avery

How Do You Use AI in Your Work? Interview Answer Guide

This question is now asked in almost every interview loop, in almost every function, including ones with no obvious technical component. Recruiters ask a soft version of it on the phone screen. Hiring managers ask a hard version of it in round two. Panels ask it sideways, as "how has your work changed in the last two years."

The standard advice is to name a tool, name a workflow, and name a metric. That advice is fine and it is also incomplete, because it optimizes for the first sentence of the answer. In real loops, the first sentence is not what decides anything. What decides it is the follow-up, and the follow-up is almost always one of three questions that most candidates have never rehearsed.

This guide gives you a way to grade your own answer before the interview, the three follow-ups to prepare, and full worked answers for people at different levels of actual AI use, including people who have barely used it at all.

What the question is really testing

Interviewers are not checking whether you can list tools. Tool lists are worthless as signal because everyone has the same list.

They are checking three things, and they are checking them in this order.

Judgment. Do you know what this technology is bad at? Someone who describes AI as uniformly great has not used it seriously, because anyone who has used it seriously has been burned by it. The fastest credibility gain in this entire answer comes from naming a limitation you personally hit.

Verification. When the output is wrong, how do you find out? This is the question that separates people who use AI from people who forward AI output. In regulated, financial, legal, clinical, and safety-adjacent work, this is the whole ballgame.

Ownership. When AI-assisted work goes out under your name and it contains an error, whose error is it? The correct answer is yours, said without hesitation. Any hedging here reads badly.

Almost every failed answer to this question fails on one of those three, not on tool knowledge.

The AI Answer Ladder

Here is a way to grade your own answer. Four levels. Say your answer out loud, then find it on the ladder. Most people are one level lower than they think.

Level 1: The tool list

"I use ChatGPT and Copilot pretty regularly. They save me a lot of time."

This is the floor. It contains no information. Every candidate that day gave a version of it. It is not disqualifying, but it is completely neutral, which in a competitive loop means it is a loss.

Level 2: The workflow

"I use Claude to draft first versions of client-facing summaries, then I edit them down. It probably saves me two hours a week."

Better. There is a specific task and a specific benefit. This clears the phone screen. It will not clear a hiring manager who is actually interested, because it only describes personal time savings, and personal time savings is the least interesting category of AI use.

Level 3: The judgment answer

"I use it heavily for first drafts and for pressure-testing my own analysis, and I do not use it for anything where I cannot verify the source. I learned that the hard way when it gave me a confident regulatory citation that did not exist and I nearly sent it."

This is where most strong candidates land, and it is enough for most roles. It has a use, a boundary, and a real failure that produced the boundary. It demonstrates all three of the things being tested.

Level 4: The systems answer

"The individual time savings matter less to me than what it changed about how the team works. We used to do competitive research once a quarter because it took a week. Now one person does a refresh in an afternoon, so we do it monthly, and the marketing plan actually responds to what competitors did last month instead of last quarter. The thing I had to build alongside it was a check step, because the first month it gave us two positioning claims that were not supported anywhere, so nothing goes into a deck now without a source link next to it."

This is the answer that gets remembered. It is not about a tool. It is about a decision that changed cadence, plus the control that was added to make the new cadence safe. Very few candidates give this answer, and it is available to more people than give it, because most people have a version of this story and describe it at Level 2.

The move from Level 3 to Level 4 is usually one question: what did this let us do that we could not do before, or could not do often enough? Time saved is Level 2. Frequency changed, quality raised, or a new capability unlocked is Level 4.

The three follow-ups that actually decide it

Prepare these. They come up constantly and they are where prepared answers fall apart, because the opening answer was memorized and the follow-up was not.

Follow-up 1: "What has it gotten wrong?"

Or: where has it failed you, when has it been unreliable, what do you not trust it with.

There is exactly one bad response here, and it is "not much, honestly." That answer tells the interviewer you have used it lightly and are describing it from marketing copy rather than experience.

Have one real, specific failure ready. Structure it in three beats: what it produced, how you caught it, what you changed afterward. The third beat is the one people forget and it is the one that matters, because catching an error once is luck and changing your process is judgment.

A working example: "It summarized a customer call transcript and attributed a commitment to the customer that they never made. It was a plausible sentence in the right voice. I caught it because I happened to have been on the call. What I changed is that summaries now get generated with timestamps against the transcript so anything in the summary can be traced back to a specific minute. If it cannot be traced, it does not go in the CRM."

Follow-up 2: "How do you check it?"

This is the verification question and it is asked most sharply in finance, legal, healthcare, and anywhere with an auditor.

Weak answer: "I always review the output." Everyone says this and it describes no mechanism.

Strong answer names the specific check and, ideally, notes that the check is proportional to the risk. "It depends what it is. For internal first drafts I just read them. For anything with a number in it, I recompute the number myself, because that is the failure mode I have actually seen. For anything client-facing, it goes through the same review it would have gone through if I had written it from scratch, since the review exists to catch errors and it does not matter who made the error."

That last clause is worth stealing. It reframes AI output as ordinary work product subject to ordinary controls, which is exactly how a hiring manager wants a new employee to think about it.

Follow-up 3: "What would you not use it for?"

The refusal question. This one is underrated and it is where you can differentiate in one sentence.

The answer should include at least one thing you refuse on judgment grounds rather than policy grounds. Saying "we are not allowed to put customer data into it" is a compliance answer, and it only tells the interviewer you can follow a rule. Saying what you would refuse even if it were allowed tells them how you think.

