How Do You Use AI in Your Work? Interview & Application Answers
Spoken or written, the answer needs four parts: the task, what the AI did, what you checked yourself, and the result. Includes 50, 100, and 150-word versions.

TL;DR: "How do you use AI in your work?" now arrives in two formats: a 60–90 second spoken answer and a 50–150 word application box. Both are graded on the same four parts: the task, what the AI did, what you checked or changed yourself, and the result. The third part is the one most candidates skip, and the one the question exists to test.
In 2026, this question is no longer only an interview question. It now appears in application forms as a free-text box, often with a second instruction attached: explain what the AI did, and explain what still required your direct input or judgment. That shift is part of a broader pattern: Greenhouse's 2026 candidate report found that 63% of job seekers have faced an AI interview, and most have not had a good experience with one. The application box is the same question with the clock removed.
A prompt in this shape showed up repeatedly in 2026 application forms:
"Please share a recent example of how you leveraged AI to improve efficiency or quality in your work. What problem were you solving, how did you use the tool, and what aspects still required your direct input or judgment?"
The wording varies. The three instructions do not: a problem, a tool action, and a judgment call. Most advice online answers only the spoken version, where you can read the room and recover with a follow-up. In a form, there is no room to recover. A hiring manager reads your paragraph next to everyone else's, and the judgment has to be visible without you there to explain it.
The two formats, side by side
| Dimension | Spoken interview | Written application |
|---|---|---|
| What they see | Your delivery, and how you handle a follow-up | A block of text, next to everyone else's |
| Length | 60–90 seconds | 50–150 words, or the box's limit |
| What wins | One story, told clearly, with a check you can defend | One story, compressed, with the judgment visible early |
| What loses | A tool list, or a rehearsed speech | A project tour with no result, or a generic AI-written paragraph |
The substance is the same in both. What changes is the order and the amount of setup you can afford.
The four-part answer: task, tool, judgment, result
Every strong answer contains four parts, in this order.
- The task and why it mattered. One sentence. What were you trying to do, and what made it worth doing? "I build a monthly competitor pricing brief for our sales team" is enough. You do not need the company's history.
- What the AI did. One sentence, one tool, one action. "I used ChatGPT to extract the pricing tiers from saved pages and draft the change log." Not a list of every tool you have opened.
- What you checked or changed. One or two sentences. This is the graded part. The draft looked right, and then you found the specific place where it wasn't. What did you compare it against? What did you correct, remove, or decide not to use?
- The result. One sentence with a number or a clear before-and-after. "The brief now takes two and a half hours instead of six." If you cannot measure it, describe the change: "The team stopped rewriting the same summary every month."

Written as a template, it looks like this:
I [task] for [who it was for]. I used [tool] to [specific action]. The first draft [what was wrong or generic], so I [check against an outside source] and [what you changed]. It now [result, with a number if you have one].
The template is deliberately plain. Its job is to keep you from spending 80% of the answer on the project and 20% on the part being scored.
Why part three carries the score
The form's second sentence usually asks what required your direct input or judgment. The interviewer already assumes AI produced a first draft; that is not the interesting part. What they cannot assume is that you noticed where the draft was wrong, generic, or unverifiable.
There is data behind that suspicion. In O'Reilly's 2026 survey of 2,000 UK workers, only 24% had confirmed an AI answer before acting on it, and 39% had already acted on one that turned out to be wrong. Employers are hiring for the 24%. An answer that describes the tool and the time saved, but never says what you checked, reads like the other 76%.
