How to Run an AI Mock Interview With a Scoring Rubric

An AI mock interview is useful when it behaves like a controlled rehearsal, not an answer generator. Give the model a verified role brief, a fixed set of competencies, rules for asking one question at a time, and a scoring rubric defined before the session. Finish the interview before requesting feedback, then require every critique to point to the transcript.

This separation matters. During the mock interview, the AI should create realistic pressure and follow-ups without rescuing the candidate. Afterward, it should act as a constrained reviewer. The candidate still decides whether the feedback is accurate and practices the revision aloud.

Keep this workflow in preparation unless the real employer explicitly permits AI assistance. As of July 23, 2026, McKinsey's candidate guidance gives a clear example of that boundary: it encourages AI for practicing interview questions but prohibits using AI to misrepresent experience or generate real-time interview responses.

What should an AI mock interview measure?

A mock interview should test whether you can understand a question, select relevant evidence, explain your own actions and reasoning, and communicate a supported answer under realistic constraints. It should not measure how quickly a model can rewrite your response.

The U.S. Office of Personnel Management defines a structured interview as a job-related assessment built around predetermined questions and consistent rating standards. A candidate cannot reproduce an employer's private process, but the same design principles make practice more useful: choose job-relevant competencies, keep the core questions stable, and compare answers against written behavioral anchors.

CareerOneStop, a service sponsored by the U.S. Department of Labor, similarly recommends researching the employer and role, practicing answers on paper and aloud, using real examples, and preparing questions. An AI mock interview should exercise those prepared inputs rather than replace them.

Define one objective for each session. Examples include:

• answer six behavioral questions with specific evidence in 35 minutes;
• complete a recruiter screen without losing the thread when compensation or availability comes up;
• explain a technical decision and handle two changing constraints; or
• deliver concise answers in a one-way video format without immediate feedback.

If the objective is merely “get better at interviewing,” the model has no stable basis for selecting questions or judging improvement.

Build a verified candidate packet

Create a small source packet before generating questions. Use information you can verify:

1. Role brief: the job description, level, location, and interview format confirmed by the recruiter or employer.
2. Competency list: four to six capabilities named in the posting or employer guidance, such as judgment, collaboration, customer focus, or a specific technical skill.
3. Evidence bank: your real projects, decisions, actions, results, and lessons, with uncertain details marked as uncertain.
4. Constraints: the planned duration, answer length, permitted materials, and any format-specific instructions.
5. Candidate questions: a short list of researched questions you may ask the interviewer.

Four to six competencies is a practical ceiling for one general session. OPM says structured interviews commonly assess four to six competencies, although the real number varies by job and interview length.

Do not ask the AI to infer secret evaluation criteria, invent a company's interview loop, or fill gaps in your work history. Microsoft's current interview tips tell candidates to understand the role, prepare examples tied to stated competencies, explain their thinking, and use the STAR(R) structure for specific answers. Those are observable preparation tasks; guessing an interviewer's hidden preference is not.

Remove confidential employer documents, customer information, credentials, and unnecessary personal data from the packet. If you plan to upload a resume or retain a transcript, use the AI interview assistant privacy checklist to inspect what the product captures, sends, stores, and lets you delete.

Define the questions and scoring rubric before the session

Ask the AI to draft more questions than you need, then review them before the mock interview. Reject questions that are generic, compound, unrelated to the role, or impossible to answer from the supplied information. Keep the final core set fixed for a baseline and a later comparison round.

Use a small rubric with behavior anchors instead of asking for an unexplained score out of ten:

Add one role-specific dimension only when the source material supports it. For example, a systems role might evaluate whether an answer identifies reliability tradeoffs, while a customer-success role might evaluate discovery and expectation management.

This is a practice rubric, not a prediction of the employer's score. OPM's structured-interview guidance says customized scales require defined proficiency levels, scoring rules, and subject-matter expertise. Unless the employer publishes its rubric, an AI-generated scale cannot credibly reproduce it.

Configure the AI as interviewer first and evaluator second

Use separate phases so feedback does not leak into the exercise. A starting instruction is:

Run a mock interview using only the supplied role brief, competencies, evidence bank, and rubric. Ask one approved core question at a time. Ask at most two concise follow-ups when an answer is incomplete, ambiguous, or missing evidence. Do not coach, suggest an example, finish an answer, reveal the rubric, or provide feedback during the interview. If the source packet does not support a company-specific claim, say that it is unknown. Keep the clearest available transcript and mark uncertain transcription. When I say “end interview,” stop asking questions and wait for a separate evaluation request.

The phrase “using only” narrows the source material, but it does not guarantee compliance or accuracy. Review the generated questions and run a short dry test before the scored session.

Follow-ups should expose an answer, not improve it. Useful probes ask what you personally did, what evidence informed the decision, what alternative you rejected, what result was observed, or what you would change. Leading prompts such as “Was stakeholder alignment the main challenge?” supply content and make the rehearsal easier than the interview.

