If you’re adding AI interviews to an ATS, HR platform or hiring product, connecting a voice API to an LLM can look like the quickest route. It gives you a voice, a model and a way to exchange messages. It does not give you a dependable interview product.
You still need to make the conversation work like an interview, assess candidates consistently, handle unexpected answers, detect possible assistance and test the system as models and candidate behaviour change. If you get these layers wrong, you risk shipping a slow or unreliable experience, making poor screening decisions, losing candidate trust and creating legal or data-protection exposure. You also need to understand how candidate data is collected, used, stored and protected, and which hiring and communications rules apply to your deployment.
That is the build-versus-integrate decision: do you want your team to build and maintain an interview engine, or keep building your hiring product and integrate an interview layer that already exists?
A voice API is not an interview engine
A basic voice-and-LLM connection can carry a conversation. A hiring interview has to do much more. It needs to understand the role, ask relevant questions, follow up when an answer is vague, keep the candidate on track and produce an assessment a recruiter or hiring manager can use.
The work between a working voice demo and a dependable hiring workflow includes:
- Voice quality and turn-taking: build a low-latency pipeline, handle interruptions and avoid awkward pauses or candidates talking over the system.
- Interview logic: ask role-specific questions, adapt follow-ups and support practical tasks such as coding exercises and role-plays.
- Guardrails and context: keep the interview focused, handle unexpected answers and reduce hallucinations or loss of context.
- Scoring and evidence: create role-based rubrics, apply them consistently and show the evidence behind an assessment.
- Integrity checks: detect signals of AI assistance or outside help and give reviewers evidence to examine, rather than treating a signal as proof of cheating.
- Evaluation and maintenance: build labelled test data, run regression checks, monitor model changes and keep improving interview quality.
These are not one-time tasks. They become part of the product your team has to test, support and maintain.
The compliance work depends on how you use the interview
There is no single checklist for every AI interview product. The questions change depending on where candidates are, whether you record or transcribe interviews, what the system infers from answers, whether it scores or ranks applicants, and how much its output influences a hiring decision. Map the candidate journey and review these areas with qualified counsel in every market where your customers recruit.
Candidate notice and transparency. Decide what candidates are told before an interview begins: that they are interacting with AI, what the system does, whether a human reviews the result and how they can ask questions or request an alternative. The notice and consent requirements depend on the jurisdiction and use.
Recording, transcription and candidate data. Identify the audio, video, transcripts, task submissions, scores and reviewer notes you collect. Define why each is needed, who can access it, how long it is retained and how candidates can exercise applicable rights. Permission to contact someone is not automatically permission to record them.
Automated assessment and human oversight. Be precise about whether the AI summarises an answer, suggests a score, ranks applicants or decides who advances. Assess applicable rules on automated employment decisions, explainability, human review and candidate rights. A human clicking approve does not necessarily settle the question if the system effectively determines the outcome.
Bias, validity and accessibility. Test whether the interview measures skills relevant to the role and whether speech handling and scoring work fairly across the candidate groups and languages you serve. Keep evidence of testing and provide an appropriate route for candidates who need an accommodation.
Security, vendors and data flows. Map where audio, transcripts and scores travel, which providers process them, where data is stored and who can access it. Review contracts, subprocessors, security controls, incident handling and cross-border transfers.
The details vary by market. For example, the EU AI Act lists AI systems intended to recruit or select people, including systems that evaluate candidates, as high-risk. NYC Local Law 144 can require an independent bias audit, public audit information and candidate notice when a covered automated employment decision tool is used. In the US, the FCC has said AI-generated voices count as artificial voices under the TCPA. In India, the DPDP Act’s main processing provisions are scheduled to commence on 13 May 2027 under the 13 November 2025 Gazette notification. These examples are not an exhaustive jurisdictional checklist; check the current rules for each market and workflow.
Using an API does not make these responsibilities disappear. Your organisation still needs to understand its role, configure the workflow appropriately and meet the obligations that apply to its use. A provider’s certifications and security controls are useful inputs to that review, not a substitute for it.
Build the hiring product. Integrate the interview layer.
If your core product is an ATS, HR platform or recruitment workflow, the interview engine may not be where you want to spend months of engineering time. You may prefer to focus on customer workflows, integrations and the parts of your product that make your platform different.
Fabric’s AI Interviewer API is designed for teams building AI interviewing into systems they already own. Fabric says it has conducted 150,000 AI interviews for customers including Meesho, CRED, Quantiphi and Sigmoid. You keep your product, brand and workflow. Fabric provides the interview layer through APIs.
The API brief describes humanlike voice conversations with less than 1.5-second latency; adaptive, role-specific interviews with follow-ups, live coding, role-plays and practical tasks; interview-specific guardrails; consistent role-based scoring with evidence; and signals of AI-assisted answers flagged for review.
Fabric’s brief also describes a curated evaluation set of more than 100,000 records and continuous evaluation. Its stated security posture includes SOC 2 and ISO 27001, AES-256 encryption at rest, TLS 1.3 in transit and role-based access controls. Review the relevant reports, technical details and contractual terms as part of your own vendor assessment.
Fabric says a scoped first integration can take around two weeks once requirements and dependencies are agreed. Treat this as an indicative starting point, not a guaranteed delivery timeline. Actual effort depends on your architecture, workflow and integration requirements.
Compare it with the interviewer you have built
If you already have an in-house interviewer, compare both systems using the same roles, candidate groups and hiring bar. Measure whether candidates participate, whether the next human round agrees with the assessment, how candidates rate the experience and how reliably the system flags suspicious assistance.
Fabric’s API brief reports reference comparisons from systems it measured: pass-through to the next human round of 20% for in-house systems versus 50% for Fabric; participation of 30–40% versus 65%; and candidate experience ratings of 6.5–7/10 versus 8.6/10. It also reports 91% cheating-detection accuracy from internal testing. These are Fabric-reported figures, not independent benchmarks or guaranteed outcomes. Results vary by role, candidate mix and implementation.
Use a pilot to test the difference in your own workflow. Agree the roles, evaluation criteria, candidate-experience measures and review process before comparing results.
What to bring to an API integration discussion
Start with one role and the hiring workflow you want to support. Be ready to discuss where the interview starts, how candidates are invited, the interview format and evaluation criteria, what information your platform needs back, who reviews it, and your technical and security requirements.
The goal is not to hand over your hiring product. It is to avoid rebuilding every part of the interview engine before you can offer AI interviews to your customers. Bring a role and your workflow to Fabric to discuss an API integration and agree a focused pilot.
This article is general product and compliance information, not legal advice. Requirements vary by jurisdiction and by how an AI interview system is designed and used. Get qualified legal and security advice before deploying it.