What "direct input or judgment" means in practice
Judgment is not a feeling, and it is not the sentence "I always double-check the output." In an answer, it shows up as a specific check against something outside the model. Pick the check that actually happened.
| If the AI produced... | The check to name | What it sounds like |
|---|---|---|
| Numbers, forecasts, market data | Recalculate from the source system | "I re-ran the totals against our CRM; two of the four figures were annualized, not monthly." |
| Code or a script | Tests, edge cases, a reproduction | "I ran the suite and added a case for the retry path the generated version skipped." |
| Customer-facing copy | Policy, legal, tone | "I checked the refund language against the policy page and removed a promise we don't make." |
| Research or a summary | Primary source, date, sample | "I traced the two claims I planned to use back to the filings; one was from 2024." |
| A plan or recommendation | Constraints, stakeholders, cost | "I cut two of the five steps because they assumed a team we don't have." |
A small, specific check beats a grand claim about verifying everything. "I checked every figure against the live pricing pages and corrected two" is more convincing than "I ensured accuracy at every stage," because the first sentence could only have been written by someone who did it.
The written version: 50, 100, and 150 words
Here is one story at three lengths. The details are ordinary on purpose; the structure is the point.
Say you are a marketing manager who used AI to build a monthly competitor pricing brief.
50 words
I used ChatGPT to draft our monthly pricing brief. It mixed annual and monthly prices and missed a new enterprise tier, so I checked every figure against the live pages, corrected two, and added the tier. The brief now takes about two and a half hours instead of six hours.
100 words
Every month I build a competitor pricing brief for our sales team. Before AI, it took about six hours across four competitors. I used ChatGPT to pull the pricing tiers from saved pages and summarize what had changed. The first draft mixed annual and monthly prices and missed a new tier, so I checked every figure against the live pricing pages, corrected two, added the tier, and left out a promotional discount that had expired. The brief now takes two and a half hours; sales used it in three renewal conversations that month. It is in the renewal prep packet.
150 words
Every month I build a competitor pricing brief for our sales team: four competitors, three tiers. Before AI, it took about six hours because most of the work was copying prices into a spreadsheet and rewriting the summary. I used ChatGPT to extract the tiers from saved pricing pages and draft the change log. The first draft looked clean, which was the problem: it had annualized two monthly prices and missed a new enterprise tier. I checked every figure against the live pages, corrected both errors, added the missing tier, and cut a promotional discount that had already expired. I also asked a sales lead to read it before it went out, because the model tended to describe every change as significant. The brief now takes two and a half hours. Sales used it in three renewal conversations that month; one rep caught a pricing gap before sending a quote.
What changes between the three versions:
- The 50-word version keeps the task, one error, the check, and the result. It drops the setup and the second check.
- The 100-word version adds why the task was expensive and who used the output.
- The 150-word version adds a second check and a consequence. Past 150 words, most answers add project background, not evidence.
If the box is shorter than 50 words, keep the error and the result. "I used AI to draft a pricing brief, caught two annualized prices and a missing tier against the live pages, and cut the monthly build from six hours to two and a half" is 33 words and still answers all four parts.
The same structure in two other roles
Software engineer (60 words)
I used Claude to draft a migration for 40 API call sites in our billing service. The patch compiled, but it skipped the retry logic on two endpoints and duplicated one helper. I reviewed every diff, added tests for both failure paths, and ran the full suite before merging. The migration took one day instead of three, with no rollback.
Customer support lead (51 words)
I used ChatGPT to draft macros for our twelve most common refund tickets. It invented a 30-day window that our policy doesn't offer, so I checked every macro against the policy page and removed that promise. First-response time dropped from nine minutes to four, and the macros became the team's default.
Neither example names a model version or a prompt. Those details feel specific but do not prove judgment. The check does.
Common application prompts, and what each one is really asking
The wording varies, but most 2026 versions fall into six families. Match the structure to the prompt instead of pasting the same answer into every box.
| Prompt family | What it is really asking | Lead with |
|---|---|---|
| "Share a recent example of how you leveraged AI to improve efficiency or quality..." | Problem, tool, judgment, result | The check: what you caught and corrected |
| "Describe how you use AI tools in your work today, or how you would apply AI to improve the role..." | Current practice plus role fit | One current example, then one sentence on the role |
| "We are looking for people who are passionate about AI... how you have put it to work..." | Sustained interest and initiative | Two areas of use, one shipped result |
| "We are looking for builders... a solution you've implemented in action." | A shipped solution and your decisions | Before and after, your role, what you would change |
| "Describe your experience using AI tools, automation, or AI-enabled technologies... level of hands-on involvement." | How hands-on you actually were | Whether you built, configured, or used it |
| "Tell me about a time you used an AI tool to complete a task. What did you use it for and what was the outcome?" | A compact story with a result | Task, tool, check, result in three sentences |
A few of these deserve their own note.
