To create an AI interview, turn a clear role profile into a fixed interview guide, permitted follow-up questions and a scorecard that quotes the candidate. The AI then runs the first conversation by chat, voice or video, or hands you the guide for a human-led interview. Criteria, approval and the decision about what happens next stay with an accountable person, and from 2 December 2027 the EU AI Act’s high-risk rules apply as well.
What is an AI interview, and what types exist?
An AI interview is a structured conversation in which an AI asks the questions, records the answers and summarises them against fixed criteria. It does not replace the conversation with the hiring manager. It clarifies what you need to know before inviting someone anyway: availability, language level, concrete experience. Structured interviews rank among the most predictive selection methods in personnel-selection research (Sackett et al., Journal of Applied Psychology, 2022). An AI interview applies that structure consistently: same questions, same scale, every rating backed by evidence.
| Type | Who runs the conversation? | Typical use |
|---|---|---|
| AI as guide author | A person, with an AI draft of questions and scorecard | Hiring-manager interviews, small teams |
| Asynchronous chat interview | AI, written answers, any time | High volume, shift work, availability check |
| Voice or phone interview | AI, spoken, with transcript | Customer contact, language level, care, logistics |
| Video interview | AI avatar or recorded | Office and specialist roles with a presentation part |
What do you need before creating an AI interview?
This guide covers two approaches: using a general language model to draft a guide for a human interviewer, or configuring a platform for an asynchronous chat, voice or video interview. The quality logic is the same in both cases.
Do not start from the job advertisement alone. Prepare a concise role profile that separates essential requirements, learnable capabilities and mere preferences, plus real work situations in which the relevant skills become visible.
| Building block | Useful input | Weak input |
|---|---|---|
| Role outcome | What should the person deliver after onboarding? | A generic task list |
| Essential criterion | An observable requirement with a reason | “Top talent” or “culture fit” |
| Work sample | A real decision or typical situation | A puzzle unrelated to the job |
| Evidence | A concrete example, approach and result | A numeric self-rating without an example |
| Exclusion | A legally and professionally reviewed minimum condition | An unverified assumption about personality |
The topic hub on AI interviews and voice recruiting places guide, channel and process step in context.
Step 1: define the purpose of the interview
Write down in one sentence which decision the interview prepares. A good goal: “The first interview clarifies shift availability, German skills for customer conversations and two concrete examples of de-escalation.” A poor goal: “The AI finds the best candidates.”
The purpose limits questions and data collection. A short screening should not pretend to be a full aptitude assessment. It collects what the next human step is missing.
Step 2: translate requirements into observable criteria
Phrase each criterion so that several reviewers rate the same answer similarly. “Strong communicator” is too open. “Explains the next step to an angry customer clearly, sets a limit and records the commitment” is observable.
Rule of thumb from practice: four to six criteria per role, three to four for a short first interview. If everything is important, the scorecard becomes arbitrary. Mark each criterion as essential, development area or additional evidence.
Step 3: design core questions with the same starting point
Every candidate gets the same core questions (more on this in building an interview guide). That makes answers comparable and stops spontaneous sympathy from deciding how deep the conversation goes. Good questions ask for a concrete example or a traceable decision.
- Experience: “Describe a situation in which …”
- Approach: “How would you handle the following case …?”
- Result: “How did you know your approach worked?”
- Reflection: “What would you do differently today?”
Avoid questions that give away the desired answer. Questions about protected characteristics and private circumstances do not belong in an automated guide.
Step 4: define permitted follow-up questions
A good AI interview is neither a rigid form nor a free conversation. Define a small fixed set of follow-up types: ask for an example, clarify the person’s own role, ask for the result or lesson. Limit repetitions and duration.
Also define stop and handover rules. If a person reports a technical problem, does not understand a question or talks about sensitive personal information, the system must not improvise. It needs an alternative channel and a human contact.
Step 5: create an evidence-based scorecard
A scorecard separates criterion, observed evidence and rating. It should never output a single number. Reviewers must be able to jump back to the relevant sentence in the transcript.
| Level | Description | Example for customer de-escalation |
|---|---|---|
| Not enough information | Question not understood or not asked | Answer breaks off, question repeated later |
| No evidence | No assessable answer | No example or irrelevant answer |
| Partial evidence | Some relevant indications | Calms the customer but names no next step |
| Solid evidence | The criterion is demonstrably met | Clarifies the issue, explains the next step, documents it |
| Strong evidence | Evidence covers approach, result and reflection | Adds a clear limit, ownership and a checked result |
These levels are not a universal template; make them concrete per role and criterion. An overall score must never hide an essential criterion: someone who cannot show a required work permit is not “compensated” by points elsewhere.
