An AI voice interview is a structured first conversation in which software asks role-relevant questions, captures spoken answers, and prepares them for a recruiter. It belongs early in a hiring process when a team needs consistent information at scale; it should not make the final hiring decision or replace a substantive conversation with a person.
The useful question is not whether an interview sounds human. It is whether the workflow gives candidates a clear, accessible way to provide relevant context and gives a qualified reviewer enough information and authority to decide what happens next. In EU hiring, that workflow also has to be designed around the GDPR and the EU AI Act; US teams need to add the rules that apply in their own jurisdictions.
What counts as an AI interview?
An AI interview is a software-mediated exchange used to collect or structure information for a hiring process. Voice interviews use spoken responses. Chat interviews use written answers. Both can cover availability, experience, qualifications, or a job-specific scenario, but they are not interchangeable for every candidate or every role.
Video is a separate choice, not a default upgrade. Once a process records or analyses a candidate's image, it can create more data, more accessibility concerns, and a harder case for necessity. A phone screen normally means a conversation with a recruiter; an AI voice interview delivered by phone is still an automated, scripted process step even though it uses the telephone network.
That distinction lets a team avoid collecting more than it needs. If the objective is to verify a license, shift availability, or an example of relevant experience, voice or chat may be sufficient. If the objective is to assess how a candidate handles a live stakeholder conversation, a trained human interviewer should take that part of the process.
Why hiring teams need a sharper standard in 2026
Application documents alone do not always create comparable evidence. A short, structured conversation can collect details that a resume cannot reliably show: what a person actually did, when they can start, what work pattern they can take on, or how they explain a relevant example. The value is the defined evidence, not an opaque claim that a system has found the best person.
EU law makes the intended use consequential. AI systems used for recruitment or selection are generally listed as high-risk under the EU AI Act. A narrow procedural or preparatory task may be outside that status when it does not materially influence the outcome, while systems that profile people require particular care. The official text therefore supports a use-case assessment, not a blanket conclusion based only on the word interview. The EU AI Act's high-risk classification and deployer obligations set out the relevant framework.
For an employer, the practical consequence is simple: document the purpose before launch. Specify the questions, the outputs, the people who review them, the cases in which the system is overridden, and the data that are retained. A transcript that helps a recruiter prepare is materially different from a ranking that silently determines who is rejected.
Should you use voice, chat, video, or a human phone screen?
Choose the format for the information you need. Voice is useful when a candidate can explain practical experience naturally and an asynchronous first conversation removes scheduling friction. Chat can be a better fit for candidates who need quiet, time to formulate answers, or a written channel. It can also be a necessary equivalent route for people who cannot or do not want to use voice.
Use video only where visual interaction has a specific, job-related purpose that can be explained. Do not make camera use a proxy for commitment or communication skill. A recruiter-led phone screen is the stronger option when a candidate has questions, needs accommodation, is discussing sensitive circumstances, or when the quality of a live human exchange is itself what the role requires.
Write down the equivalent path before candidates receive an invitation. A fair process can standardize the evidence sought without forcing every person through the same modality. That is better for accessibility and gives the team a defensible answer when it reviews candidate experience.
Which channel works for the role and the candidate?
Channel choice is part of process design. A portal can work well when an application already starts online and candidates need a single place for documents and follow-up steps. WhatsApp can reduce friction for mobile-first audiences. A phone call can be the practical route for frontline and non-desk hiring, where email and lengthy forms are a poor fit.
None of these channels removes the need for notice. Before the conversation starts, explain that AI is involved, what the session covers, how long it is expected to take, what data will be created, and whom the candidate can contact. A link sent without context does not create informed participation.
Test the invitation and channel with the actual target group rather than assuming one channel fits all jobs. The best route for a warehouse shift may not be the best route for a specialist role. Convenience should never mean that a candidate who needs an alternative receives less consideration.
What should the interview ask?
Every question should map to a documented job requirement. Good early-stage topics include relevant experience, work authorization where applicable, required qualifications, availability, start date, shift constraints, and a focused example of a task the role genuinely involves. Decide in advance why each answer is needed and what a reviewer may do with it.
Separate hard requirements from exploratory evidence. A mandatory certification can be handled as a clear check. An open answer about collaboration or motivation should not become an unexplained universal score. Structure means comparable, role-relevant prompts; it does not mean requiring candidates to have the same background, communication style, or life circumstances.
Do not add questions merely because an agent can ask them. Avoid collecting health, family, religion, origin, age, or other information that is not relevant to the job. Review the script question by question with recruiting, the hiring team, and privacy stakeholders before a pilot starts.
Where must human judgment remain?
Software can ask a consistent set of questions, create a transcript, organize answers against pre-set criteria, and flag missing information. It should not replace the person accountable for an invitation, a rejection, or a hire. A reviewer needs the competence, time, and authority to disagree with the system, not just a button that records approval.
