How to prepare for a job interview with AI

AI can ask practice questions, test whether an answer is clear and spot missing evidence. Use it as a rehearsal partner, not as a stand-in interviewer: give it a small fact set, answer in your own words and judge the reasons behind its feedback.

Prepare the source material first

Collect four short inputs:

1. The current job advert.
2. Your submitted CV and letter.
3. Three to five examples from your real experience.
4. Checked notes about the employer and role.

Do not upload confidential work documents. Remove addresses, identification numbers, references’ details, customer names and other information the exercise does not need. Check the service’s privacy settings and terms.

The University of Zurich describes interview preparation as knowing your application, researching the employer and preparing questions. Review its interview guidance.

Make an evidence bank

Start with real examples, not model-written stories.

EXAMPLE NAME:
Situation:
My responsibility:
Action I personally took:
Other people's contribution:
Result or what changed:
What did not work:
What I learned:
Facts I can quantify:
Details that are confidential and must stay out:

Choose examples that cover different skills: solving a problem, working with another person, handling a mistake, learning something and making a decision with incomplete information.

Use a controlled practice prompt

Act as an interview practice partner, not as the employer.

Use only the JOB ADVERT and CANDIDATE EVIDENCE below. Do not invent facts about
the role, employer or candidate. If information is missing, say so.

Ask one interview question at a time. After my answer:
1. Summarise the claim I made.
2. Identify which supplied evidence supports it.
3. Flag vague wording, unsupported scope or a question I did not answer.
4. Suggest one follow-up question.
5. Do not write a polished replacement unless I ask.

JOB ADVERT
[paste relevant parts]

CANDIDATE EVIDENCE
[paste your evidence bank]

Work through one question at a time. A follow-up tests your example better than a list of 30 questions.

Worked practice round

This fictional example was tested on 8 September 2026 with a GPT-5-based Codex agent. Another model—or another run—may respond differently.

Advert requirement: Coordinate process improvements across warehouse and customer-service teams.

Candidate evidence: Elias created a shared delayed-order list after the two teams used separate email threads. He proposed the fields, tested the list with two colleagues and ran a fortnightly review. He did not manage either team. After six weeks, both teams used the list; no reliable time-saving figure was measured.

Practice question: Tell me about a process you improved across teams.

First answer:

I led a major transformation that aligned warehouse and customer service, dramatically improving efficiency.

Useful feedback:

• “Led” may overstate ownership; the evidence says Elias proposed and ran the process, not that he managed the teams.
• “Major transformation” and “dramatically improving efficiency” are unsupported.
• The answer omits the original problem and the concrete change.
• Follow-up: how did Elias persuade colleagues to use the list?

Second answer in Elias’s own words:

Warehouse and customer service were tracking delayed orders in separate email threads, so neither team had a complete view. I proposed one shared list, agreed the fields with a colleague from each team and ran a review every two weeks. Both teams were using it after six weeks. We did not measure time saved, but the list gave each delayed order one visible owner.

The useful part of the feedback was specific: it found inflated claims and missing evidence. Elias wrote the final answer and did not invent a time saving.

Judge feedback instead of obeying it

Use a fixed rubric:

Criterion	Check
Relevance	Did the answer address the question asked?
Ownership	Is “I” used only for the candidate’s actions?
Evidence	Can each factual claim be traced to the evidence bank?
Clarity	Can a listener follow the situation, action and result once?
Brevity	Could one detail be cut without losing the proof?
Limits	Are uncertainty, team contribution and missing measures preserved?

Reject advice to sound grander, insert a number, claim a different motivation or guess what the employer “wants to hear.” OpenAI advises users to verify important output with reliable sources. Read the accuracy guidance.

Practise delivery without memorising a script

Turn the answer into four cue words, not a paragraph:

separate emails → shared list → two colleagues → visible owner

Answer again from the cues and let the wording change. A memorised paragraph is harder to adapt when the interviewer follows up.

For video interviews, test the equipment and setting you will use. UZH distinguishes live interviews from recorded formats and advises candidates to prepare the technical setup. See its interview-format guidance.

Ask the model to find gaps, not employer secrets

Good requests:

• Which advert requirement has no example in my evidence bank?
• Which sentence makes my role sound larger than the facts?
• What follow-up question would test this example?
• Which term might be unclear to someone outside my last employer?

Poor requests:

• Predict the exact questions this employer will ask.
• Tell me the answer that guarantees an offer.
• Invent a stronger leadership story.
• Score my chance of being hired.

Common mistakes

• Uploading more personal or confidential information than the task needs.
• Letting the advert and candidate facts blend into one block.
• Rehearsing polished text before checking the claims.
• Accepting a fit score with no transparent criteria.
• Practising only friendly questions.
• Memorising a script instead of understanding the example.
• Treating one model run as stable or authoritative.

Check an AI interview practice session

• ✓I prepared the advert, submitted application and real examples
• ✓I removed unnecessary personal and confidential information
• ✓The tool was told not to invent employer or candidate facts
• ✓I practised one question and follow-up at a time
• ✓I checked ownership, evidence, clarity and limits
• ✓I rejected unsupported numbers and inflated claims
• ✓I rewrote answers in my own spoken language
• ✓I prepared my own questions for the employer

First build examples with the general interview worksheet. If your application also needs editing, use the fact-checked AI application workflow.