New jobs created by AI: role map, PDLC and how to pass the interview

“AI will replace jobs” is a useless conversation: it produces anxiety and nothing else. A more practical frame: inside companies the process changes, the task list changes with it, and roles appear that did not exist three years ago. Those are the roles being hired right now.

What changes inside a company

The classic product development life cycle repeats the same steps: understand the problem, design a solution, build it, test it, ship it, measure the effect. AI enters every step, but differently.

• Discovery. Interview synthesis, feedback and ticket analysis, hypotheses — faster, but with new work: checking whether the model invented a pattern.
• Design. Mockups and copy appear in minutes, so value shifts to problem framing and quality criteria.
• Build. Code is written faster, but review, tests and reading generated code take a larger share of the time.
• Test. A separate discipline emerges: model quality evaluation — test sets, metrics, manual and automated runs.
• Launch and measure. Someone has to track cost, watch quality drift and explain to the business what actually changed.

New professions grow out of this shift: the work does not disappear, it is redistributed — part of the routine moves into the model, and the freed time moves into framing, verification and explanation.

The roles that appeared

AI / LLM engineer

Builds working scenarios on top of models: model choice, context, tools, caching, error handling. Unlike a classic ML engineer, this role usually integrates existing models rather than training new ones.

Evaluation engineer

Builds the check system: which cases must work, how to measure regressions, what an error costs. Essentially a tester who tests model behaviour instead of code.

AI product manager

Decides where AI is appropriate at all: estimates effect, risk and cost, works with legal and security, and says no to impressive demos that solve nothing.

Data curator

Owns input quality and labelling: what goes into the context, what must never be used, how correctness is verified. Often underestimated, yet it sets the quality ceiling.

AI Ops / MLOps

Operations: model versions, monitoring, request cost, latency, rollbacks. Without it a pilot falls apart the moment it becomes a product.

Adoption and enablement lead

Helps people rebuild their work: guidelines, training, error reviews, change support. Technically simple and the most conflict-prone part of any rollout.

The skills that actually get tested

Job ads are full of tool names, but interviews test something else: whether you can carry a scenario to a result and explain its limits.

1. Systems thinking. Can you see the data path: where the request comes from, what enters the context, where the answer is stored, who owns the failure.
2. Comfort with uncertainty. The stack changes every six months; interviewers look for how fast you learn, not which library you memorised.
3. Defining criteria. “Works well” is not a criterion. People who propose measurable checks stand out.
4. Economics of the solution. Request cost, latency and volume are part of the engineering decision, not “a question for finance”.
5. Security and data. What must never be sent to a model, how personal data is handled, where the legal boundary sits.
6. Communication. Explaining to a manager and a team why the solution works and where it breaks.

How to pass an AI role interview

1. Prepare stories, not answers

You will almost always be asked about a real case. Structure it in four parts: the task and constraints, what you did, what happened, and what you would do differently. Without the last part the answer sounds like a report, not experience.

2. Expect the “bring AI into this process” case

A typical prompt: “support gets 3,000 tickets a day, what would you propose”. A strong answer starts with questions, not with a model: what volume, what does an error cost, who verifies the result, what data exists. Then a narrow pilot and a metric.

3. Name the limitations before you are asked

Interviewers value honesty about where the solution fails: hallucinations, stale data, cost, legal risk. A candidate who states the boundaries looks more reliable than one who promises everything.

4. Do not argue with terminology

If a question is loosely worded, restate it in your own words and move on. They test thinking, not vocabulary — debating whether “prompt engineer” is a real job is the worst use of interview time.

In these interviews the winner is not the person who knows more tools, but the one who clearly explains the task, the constraints, and how the result will be verified.

Preparation checklist

1. Write three stories in “task — action — result — lesson” format.
2. Describe one AI rollout along the PDLC steps and sketch a metric.
3. Define five quality checks for an imagined scenario.
4. Estimate cost: daily requests × price × headroom.
5. Name three risks and three ways to reduce them.
6. Say it all out loud — in English, and in the local language if the role is local.
7. Rehearse the conversation with an AI coach: it asks follow-ups and shows where your answer gets vague.

New professions look intimidating because of the titles, but the requirements are down to earth: understand the process, define criteria, carry it to a result and explain it to people. These skills are trainable — including in conversation: the more often you say your story out loud, the calmer the real interview will be.