Article 22 GDPR gives people the right not to be subject to a decision based solely on automated processing when it produces legal or similarly significant effects. In the limited cases where sole automation may be permitted, the Regulation still requires safeguards including human intervention, the ability to express a point of view, and a way to contest the decision. Read Article 22 GDPR in the official text.
Build a real review protocol: identify the evidence visible to recruiters, the evidence deliberately withheld, the conditions for overriding a recommendation, and the record kept when an override occurs. This is also how a team guards against automation bias—the tendency to treat a machine output as more certain than it is.
What does GDPR-ready use look like?
Audio, transcripts, and interview metadata are personal data. Before rollout, establish the controller and processors, the legal basis, recipients, access controls, retention period, and deletion path. Assess whether the intended processing requires a data protection impact assessment under Article 35 GDPR; high-volume or consequential hiring uses warrant careful assessment rather than a generic privacy notice.
Give candidates meaningful information before the session. They should know that they are interacting with AI, whether the conversation is recorded and transcribed, whether an assessment is generated, how the information affects the next step, and how to reach a person. Consent is not a substitute for purpose limitation, data minimization, or a defensible legal basis.
Keep the data needed for the next hiring decision, not every artifact that the system can produce. If a reviewer can work from a short, controlled summary, retaining the full audio indefinitely may be difficult to justify. Privacy practice has to match the actual workflow, including exports into an ATS and access by hiring managers.
How do you handle languages, dialects, and candidate acceptance?
Support for a language does not prove equal performance across accents, regional dialects, technical vocabulary, noisy environments, or speech disabilities. Test the exact role, language, and channel with realistic sample answers. Measure practical failures such as misunderstandings, unusable transcripts, repeated prompts, and drop-offs; do not publish broad accuracy claims you have not measured.
Sprad states that its voice interviews support more than 30 languages, based on product documentation dated 20 August 2026. For EU and US audiences, its stated privacy position is compliance with relevant requirements and EU hosting as an available option. Those are useful procurement inputs, but they do not remove the employer's responsibility to validate the language experience for its own candidate population.
Candidate acceptance is earned in the invitation and the process. State the reason for the step, the expected time, the role of a human reviewer, and the available alternative. Ask a short post-interview question about clarity and effort. A completion rate alone cannot tell you whether the issue was trust, language, accessibility, the channel, or an unnecessarily long script.
How should you assess an AI interview tool?
Start with evidence and governance rather than a polished demo. Ask a provider to describe the intended use, the data generated, the role-specific criteria, the explanation available to reviewers, the human override mechanism, data deletion, and ATS export. For EU use, also ask for the information needed to assess high-risk classification, deployment responsibilities, and the relationship between the tool's records and your privacy documentation.
Cost should be understandable in the same terms. Under Sprad's pricing as of 19 August 2026, a five-minute voice interview uses 28 credits, or about €1.96 at the stated credit rate. That pays for a structured first-pass task; it does not buy an employment judgment or eliminate the need for a human conversation. The voice interview use case explains the available portal, WhatsApp, and call-based paths.
- Evidence: Can reviewers see what supports an output rather than only a ranking?
- Control: Can you configure questions, criteria, permissions, retention, and escalation by workflow?
- Accessibility: Is there an equivalent alternative for language, disability, or channel barriers?
- Data flow: Can only the necessary information return to the ATS?
- Operational accountability: Does the provider supply the documentation your privacy, HR, and legal teams need?
Frequently asked questions
Are AI voice interviews legal?
They are not categorically prohibited. Lawfulness depends on the jurisdiction, the role of the system in a decision, the data processed, transparency, human oversight, and the employer's wider hiring process. This page is a process guide, not legal advice for a particular deployment.
Does an AI interviewer have to disclose itself?
For EU high-risk AI systems that make or assist decisions about natural persons, deployers must inform the affected people that they are subject to the system. Even outside that specific obligation, clear notice is the sound operational choice. A candidate should not discover midway through a call that the interviewer is automated.
Can a voice agent score confidence, personality, or cultural fit?
Do not treat inferences from voice, accent, or manner as a standalone hiring basis. Use concrete, job-related answers and a human review instead. The more an output ranks or excludes people, the more important it is to assess its legal status, validity, and risk of discrimination.
How long should an AI first interview take?
Only as long as the defined first-stage questions require. Give the expected duration before the candidate starts and test whether it is realistic. If a pilot produces thin answers, improve the questions before extending the session.
What should be stored in the ATS?
Store the minimum evidence needed for the next step, such as a verified requirement, availability, or a human-reviewed summary. Do not automatically make raw recordings, incidental comments, and full transcripts part of a permanent candidate record. A structured CV-screening workflow can collect further context without treating the resume or the interview transcript as the whole person.
Continue with the decision that matters next
A sound pilot begins with one role, a short script, a tested alternative channel, and a named human reviewer. From there, the deeper questions are specific: video versus voice, non-English interviewing, candidate access, works council involvement in Germany, bias checks, and ATS handoff. The useful boundary is clear: AI can make the first conversation more structured; people remain responsible for what the conversation means.
