AI screening interview is an initial interview conducted by an AI agent in chat, a messenger app, or by phone: it responds to the application, explains the job opening, asks scripted questions, checks basic requirements, transcribes the conversation, and sends the recruiter an assessment. In high-volume hiring, this takes routine top-of-funnel work off the team. The decision to reject or hire, however, must be made by a human: Russia’s personal data law prohibits decisions that create legal consequences for a person based solely on automated processing, except in cases with the person’s written consent or when expressly provided by law.
Prepared by AIrassvet using publicly available sources, verified on October 8, 2026. It is intended for business owners, HR leaders, and heads of high-volume recruiting in retail, call centers, logistics, manufacturing, and services. This article is not legal advice: the handling of candidate data and hiring decisions should be reviewed with legal counsel and aligned with company policy. AIrassvet is not affiliated with the companies mentioned, and the section at the end describes the publisher’s services.
Contents
- Why this topic became a hot issue in October 2026
- What an AI agent does in initial screening
- In brief: funnel stages and AI’s role
- What metrics to define: the Beeline tender example
- Where the legal line is
- Key risks and how to reduce them
- How to set up the system technically
- Six-week pilot plan
- When AI screening is not needed
- Frequently asked questions
- Primary sources and verification method
Why this topic became a hot issue in October 2026
October 5, 2026 CNews reported, that Beeline collected commercial proposals from vendors of an AI agent platform for high-volume recruiting in its call centers. According to the procurement documents cited by the publication, the platform must:
- handle more than 130,000 applications, conduct at least 10,000 interviews, and initiate no fewer than 40,000 conversations per year, with support for 100 concurrent sessions;
- communicate with candidates wherever they submitted their application: on job sites, in messaging apps, on the careers site, and by phone;
- analyze resumes, test candidates, and conduct voice interviews;
- integrate with the company’s ATS: move candidates through the funnel, apply tags, and export transcripts and audio recordings to the candidate profile;
- host all servers in Russia in accordance with Federal Law No. 152-FZ.
The procurement is a request for proposals and does not obligate the company to sign a contract. Beeline emphasized that this is not about handing hiring decisions over to AI. According to the same report, T2 also says the final screening decision is made by a human, while MegaFon uses scripted bots and AI to process applications in high-volume hiring.
Experts quoted in the piece disagree on how widespread the technology is: according to Skolkovo, only 5% of companies use AI in hiring, while a Get Experts survey puts the figure at 49%. The difference, says Igor Terekhin (Dar, Korus Consulting Group), is in what counts as AI: a recruiter’s tool or an agent that manages the funnel on its own. The main shift in 2026, in his view, is the move from experiments to tenders with SLAs.
What an AI agent does in initial screening
An AI recruiting agent is a software program based on a language model that holds a conversation with a candidate and performs actions in the recruiting system according to predefined rules. Unlike a scripted bot, it understands free-form answers, responds to candidate questions, and can steer the conversation back to the script.
Typical tasks:
- responding to an application within minutesrather than the next day, in the same channel where the candidate applied;
- explaining the job opening— schedule, conditions, location, and screening stages — using an approved knowledge base;
- checking basic requirements— availability for the schedule and work format, required documents and clearances — only against criteria tied to business qualifications;
- competency-based interviewing— standard questions, situational tasks, and basic skills tests;
- transcription and summary for the recruiter: answers, quotes, and a score for each criterion with a reference to the relevant part of the conversation;
- scheduling the next step and sending reminders about the meeting;
- re-engaging “lost” candidates— following up again with people who did not respond or did not show up.
A CNews expert notes that, based on the tender metrics, the main funnel loss is not rejections, but candidates who stopped replying or never showed up. A fast response and confirmed attendance deliver more value than perfect questions.
In brief: funnel stages and AI’s role
| Stage | What AI does | What remains with a human | Risk of error |
|---|---|---|---|
| Application | Responds in the candidate’s channel, explains the job opening | Configuring the response base | Low |
| Resume screening | Matches the resume to the requirements, flags mismatches | Decision on edge cases | Medium |
| Initial interview | Asks questions, records and transcribes answers | Reviewing scores, borderline cases | Medium |
| Assessment | Assigns scores by criteria with quotes from the conversation | Decision to advance or reject | High |
| Scheduling and attendance | Schedules the meeting, sends reminders, re-engages lost candidates | Meeting with the candidate | Low |
| Analytics | Shows where candidates are dropping off, by job opening and channel | Management insights | Low |
Table — AIrassvet editorial classification; risk is assessed based on the consequences of an error for the candidate.