"Passionate about AI" rewards breadth, but breadth without evidence reads as a hobby list. Name two areas you have actually used AI in, then spend most of the words on one result. If your use is personal rather than professional, say so. A personal project with a real outcome is stronger than an inflated work claim.
"Level of hands-on involvement" is there to separate people who built a workflow from people who used a chatbot once. Be precise. "I built a prompt library and a review checklist for our team" is a different level from "I use ChatGPT for emails." Neither is disqualifying; vagueness is.
"A solution you've implemented in action" is not asking for a tool recommendation. It is asking what you changed, who used it, and what you would do differently. If the solution was small, say so and show the before-and-after. Small and shipped beats ambitious and hypothetical.
The spoken version: 60–90 seconds
Out loud, the same four parts, with two differences: you can read the room, and you will get follow-ups.
- Open with the task, not the tool. "At my last job I was spending six hours a month on..." gives the interviewer a reason to listen. "I use ChatGPT a lot" does not.
- Say the tool once. Then move to what you did with the output.
- Spend the most time on the check. "The draft looked right, but..." is the sentence that makes the story yours.
- End on the result, then stop. Do not fill the silence with a list of other tools.
Rehearse it twice out loud, not in your head. The first run is usually 30 seconds too long; the second is where you find the sentence you can cut.
Prepare a second story about quality rather than speed. The first answer usually proves you can go faster. The follow-up asks whether you can be trusted when speed is not the point. If you need a structure for that follow-up, our guide to catching an AI mistake covers the same four parts from the failure side.
When the interview tests AI use live
A third format has appeared in technical rounds: the interviewer watches you use an AI tool during the interview. Meta's AI-enabled coding round, rolled out in late 2025 and expanded through 2026, is the most-discussed example. Candidates get a CoderPad environment with an AI assistant and a codebase they did not write.
The scoring still includes problem solving and code quality, but two of the four competencies are verification and communication: whether you check the output, and whether you can explain what you are doing while you do it.
Three habits matter in that format:
- Narrate the plan before you prompt. Say what you are trying to produce and what you will check first. Some phases restrict the AI entirely, so read the instructions before you start prompting.
- Say what you are checking as you check it. Silence looks like copying, even when it isn't.
- When the tool is wrong, say so and show the correction. The recovery is the signal.
If you sit silently and paste, the interviewer cannot see the judgment they are scoring. For the format basics, our guides to HackerRank's AI-assisted interview mode and Google's AI interview policy cover what different platforms currently allow.
Follow-up questions to expect
Interviewers rarely stop at the first answer. The follow-ups are where they test judgment:
- "How do you decide when not to use AI?" This is checking for a boundary, not a rule against AI. Our guide to when you would not use AI has the structure for that answer.
- "How do you verify what it gives you?" Answer with the specific check you used in your story, not a general habit.
- "Tell me about a time AI got something wrong." Use the same story if it fits, or a second one. The structure is here.
- "What would you do differently next time?" This is a judgment question in disguise. Name one change to the process, not a change to the tool.
If the conversation moves toward agents and delegation, how you work with AI agents is the adjacent question, and it is scored differently: there, the interviewer is looking for how you scope a task, not just how you check an output.
Answers that backfire
Five patterns come up again and again, and all five are avoidable:
- "I don't really use AI" or "I use it for everything." Both are non-answers. The first reads as resistance; the second gives the interviewer nothing to evaluate.
- A tool list with no outcome. "I've used ChatGPT, Gemini, and Copilot" answers a different question than the one being asked.
- A project tour with no result. In a written form, this is the most common failure: 300 words about the project and one clause about AI.
- A number you cannot defend. If you say 40%, be ready for "how did you measure that?" A smaller, verifiable number is stronger than a round one you invented.