Step 6: use a controlled prompt
A general language model works for a first guide. Do not enter real applicant data until use, contract and data protection are settled. A useful prompt contains goal, role, criteria, output format and clear prohibitions.
Prompt template:
You are an interview designer. Create a structured first-interview guide for [role]. Purpose: [decision]. Essential criteria: [list]. For each criterion write one open core question and a small fixed set of neutral follow-ups. Create a scorecard with clearly named evidence levels, visible evidence anchors and “not enough information” as its own option. Do not ask about private or protected characteristics. Do not make a hiring decision. Output questions, follow-ups, scorecard and review notes separately.
Check every output professionally. A model can misread requirements, produce duplicate questions or suggest criteria that are irrelevant for the role.
Step 7: configure channel and candidate experience
On a platform you choose between web form, chat, video, voice, messenger or phone. The best channel is the one your audience can use reliably: browser video can work for office roles, while phone or messenger is often more accessible for non-desk roles.
The invitation explains purpose, expected duration, channel, evaluation, next step and contact person. Offer an alternative for technical, language or accessibility barriers. Under Article 50 of the EU AI Act, people must be able to recognise that they are talking to an AI (transparency duty from 2 August 2026). The candidate portal shows how several paths merge into one consistent process.
Step 8: test the interview before launch
Test at least: an ideal answer, a weak but honest answer, an unexpected career path, a technical failure and an answer with sensitive information. Use invented test profiles, not real applicant data.
- Have several reviewers score the same answers independently before they compare notes.
- Compare score, rationale and missing information.
- Check transcription, accents and technical terms.
- Test drop-off, resumption and the alternative channel.
- Correct guide and scorecard before real invitations go out.
If reviewers repeatedly reach different results, the cause is usually vague criteria or a poor scale. More AI does not solve that design problem.
How do you run the AI interview once it is built?
A finished guide is only half of it. The run has a fixed order, and most acceptance problems appear in the first two steps.
- Invitation: say that the conversation is with an AI system, how long it takes, who reviews the answers and how to ask for a human alternative.
- Channel choice: let candidates pick chat or voice where the role allows it; chat or voice depends on the audience.
- Interview: the AI follows the approved questions and permitted follow-ups only.
- Review: a named recruiter reads the scorecard with the quotes behind it and decides the next step.
- Reply: every candidate gets an answer, also those who do not advance.
Not every vacancy should start with an AI interview. Check which roles fit an AI interview first, and compare tools in our overview of AI interview software. The product behind this workflow is AI interviews in Sprad screening.
Which mistakes happen most often?
- Too many criteria: a first interview with eight criteria becomes an interrogation, drop-off rises and the scorecard turns arbitrary.
- A fit score instead of evidence: a single number can be neither checked nor corrected.
- Different questions per person: if the AI probes freely, you end up comparing conversations, not candidates.
- No alternative channel: anyone who cannot use the microphone or the language silently drops out.
- Never testing weak answers: ideal answers pass almost any scorecard; weaknesses show only with honest, patchy answers.
- Automatic rejections: the biggest legal risk (see below).
What a human reviewer actually has to check
Human oversight only works if the reviewer sees more than a recommendation: the relevant answer, the criterion, the evidence anchor and the uncertainty of the evaluation. They must be able to correct the output and reject the proposed next step.
Separate three cases: the criterion was not evidenced, the question was not understood, or the information was never collected. They must not receive the same rating. Document the reviewer’s own reasoning, otherwise control turns into nodding through a summary. If relevant information is missing, it becomes an open question for the next personal step, not a new selection criterion added during review.
What does a complete interview blueprint contain?
| Phase | Content | Responsibility |
|---|---|---|
| Invitation | Purpose, duration, data notice, alternative | Recruiting |
| Opening | Process, notice that it is an AI, consent to the channel | System and recruiting |
| Core questions | The same job-related questions | Interview owner |
| Follow-ups | Only approved clarifications | Configuration |
| Scorecard | Criterion, evidence, rating, uncertainty | Human reviewer |
| Handover | Transcript, summary and open points into the ATS | Integration |
| Decision | Next step and documented reasoning | Accountable person |
How do you monitor quality after the pilot?
A passed test run is not a permanent release. Transcription, model behaviour, job requirements and candidate paths change. Assign fixed owners for approval, error analysis and shutdown. Every change to prompt, model, channel or scoring anchor creates a new version.
