clean, well-structured, and lists a medication at a dose meant for an adult, roughly triple what a child that size should receive. The model wrote it with total confidence. Pri's pipeline ran perfectly. The grammar is flawless.
Bina catches it in four seconds, because she knows roughly what a child's dose should look like.
That moment is the whole argument for domain expertise as a career-changer's strongest lever. A pure coder reviewing that note has no reason to flinch; nothing about it looks broken. You can only catch a wrong medical answer if you know medicine, a wrong legal answer if you know law, a wrong lesson plan if you've taught. The AI supplies fluency; you supply the truth check.
The AI can sound right about anything. Only someone who knows the subject can tell when it actually is.
Bina flags the draft, loops in Dr. Soledad Marchetti at the pilot clinic to confirm the safe range, and works with Marc Devlin, the AI product manager, on a guardrail so Scribe stops doing this. No one on that thread wrote code. They applied judgment, domain knowledge, and clear communication, and kept a confident hallucination away from a real child.
Bina has been at Trellix five days. Watch the category come to life.
Monday — she applies AI. She runs a batch of test patient messages through Scribe and reads every reply the way a clinic would. No code; just careful reading against what a good reply looks like.
Tuesday — she judges the output. She hits the pediatric dose error. She doesn't fix the model. She flags the failure, documents it clearly, and pulls in Dr. Marchetti to confirm the safe range. This is evaluation, the beating heart of the product.
Wednesday — she guides the behavior. With Marc, she rewrites part of Scribe's instructions so it never states a dose without flagging it for human confirmation. That's prompt work — clear writing aimed at the model instead of at students.
Thursday — she helps others apply it. She joins Theo on a clinic call and turns a doctor's worry ("will this thing make things up about my patients?") into a concrete answer about how review works. Enablement.
Friday — she adapts. A new model version ships overnight, so Bina re-runs her test set to see what changed. The field moves; standing still isn't an option.
One person, one week, every face of the role family: apply, guide, judge, enable. Not a line of Python in sight, and indispensable to the product.
Open any AI chatbot and ask it a question in a field you know well, like your old job or a hobby. Read the answer like Bina reads a Scribe draft: does it sound polished? Now hunt for the one thing that's subtly off or oversimplified. Write down what you caught and how you knew. That feeling, the "this sounds right but isn't quite" itch, is the core skill of an AI-adjacent role, and you just used it.
Preparing your quiz…