We read every AI question from 1,720 developer interviews. Here's what to prepare.

"How do you use AI in your daily development work?"

It sounds like a question about tools. You name yours, say what you use them for, and stop.

That is exactly where many answers went wrong. We read the reviews of 55 answers to this kind of question. In 25 of them, the person named tools or tasks and never gave one real example.

We looked at 1,720 real interviews for developer and QA jobs. AI came up in 246 of them: 14%, or about 1 in 7. We read every one of those questions. They came in two very different kinds:

• AI as your tool. How do you work with it, and how do you check its work?
• AI in your product. How do you build something that uses it?

They need different preparation. For the first, bring one real task. For the second, bring a design you can explain.

Where this data comes from. We build MeetAssist, an AI assistant that people use during job interviews. Users can save their interview transcripts. We take only the interviewer's questions, rewrite them, remove all names, and publish them in a free interview question bank. Users can opt out.

This sample: 1,720 interviews for developer and QA roles, from 785 people, September 2025 to October 2026 (1,160 of them from July 2026 on). These are people who chose to use an AI helper, so it is not a random sample. We only hear the call, and we guess the role from the questions. The counts show what interviewers asked, not who got the job. About 1 in 5 of the interviews with AI questions were not in English.

A: "How do you use AI?"

Questions about AI as your tool came up in 140 interviews.

In 78 of them (56%), the interviewer asked only the simple version: do you use it, which tools, how in your daily work. In the other 62 (44%), they went deeper: how exactly you work, how you check the output, a real example, a time it was wrong.

So the simple question is often a door. Almost half the time, the interviewer walked through it.

The simple versions sounded like this:

• What is your AI tool of choice?
• Do you ask AI to explain code, or to write it for you?
• How much of your code do you write by hand, and how much with an agent?

When an interviewer named a specific tool or model, it was most often Claude (19 interviews), then Copilot (12), Gemini (8), Cursor (7) and ChatGPT (6).

The deeper questions were about how you work:

• When you give a coding agent a small feature, what context do you give it up front?
• When you write a prompt, what parts does it have?
• What does an "AI-friendly" repository mean to you?
• Have you written rules for an agent? Give an example of one rule.
• Pick one real example where you used AI to build, debug, refactor or test something. What did the AI produce, and what did you do yourself?

And about how you check its work. These ask for details you only know if you have really done it:

• If Claude writes a test and says it passes, how do you know it actually tested anything?
• When the agent confidently invents classes or endpoints that don't exist, what do you do?
• Do you just accept the first answer AI gives you?
• Do people still review the pull requests your agents write?
• Tell me about a time an agent added a subtle bug. How did you catch it?
• How would you reduce the AI slop?

One question came up in three interviews, with three different people, months apart:

"Describe a problem you worked on that AI couldn't solve without help. Why couldn't it solve it?"

Have an answer ready for this one.

Where answers fell short

For many answers we also have an automatic review: a short note on what the person said and what was missing. A model writes these notes. They are not the employer's opinion, and they are not hiring results. We only report counts, never individual answers.

We read all 55 reviews of answers to "How do you use AI?" and "Which tools?" questions:

• 25 of 55 named tools or tasks, but gave no real example.
• 17 of 55 said they always check what the AI gives them. Only 3 of them said what they actually check.
• 4 of 55 said AI makes them much faster, with no example behind it.

A typical answer of the first kind (our example, not a quote):

"I use Claude, Cursor, Copilot and ChatGPT. For coding, tests, docs and debugging."

The interviewer learns almost nothing from this. And "I always check it" leaves the obvious next question open: check what, and how? That is the test question above, in a different form.

Prepare a task, not a tool list

Answer the simple question as if the deeper one is already coming. Build it around one real task:

1. The task. What you built or fixed.
2. The context. What you gave the AI: which files, which rules, an example.
3. The steps. A plan first, then small changes you could review.
4. The check. What you looked at or ran yourself: the diff, the tests, the edge cases.
5. The miss. One time it was wrong, and what you changed after that.

Here is an example we wrote in that shape. It is made up, not a candidate's answer:

"Last week I added CSV export to our admin page with Cursor. I gave it our existing PDF export, the API route and the tests. I asked for a plan first. Then I let it make one small change at a time, so each diff was easy to review.

