What does an AI engineer do?
An AI engineer builds software that uses machine learning models, then tests, deploys and monitors it. No government job profile uses the title yet. This page uses the nearest ones instead: software developers at the Bureau of Labor Statistics, and data scientists at O*NET OnLine, plus Microsoft's published description of the role.
The BLS page on software developers, quality assurance analysts and testers, last modified August 27, 2026, says "software developers design computer applications or programs." It gives a median annual wage for software developers of $135,980 in May 2025. It projects employment to grow 10 percent from 2025 to 2035 and expects demand partly "due to the continued expansion of software development for artificial intelligence (AI)."
O*NET OnLine's profile for Data Scientists (15-2051.00) covers the model side of the work. It says the job applies "data mining, data modeling, natural language processing, and machine learning" to large datasets. Its tasks include to "test, validate, and reformulate models to ensure accurate prediction of outcomes of interest" and to "compare models using statistical performance metrics, such as loss functions or proportion of explained variance."
Microsoft's Azure AI engineer page describes the job in plain steps: requirements definition and design, development, deployment, integration, maintenance, performance tuning and monitoring. It says the engineer works with solution architects, data scientists, data engineers and other software developers, and should "be able to apply responsible AI principles." Microsoft has since retired that certification, though the role description is still published.
In Canada, Job Bank has no separate AI engineer profile. The nearest is NOC 21231, software engineers and designers. Its national summary, modified September 24, 2026, lists a median wage of $56.49 an hour and says the occupation "usually requires a university degree."
How were these AI engineer questions chosen?
They come from the duties in the O*NET profile and Microsoft's role description, matched to the question each duty tends to produce. They are drawn from the job's published duties, not from a survey of real interviews. Nothing here claims to be the most asked, or to match what any named company asks.
| Duty (O*NET and Microsoft) | Question it tends to produce | What the interviewer is checking |
|---|---|---|
| Requirements definition and design | "How would you build a tool that answers customer questions from our help docs?" | Whether you ask about users, data and the cost of a wrong answer before picking a model |
| Test, validate and reformulate models | "How would you know this model is good enough to ship?" | A test set, a metric tied to the task, and a simple baseline to beat |
| Compare models using performance metrics | "Model A scores higher offline, yet users prefer model B. What now?" | Whether you question the metric and look at real use |
| Deployment and integration | "Walk me through putting this model behind an API." | Latency, cost, versioning and a way to roll back |
| Monitoring and performance tuning | "How would you notice the model getting worse after launch?" | Logging, quality checks over time, and alerts with an owner |
| Apply responsible AI principles | "What would you do about a model that gives wrong answers with confidence?" | Risk thinking, human review and clear limits on use |
The last row is where many loops ask about risk. The NIST AI Risk Management Framework, released on January 26, 2023, is "intended for voluntary use" and is organised around four functions: Govern, Map, Measure and Manage. NIST added a Generative AI Profile on July 26, 2024. You don't need to name the framework in an interview, though the habit it describes, finding and limiting risk, is worth showing.
Which behavioral questions should you expect?
Behavioral questions ask about past work, so answer each one with a single real project in STAR order: situation, task, action, result. Interviewers listen for your own part in the work, how you judged whether a model was good enough, and what you changed after something went wrong. Keep the model details short unless they ask for more.
The STAR method guide shows the structure, and the behavioral interview questions guide covers the wider format.
1. Tell me about a model or AI feature that didn't work as expected after launch.
What they listen for: how you found out, how quickly you limited the harm, and what you added so the same failure would show up earlier next time. Blaming the data without saying what you did about it scores poorly.
Made-up example answer. Situation: our team launched a tool that sorted incoming support tickets by topic, and agents started complaining that billing tickets landed in the wrong queue. Task: I owned the model and the service around it. Action: I pulled a sample of misrouted tickets and found that a new billing product name had appeared after we collected our training data. I added a rule to send low-confidence tickets to a person, labelled a fresh batch of recent tickets, retrained, and added a weekly check on accuracy by topic. Result: misrouted billing tickets dropped back to the level we saw in testing, and the weekly check now warns us when new topics appear.
2. Tell me about a time you disagreed with a product manager or data scientist about how to measure success.
What they listen for: the evidence you brought, whether you understood their goal, and how the team settled it. A good answer ends with a metric both sides could accept.
3. Describe how you explained a model's limits to people who don't build models.
What they listen for: plain words, an example of a wrong answer the model might give, and what you asked them to do about it. Avoiding jargon matters more than covering every detail.