Categories that work: performance feedback for a person on your team, anything where the point is that a human paid attention, decisions that require accountability you cannot delegate, and first-pass thinking on a problem you do not yet understand, because outsourcing the first pass means you never build the intuition.

That last one is a strong, slightly contrarian position and it plays well: "I do not use it for the first pass on a problem I do not understand yet. If I let it frame the problem, I get its frame, and I lose the part of the process where I actually learn the domain. I use it after I have my own read, to attack that read."

Three worked answers by situation

Full answers, sixty to ninety seconds each. Adapt the specifics, keep the structure.

If you are a heavy user in a non-technical role

"Two ways, and they are pretty different. The smaller one is drafting. Client updates, internal briefs, the first version of a deck narrative. That saves me maybe four hours a week and it is not the interesting part.

The bigger one is that it changed how often we do things. Our competitive teardowns used to be quarterly because each one took a person the better part of a week. Now it is a monthly refresh in an afternoon, so the plan actually reacts to the market in something close to real time. That was a real shift in how the team operates, not a time saving.

The thing I had to add was verification. Early on, a teardown included two claims about a competitor's pricing that were confidently stated and not supported anywhere I could find. So now nothing goes into a deck without a source link beside it, and if there is no link, the claim comes out. That rule cost us some speed and it is the reason I trust the output enough to put my name on it."

If you are a moderate user

"Mostly two things. I use it to pressure-test my own thinking, so before I take a recommendation to my director I will ask it to argue the other side and see what I have not considered. That has genuinely caught gaps, maybe one time in four.

And I use it for the tedious middle of analysis. Cleaning up data structure, writing the formula I know exists but cannot remember, turning notes into a structured summary.

What I do not do is let it near the numbers themselves without recomputing. I had it produce a variance table that looked completely right and had one figure that was wrong by a factor of ten, and it was wrong in a way that would have passed a quick read. So I recompute anything I am going to present. It is faster than it sounds and it is the only reason I am comfortable using it on anything that goes upward."

If you have barely used it

Do not fake this. Fabricated AI fluency collapses on follow-up one, every time, because you have no real failure story and it shows immediately.

The honest answer works if it has three parts: the true constraint, what you have actually done, and a specific view about where you would apply it here.

"Honestly, less than I would like, and the reason is that we could not put anything with customer data into an external tool and we did not have an approved internal option until recently. So my use has been on the personal side, mostly research and drafting.

What I have done is spend time on where it would fit in my actual work. The clearest one is our intake process. We get about forty requests a week that come in as unstructured email, and someone spends most of a day categorizing and routing them. That is a classification problem and it does not need any judgment I would be worried about handing over, as long as anything ambiguous gets escalated rather than guessed.

What I would want to understand here is where the line sits, because the useful question is not which tool, it is which decisions the team is comfortable letting it touch."

That answer is often stronger than a mediocre Level 2 answer from a heavy user, because it demonstrates the thing being tested. Judgment about fit, not tool familiarity. More on the broader interview shift in how to prepare for AI interviews.

Five ways this answer goes wrong

Inventing a metric. "It made me 40% more productive" invites the question of how you measured that, and there is no good answer. If you have a real number, use it. If you do not, describe the change qualitatively and specifically. A vague true statement beats a precise invented one.

Enthusiasm with no boundary. Pure advocacy reads as inexperience. The people who have used these tools the most are the most specific about their limits.

Suspicion with no use. The mirror error. Leading with risk and never getting to a use case reads as someone who has decided not to engage. You can hold a critical view, but it has to be a critical view built from use.

Describing the tool instead of your work. Explaining what a model does is not an answer. The interviewer knows. They want to know what you did.

Missing the compliance dimension in a regulated role. If you are interviewing in healthcare, financial services, legal, or government, and your answer never touches data handling, that is a mark against you regardless of how good the rest was. One sentence on where the data lives is enough.

How to build your answer in twenty minutes

Do this once and it covers every version of the question you will get.

  1. List your last ten weeks of work. Not your job description. What you actually did.
  2. Mark where AI touched it. Be honest, including the small stuff.
  3. Find the frequency change. Of everything marked, which one let something happen more often, faster, or at all? That is your headline, not the biggest time saver.
  4. Write the failure. One real instance of a wrong output. What it was, how you caught it, what you changed. This is your follow-up one answer and it is the highest-value thirty seconds in the whole conversation.
  5. Write the refusal. One thing you would not hand over even if permitted, and the reason. This is your follow-up three answer.
  6. Say it out loud twice. Aim for ninety seconds. If it runs past two minutes it is a monologue and the interviewer stops listening around the ninety second mark anyway.

One more thing worth doing before the interview. If you are claiming meaningful AI fluency in a conversation, your LinkedIn profile should not be silent about it, because the interviewer very often has your profile open while you are talking. The gap between what you say and what your profile shows is noticeable. If you are not technical, that is still solvable, and how to show AI skills on LinkedIn without being technical covers exactly how.

The short version

Get off the tool list. Lead with what changed about the work, not what you saved. Have one real failure and one real refusal ready, because those are the two follow-ups that decide the answer. And if you have not used it much, say so and bring a specific view about where it fits, which is a better answer than a fluent one with nothing underneath it.

If you want to know whether your LinkedIn profile backs up the story you are telling in interviews, an Avery audit scores exactly that gap and tells you which fixes close it fastest. Start at heyavery.ai/pricing.

Ready to put this into action?

Get a personalized LinkedIn audit from Avery — built on 12,000+ real coaching sessions.

Get My Audit