I checked it like a teammate's code. I read the diff, ran the tests, and tried the edge cases myself: an empty table, a comma inside a name, 50,000 rows.

It did get one thing wrong. It loaded all the rows into memory at once, and the big file failed. I asked it to stream the rows instead, and I added a test with a large file. Now I always ask for a plan before any code."

That takes under a minute to say. It answers "which tools", "how" and "how do you check" in one go, and it gives the interviewer something real to dig into.

Use your own task, though. Don't borrow a failure you never had. The follow-up questions will find out.

B: "You built something with AI"

The second kind came up in 94 interviews. At least 52 of them asked about the candidate's own AI work: a project on the CV, a feature they built.

Knowing the words helps. A large language model (LLM) is the model that writes the text. RAG (retrieval-augmented generation) means the system first finds the right documents, then gives them to the model to answer from. RAG was named in 16 interviews. But these questions went past "what is it". They asked what happens when the system is slow, expensive, careless with private data, or wrong.

Cost and speed

• How would you add an LLM to a live chat, where speed and cost matter?
• How would you change the design to cut LLM costs for a chatbot with 14,000 requests a day?

Be ready to say where your system calls the model, what each call does, and what you would change to make it cheaper or faster.

Private data

• How do you handle private data before you send it to an LLM?

Be ready to follow the data: what reaches the model, what must not, and what you check before sending it. Naming your model provider is not an answer.

Wrong answers, and answers that change

• Your RAG system finds the right document, but the model still gives a wrong answer. What happened?
• How do you test an LLM's answer when it is different every time?
• How do you know the LLM's answers are not getting worse over time?

Be ready to treat "finding the documents" and "writing the answer" as two separate steps that can each fail. And be ready to say how you judge an answer, and how you compare results after a change. Be honest about which tests you actually ran and which you would add.

The rest: opinions, and AI inside the interview

Some interviewers asked what you think:

• Will AI replace developers?
• Do AI tools really make developers faster, or slower?
• Have you coded before ChatGPT?
• You use AI a lot in your work. Isn't that a risk for us?

That last one shows the other side. Some teams worry about too much AI, not too little. Be ready to say where you don't use it.

And some interviews let you use AI on the task itself, and then asked about it:

• Did you use AI for this exercise? How did you make sure the result was correct?
• You get a ZIP file in Claude Code with no other context. Where do you start?
• How would you solve this with AI?

Who got asked, and how much

Role	Interviews	With at least one AI question
QA automation	135	23%
Full-stack	349	21%
Backend	693	12%
Frontend	162	8%
DevOps	106	8%

QA automation came out on top. Many of their questions were about tests written by AI. (We also had cloud and mobile interviews, but too few to compare.)

Often AI was a single question: in 131 of the 246 interviews, it came up once. But in 31 interviews there were four or more AI questions. When it goes deep, it goes deep.

Before your next interview

• Prepare one real task you did with AI: the context, the steps, your checks, and one miss.
• Know what exactly you check when AI writes code or tests.
• Know where you don't use AI, and why.
• If AI is on your CV, prepare the design of what you built: where the model is called, cost, speed, private data, and how you test the answers.

We keep full lists of real questions, with how often each one was asked: Generative AI, RAG, AI agents, and 78 more topics and roles in the interview question bank.

Have you been asked about AI in an interview? What was the question? Tell us in the comments.

---

How we counted. We searched the interviewers' questions, in all languages, for AI words: AI, LLM, RAG, ChatGPT, Claude, Copilot, Cursor, Gemini, agent, prompt, MCP (Model Context Protocol, a way to connect AI tools to other tools and data), embeddings and a few more. That found 536 questions. 48 were false hits, such as C++ vectors, monitoring agents, cursor pagination, or a company with "AI" in its name. A model sorted the rest into kinds, and we checked the result by reading every question. One interview can have several kinds. "Developer and QA roles" means backend, full-stack, frontend, mobile, DevOps, cloud, QA automation, QA testing and CRM/ERP developers. We left out AI and machine learning engineer roles, where AI questions are expected.