For behavioral practice, prepare the underlying examples with the behavioral interview story-bank workflow. The mock interviewer should retrieve and test those stories through questions; it should never create a new achievement for you.

Run a baseline round without rescue

Match the real format as closely as practical: same duration, audio or video mode, note-taking limits, response timer, and transition between questions. OPM's Structured Interview Guide recommends pilot testing questions in a trial that mirrors the intended interview so unclear wording and weak prompts can be found before use. For a candidate, the analogous goal is realism, not imitation of a confidential employer process.

During the baseline:

• answer aloud rather than silently composing text;
• do not restart a weak answer unless the real format permits restarts;
• ask for clarification when the question is unclear;
• let silence remain instead of asking the AI for a hint;
• mark genuine tool or transcription errors separately; and
• finish the full round before opening the rubric.

If the real interview is virtual, test the microphone, camera, meeting application, and any permitted workflow before the scored round. The AI interview assistant preflight checklist covers permissions, audio, capture behavior, screen sharing, and recovery using synthetic material.

Save the prompt version, question set, rubric, duration, and transcript together. Otherwise, a later score can change because the AI used different questions or standards rather than because your answer improved.

Demand transcript-grounded feedback

After the interview ends, evaluate one question at a time. For each rubric dimension, require four fields:

1. Observation: the exact transcript passage or timestamp.
2. Rubric anchor: Missing, Partial, or Supported.
3. Effect: why the observed answer helped or weakened clarity.
4. Next attempt: one concrete behavior to practice, not a replacement script.

Then ask the evaluator to separate facts from inference. “The answer never states a result” is observable. “The interviewer would reject this candidate” is an unsupported prediction. “You sounded unconfident” is too vague unless the transcript or recording shows a specific pattern such as repeated qualifiers, an abandoned sentence, or a long pause.

The National Institute of Standards and Technology identifies confabulation as a generative-AI risk: models can confidently produce erroneous content, inconsistent logic, or false supporting citations. That means a polished critique is not evidence. Check every finding against the recording, transcript, source packet, and rubric before accepting it.

Review the round in three passes:

Pass one: question coverage

List each material part of the question and where the answer addressed it. Flag omissions, but do not grade style yet.

Pass two: evidence and reasoning

Verify that every project, action, number, outcome, and lesson came from the candidate packet or the spoken answer. Identify where reasoning skipped an assumption, alternative, tradeoff, or causal link.

Pass three: delivery

Measure answer length, structure, repeated phrases, unclear references, and whether the conclusion arrived before time expired. Use audio or video only with consent and a clear retention plan; text alone cannot support claims about eye contact, voice, or facial expression.

Improve one behavior, then rerun the same test

Choose one or two changes from the review. Rewrite only a compact outline, then answer the question aloud again without reading a script. Examples of useful changes are:

• lead with the decision before explaining background;
• replace “we” with your specific responsibility;
• name the rejected alternative and the constraint that ruled it out;
• state a bounded result without inventing precision; or
• stop after the reflection instead of repeating the conclusion.

Run a comparison round with the same core questions, time limit, rubric, and prompt version. Add new follow-ups only after the fixed questions are complete. This controls enough of the exercise to make the before-and-after transcript meaningful.

Keep a simple change log: question, baseline anchor, practiced behavior, comparison anchor, and remaining uncertainty. Do not average all rubric dimensions into one “interview readiness” number. A single score hides whether the candidate lacks evidence, misses questions, or simply needs a tighter delivery.

Specialized formats need additional exercises. Use the AI case interview practice protocol for exhibit handling, calculations, and synthesis; do not stretch a general behavioral rubric over case-solving work.

Know what an AI mock interview cannot tell you

An AI mock interview cannot know the employer's private question bank, predict a hiring decision, verify an experience that exists only in your notes, or reliably judge human rapport. It may ask unrealistic follow-ups, reward formulaic answers, miss domain errors, or change its standards between runs.

Use a human reviewer for high-stakes calibration when possible, especially someone who understands the role. Give that reviewer the same transcript and rubric before showing the AI's findings. Agreement on a specific observation is more useful than agreement on a global score.

Do not record another person without permission, upload confidential recruiting material, or reuse questions obtained in violation of an employer's rules. Before the real interview, read the employer's current instructions. Preparation with AI does not imply permission for AI, transcription, recording, notes, or outside assistance during the actual session.

Turn the mock interview into a repeatable practice loop

A useful AI mock interview has a verified source packet, fixed job-related questions, a behavior-anchored rubric, a no-coaching interview phase, and a transcript-grounded review phase. Run a baseline, correct one observable behavior, repeat the same test, and keep only feedback you can verify.

If you want to test a permitted desktop workflow after the practice method is stable, download Control and use synthetic material with the free message and voice allowance first. If the employer does not clearly allow assistance, keep Control in the rehearsal and bring only your improved answers into the real interview.