- An answer that sounds generated. In a written form, this is now a real risk: generic transitions, no specifics, every sentence the same length. Write the draft yourself, then use AI to tighten it, not to invent it.
Using AI to edit your answer is normal. Using it to invent an example is the one mistake that can end the process, because the follow-up will expose it.
How AceRound helps with this question
For the spoken version, AceRound's mock interview can run this question and the follow-ups, so the 90-second answer is something you have said out loud before it counts. In a live round where AI use is allowed, Interview Copilot surfaces structure and key points while you direct the conversation; it does not type the answer for you. For a written application box, the four-part structure above is the tool. The words should be yours, because the interview that follows will ask about them.
If you are preparing several AI-literacy questions at once, behavioral interview questions about AI maps the cluster and shows where this one sits.
FAQ
What is this question really asking?
Whether you can direct an AI tool and still own the output. The interviewer or the form is not checking which tools you have heard of; they are checking whether you can name a task, describe what the tool did, and show what you checked or changed yourself. The strongest answers spend more words on the third part than on the first two.
How long should the written answer be?
Follow the box's limit first. If there is no limit, 100 to 150 words is enough for one story: the task, the tool, the check, and the result. A 50-word box can still work if you keep one specific error and one number. Anything past 200 words usually turns into project background.
How do I answer 'what aspects required your direct input or judgment'?
Name a check against something outside the model, and say what it changed. For example: 'I re-ran the totals against the CRM and found two prices had been annualized,' or 'I checked the refund wording against our policy page and removed a promise we don't make.' 'I always review the output' is not an answer; it describes a habit without evidence.
What if I don't use AI at work yet?
Use the closest real substitute: a course project, a job search task, or a personal workflow, and say plainly that it was not part of your job. Then spend the answer on what you checked. Interviewers are testing judgment and curiosity, not whether your employer has adopted AI. A made-up work example is worse than an honest smaller one.
Can I use the same answer in the interview and on the application form?
Use the same story, not the same wording. Spoken, you can pause, adjust, and take a follow-up. Written, the reader sees only the text, so the judgment has to appear earlier and the result has to be explicit. Compress the spoken version to 100 to 150 words and cut the setup first.
Should I name specific tools?
Name the tool once, then move on. 'I used ChatGPT to extract the pricing tiers' is enough; a list of five tools adds no evidence about your judgment. If the tool is unusual or internal, say what it does in a few words. The exception is a prompt that explicitly asks which tools you use; then answer the question directly and return to the example.
What if the form asks 'Did you use AI to complete this application?'
Answer directly, then explain the boundary. If you used AI to organize your thoughts or tighten a sentence, say that; if you did not use it for the substance, say that too. Do not claim you wrote every word yourself if you didn't, and do not describe using AI to invent experience. The follow-up question is usually about what you would not delegate.
Do some companies now test AI use live instead of just asking about it?
Yes. Some technical rounds now give candidates an AI tool and watch how they use it. Meta's AI-enabled coding round, rolled out in late 2025 and expanded through 2026, is the most-discussed example. The interviewer is scoring how you direct the tool, check the output, and recover from a wrong answer, alongside the usual problem-solving and code-quality signals.
Author · Alex Chen. Career consultant and former tech recruiter. Spent 5 years on the hiring side before switching to help candidates instead. Writes about real interview dynamics, not textbook advice.
Related Articles

"Describe a Situation Where You Identified an Issue in AI Output" — Answer Guide
Two things get graded: what the AI got wrong, and the steps you took to fix it. A four-step structure, three worked examples, and a 90-word form-box version.

Xinterview AI (X Interview): What the Recorded Interview Is and How to Prepare
Xinterview AI sends candidates a link to record answers on their own time. What the timers, retakes and AI scoring actually do, plus how to prepare.

Big Interview Alternative: What $39 Buys, and Where It Stops
Big Interview charges $39 for 30 days, $99 for three months and $299 for life. What those plans include, what they leave out, and the one-time alternative.