How to prepare for an AI interview

What you are actually being assessed on

The mental model that helps: an AI interview is closer to a written exam marked against a rubric than to a conversation.

In a human interview, a lot of the signal is relational — rapport, warmth, whether the interviewer enjoys talking to you. Those things are genuinely, if unfairly, part of the outcome.

A well-built AI interview strips most of that out. What is left is the content: did you address the question, is your reasoning sound, did you consider trade-offs, is there specific evidence behind your claims. Warmth is not being measured. Neither is confidence.

This changes what preparation means, and it is good news for many candidates. If you have ever been passed over for someone more charming, this format is closer to a fair fight.

The five rules

1. Say the reasoning out loud

The most common way strong candidates underperform. You consider three approaches, dismiss two in a second and a half, and describe the third. A human interviewer might infer the process. An assessment cannot score what you did not say.

Make it explicit:

"There are three ways to approach this. Caching at the edge is fastest to ship but the invalidation problem is real given the data changes per user. A read replica handles the load but not the latency, which is the actual complaint. So I'd start by measuring where the time is going, because if it's a single N+1 query then the first two options are premature."

Same thinking either way. One version gets assessed.

2. Be specific, not polished

Specificity is the strongest signal available to you, and it is very hard to fake.

• Vague: "I improved the performance of our checkout."
• Specific: "Checkout was taking about four seconds at p95. Profiling showed most of it was a per-item tax lookup running serially — 30-odd calls. I batched them and added a short cache on the tax-rate table since rates change monthly. It came down to roughly 600ms. The trade-off was staleness, so I set the cache TTL to an hour and added an invalidation hook to the admin tool."

The second answer contains numbers, a diagnosis, a decision, a trade-off, and a mitigation. That is what a rubric is looking for. It also cannot be produced by someone who did not do the work.

3. Answer the question that was asked

Prepared monologues score badly. If you are asked how you would handle a specific failure mode and you deliver your rehearsed story about a project you are proud of, you have not answered, and a rubric will score it as not answered.

Read or listen carefully. Answer the actual question. Then, if relevant, add the example.

4. Use structure

Not a rigid formula, but a shape:

• Behavioural questions ("tell me about a time..."): the situation and its constraint, what you decided, why, what happened, what you would change. Real instance, not a general policy.
• Situational questions ("how would you handle..."): what you would establish first, the options, which you would choose and why, what would make you change your mind.
• Technical questions: what you understand the problem to be, your approach, the trade-off, and the failure mode you are most worried about.

That last clause — naming what worries you — is disproportionately valuable, because acknowledging a weakness in your own answer is something confident-but-shallow answers never do.

5. Treat trade-offs as the point

Almost every substantial interview question is really asking: do you understand that choices have costs?

"I'd use X" is a weaker answer than "I'd use X, accepting that we lose Y, which is fine here because Z." Every time you make a choice, name what it costs. This single habit lifts more answers than any other.

Preparing properly

Write out five or six real situations in detail. Not stories — material. For each: the constraint, the decision, the reasoning, the outcome, and what you would do differently. Twenty minutes each, written down. This is the highest-return preparation there is, because it converts hazy memory into retrievable specifics.

Map the requirements. Read the role requirements and, for each, note which situation demonstrates it. Any requirement with no match is a gap to think about in advance rather than discover live.

Practise thinking aloud. Genuinely unnatural at first. Record yourself answering one question and listen back for the steps you skipped.

Get your numbers. Rough is fine — "about 40% of traffic", "roughly 200 requests a second". Specific beats vague even when approximate, and vagueness where a number should be reads as absence.

Check the practical setup. Time, expected duration, format — text or video, whether you can pause, whether there is a retake policy. Ask if it is not stated.

Your rights, and what to ask

You are not simply a subject of this process.

You are generally entitled to know. In several jurisdictions you must be told in advance that an automated employment decision tool will be used, and what qualifications and characteristics it assesses.

You can usually request an alternative. A human-only route, or an accommodation if you have a disability that the format disadvantages. Ask in writing.

You can ask what data is collected, where it came from and how long it is kept.

Reasonable questions to ask the recruiter:

• What does the assessment measure, and how is it weighted?
• Is it text-based or video? Is anything inferred from video?
• Will I receive feedback either way?
• Is there a human reviewer, and can they override the result?
• Is there a retake policy?
• What happens to the recording or transcript, and for how long?

A good employer answers these easily. Reluctance is itself information about how they will treat you as an employee — and if a process scores your facial expression or perceived personality from video, you are entitled to be sceptical of it. That basis is contested, and it is worth knowing before you invest an hour.

What the CalHire assessment is like

Concretely, so you know what to expect:

• One assessment, not one per employer. A skills test plus a text-based interview, generated for your declared skills, producing a verified composite score — test 40 / interview 35 / role-fit 25 by default, adjustable by the employer per role.
• Text-only for the AI interview. No video, no webcam proctoring, nothing inferred from your face or voice. Video is human-only and only after you have consented to an identity reveal.
• The AI never learns who you are. Your personal information is stripped before any model sees your work, so your name, photo, age, school and employers cannot influence your score.
• Valid 90 days, retake after 30. One score, portable across every application, with a freshness badge.
• Nothing is auto-rejected. If you score below an employer's threshold you are flagged for human review, not removed. Integrity signals also route to a human, never an automatic fail.
• You always get a report card. Growth-framed, showing where you stood and the highest-value thing to improve — which makes the assessment useful even when the outcome is no.
• Free, always.

See for candidates for the whole journey and /features for how the anonymisation boundary works.

The summary: be concrete, make your reasoning visible, name the trade-offs, and answer what was asked. Those four habits are what a rubric rewards — and they happen to make you better in human interviews too.