Do not only track the completion rate of interviews. Check transcript corrections, unsuitable follow-ups, loops, drop-offs, use of the alternative channel, contradictions between system and review, and candidate feedback. Look for patterns across languages, devices and audiences without drawing hasty conclusions from small numbers. Stop the affected version if an error can influence selection decisions, and reinstate it only after a documented fix and a new test.
Which legal questions need review?
EU AI Act (as of 31 October 2026): recruiting and selection AI is classified as high-risk in Annex III. The high-risk obligations (including human oversight, documentation, data quality) apply from 2 December 2027 under the Digital Omnibus, Regulation (EU) 2026/1744. Transparency duties under Article 50 apply from 2 August 2026; the Article 5 prohibitions (for example emotion recognition in the workplace) have applied since 2 February 2025. Build for the later duties now instead of retrofitting.
GDPR: the GDPR requires transparency, data minimisation and storage limitation. Article 22 protects against decisions based solely on automated processing. In the SCHUFA ruling (C-634/21, 7 December 2023) the Court of Justice held that even an automatically calculated score can fall under Article 22 if the recipient relies on it decisively. For AI interviews: a score recruiters adopt unchecked is risky; a documented human review with its own reasoning is the safe route.
Works council (Germany): Section 87(1) no. 6 BetrVG (technical devices for monitoring behaviour or performance) and Section 95 BetrVG (selection guidelines) can apply. The article on AI in recruiting and the works council covers the practical steps.
What does an AI interview cost? A worked example with Sprad
At Sprad you pay in credits only for the steps you switch on, per application: CV review 2, chat interview 10, voice 30, phone 40, video 80 credits. A free account includes 100 credits a month; Atlas Base is 5,000 credits for €450 a year (€0.09 per credit). Prices net, as of 29 October 2026, per the pricing page.
| Assumption: 100 applications a month | Credits per application | Credits total | Cost at €0.09 |
|---|---|---|---|
| Review + chat interview | 2 + 10 = 12 | 1,200 | about €108 |
| Review + voice interview | 2 + 30 = 32 | 3,200 | about €288 |
| Review + video interview | 2 + 80 = 82 | 8,200 | about €738 |
The example is an assumption: it presumes every application goes through the interview. In practice you often invite only a shortlist, and costs fall accordingly.
When is an AI voice interview useful, and what can Sprad do?
A voice interview pays off when spoken answers add relevant information and scheduling would otherwise slow you down, for example for customer contact or language level. Sprad runs such structured conversations by chat, voice, phone and video in many languages, and on request over WhatsApp. Every application is read and scored, with a rationale per criterion and the quote behind it. Nothing is rejected automatically: the tool collects and structures evidence, people own the next step. The screening and voice interview overview shows the flow.
What Sprad is not: a replacement for the conversation with the hiring manager, or a tool that takes the decision off your hands. If you also want to find candidates, the same platform searches your own talent network.
Frequently asked questions about creating AI interviews
Can ChatGPT create an interview guide?
Yes. It drafts questions, follow-ups and a scorecard. You must check the output for job relevance, fairness, data protection and accuracy, and must not enter real applicant data without a settled legal basis.
How many questions should an AI interview have?
For a short first conversation, three to five core questions are enough, each serving exactly one criterion. A long catalogue raises drop-off without adding predictive value.
How long should an AI interview take?
As a guideline, 10 to 15 minutes. State the duration in the invitation and cap it in the configuration.
Can the AI ask follow-up questions?
Yes, if permitted types and limits are defined: ask for an example, clarify the person’s role, ask for the result. Free-form probing is unsuitable.
Should the AI reject candidates automatically?
No. Automatic exclusions carry legal risk (GDPR Article 22, EU AI Act) and lower quality. Use human review, documented criteria and a way to correct errors. See also Can AI reject candidates automatically?
Is an AI interview legal in the EU?
Yes, with conditions: informing applicants, a GDPR legal basis, a human decision and, where applicable, works-council involvement. The AI Act’s high-risk obligations apply from 2 December 2027.
What makes a good result?
A good record shows the relevant answer, the rating and the remaining uncertainty per criterion. A blanket fit score without source reference is not enough.
What does an AI interview cost?
At Sprad from 2 credits for the review plus 10 for a chat interview or 30 for voice. In the example above that is roughly €1.08 to €2.88 per application (as of 29 October 2026).
Jürgen Ulbrich
Jürgen Ulbrich has more than a decade of experience in developing and leading high-performing teams and companies. As an expert in employee referral programs as well as feedback and performance processes, Jürgen has helped over 100 organizations optimize their talent acquisition and development strategies.