4. Tell me about a time you had to learn a new tool or technique quickly for a project.
What they listen for: how you learned it, how you checked your understanding before relying on it, and what you would do differently. This field changes fast, so interviewers want a learning habit they can trust.
What role-specific AI engineer questions come up?
Role-specific questions test the work itself: designing a system around a model, judging quality, shipping it and keeping it healthy. Interviewers check whether you clarify the problem first, choose a simple baseline, and say what could go wrong. Naming the tradeoffs you made, such as cost against quality, matters as much as the final design.
5. Design a system that answers employee questions from internal documents.
What they listen for: who asks, how sensitive the documents are, how you find the right passages, how you show sources, and what happens when no good answer exists. Ask about permissions early, because some employees should not see some documents.
Made-up example of thinking out loud: "First, who are the users, and are any documents restricted to certain teams? I'd split the documents into passages, index them, and fetch the most relevant passages for each question before the model writes an answer. Every answer would cite the passages it used, and if none score well I'd have it say it doesn't know. I'd check permissions at fetch time, so nobody sees a passage they can't open today. Then I'd build a test set of real questions with known answers before launch."
6. How would you evaluate a feature built on a large language model?
What they listen for: a test set drawn from real use, clear pass and fail rules, checks for the failures that matter most, and a plan to repeat the tests after every change. Say where you would use people to review outputs and where automatic checks are enough.
7. What is data leakage, and how do you prevent it?
What they listen for: a correct definition and a concrete habit. The scikit-learn guide to common pitfalls defines leakage as using "information that would not be available at prediction time" when building the model. It advises you to "always split the data into train and test subsets first, particularly before any preprocessing steps."
8. The model is accurate but too slow and too costly. What are your options?
What they listen for: a smaller model, caching repeated requests, shorter inputs, batching, or sending only hard cases to the larger model. Say how you would measure the quality you lose with each option.
9. How would you monitor a model in production?
What they listen for: logging inputs and outputs where privacy rules allow, tracking quality on a sample, watching for changes in the kind of input you receive, and alerts that reach a named owner. Microsoft lists monitoring and performance tuning as phases of the job, so expect detail here.
Which situational questions test judgment?
Situational questions describe a problem you have not faced yet and ask what you would do. Interviewers want an order of steps: protect users first, gather facts, tell the right people, then fix and follow up. Asking one or two clear questions before you answer often scores better than a fast plan that guesses.
The guide to situational interview questions explains how these differ from behavioral ones.
10. A customer reports that your AI feature gave them confidently wrong advice. What do you do?
What they listen for: limiting harm before you investigate, keeping the evidence, telling the right people, and adding a test for the failure.
Made-up example answer: "First I'd find the exact input and output in our logs and check whether other users got similar answers. If it could cause real harm, I'd turn off that feature or add a warning while we looked into it, and tell support and my manager what I had done. Then I'd add the case to our test set, find out why the model answered that way, and fix the prompt, the data or the guard around it. I'd rerun the full test set before turning the feature back on."
11. A product manager wants to launch on Friday, and your evaluation isn't finished.
What they listen for: a clear statement of what is known and unknown, a smaller launch option such as a limited group of users, and a decision made by the right person with the risk written down.
12. You're asked to train a model on customer data, and you're unsure the company is allowed to use it that way.
What they listen for: stopping to ask, going to the privacy or legal team, and not guessing. Interviewers want to hear that you would raise the question before any data is copied.
What follow-up questions tend to come next?
Follow-ups test whether your first answer holds up under detail. Expect the interviewer to change one condition, such as ten times the traffic or a stricter budget, or ask how you measured a result you mentioned. Prepare by asking yourself "how do you know?" about every claim in your stories.
- "What was your baseline, and how much better was the model?"
- "How did you build the test set, and who labelled it?"
- "What would you do if the cost per request had to drop by half?"
- "Which part of this did you build yourself?"
- "What would make you roll the model back?"
How should you prepare for an AI engineer interview?
Pick four or five real projects and write one STAR story for each, including one where a model let you down. Practise a system design out loud with a timer, ending with how you would test and monitor it. Read the posting again and match each listed tool and duty to a story or project you can describe.
The free interview question generator lets you paste the job and pick the round, and it returns likely questions, why they come up, and what the interviewer listens for. AI engineers often interview alongside other technical roles, so compare this page with the questions for a software engineer, a data scientist or a data engineer.
Before the loop, check that your resume bullets name results you can explain in detail. The guide to resume bullet points shows how. Prept's Mock room lets you answer out loud on a clock, scores each answer out of 10 and checks it against your resume, and it is included in the free plan.