What metrics to include: Beeline tender example
The value of Beeline’s tender is that it defines measurable requirements for the AI agent. According to CNews:
| Requirement | Value in the tender |
|---|---|
| Speech and response recognition accuracy (ASR/NLU) | At least 99% |
| Agreement between the agent’s competency assessment and the HR expert’s assessment | At least 85% of cases |
| Initial resume screening accuracy | Above 90% |
| Response time to an applicant’s message | No more than 10 seconds |
| Upload of the transcript and audio to the candidate profile | Within 5 minutes after the conversation |
In the same piece, Igor Terekhin describes funnel business metrics that the client sets for the agent as targets: at least 30% of candidates are booked for a job presentation, 25% of those who “dropped out” are brought back, and no more than 20% abandon the dialogue. Verification is a blind comparison of the agent’s assessments with those of HR experts.
The experts in the article consider some of the requirements ambitious: recruiters often disagree with each other in their assessments, and a 10-second pause in a live voice conversation, conversely, is too long. For your own project, these figures are useful as a metric framework, not as ready-made thresholds: the thresholds need to be set based on your baseline.
Metrics worth tracking in your pilot:
- time to first response to the candidate;
- share of candidates who move to the next stage, and show-up rate;
- share of abandoned conversations;
- agreement with HR on a blind sample;
- share of false rejections: candidates whom the agent recommended screening out, but whom the recruiter would have moved forward after review;
- recruiter time per hire.
Where the legal boundary is
Three rules determine where a human must be involved.
1. Decisions based solely on automated processing. Article 16 of Federal Law 152-FZ prohibits making decisions that produce legal consequences for a person or otherwise affect their rights and legitimate interests solely on the basis of automated processing of personal data. Exceptions include written consent or cases provided for by federal laws. The controller must explain the procedure for such a decision, give the person an opportunity to object, and review the objection within 30 days.
Practical takeaway: the agent can recommend, but a human must confirm the rejection, with access to the transcript and the ability to disagree with the assessment. Fully automated rejection without human involvement is an area that requires separate legal review.
2. Ban on unjustified rejection and discrimination. Article 64 of the Labor Code of the Russian Federation prohibits unjustified refusal to enter into an employment contract and any restriction of rights based on sex, age, nationality, place of residence, or other circumstances unrelated to job qualifications, except in cases provided for by federal laws. Upon the candidate’s written request, the employer must provide the reason for the refusal in writing no later than seven business days. Therefore, the agent’s criteria should concern only job-related qualifications, and the assessment should be explainable through the candidate’s specific answers.
3. Candidate personal data. Resumes, voice, and transcripts are personal data. You need a legal basis and consent for processing, clear retention periods, and a processing location. In its tender, Beeline requires servers to be located in Russia. For more on sending data to external models — see the article on rules for sending customer data to AI.
Abroad, regulation is moving in the same direction. According to a Habr report from October 2, 2026, the governor of California signed laws that prohibit employers from making termination decisions based on AI recommendations and from using biometrics to assess employees’ emotional state.
That adds another rule: do not assess a candidate’s emotions based on voice or video. This is only weakly related to job qualifications, is hard to explain, and creates a discrimination risk.
Main risks and how to reduce them
| Risk | How it shows up | How to reduce it |
|---|---|---|
| Scaling of criteria errors | A good but atypical candidate is automatically screened out | A human reviews all rejections or a sample of them; borderline cases go only to a human |
| Bias | The assessment depends on accent, diction, or speaking style | Criteria based only on the content of answers; check alignment with HR by candidate groups |
| Speech recognition errors | An answer about schedule or experience is misunderstood | Confirm key answers, ask follow-up questions, offer text as an alternative channel |
| Poor candidate experience | The candidate abandons the conversation or declines the job | Clearly state that they are speaking with AI, and give them a way to reach a person |
| Made-up answers about the job | The agent promises a salary or conditions that do not exist | Answers only from an approved knowledge base; no improvising on terms and conditions |
| Data leakage | Resumes and recordings go to an uncontrolled service | Processing in Russia or in your own environment; access logs |
| Channel abuse | Messenger account blocks, session failures | Channel monitoring, a backup channel, message rate limits |
In the CNews article, Elena Agupova (Softline Solutions) named scaling of criteria errors as the main downside of AI agents. She also noted that with tens of thousands of outbound contacts, channel operations can turn out to be more complicated than the AI component itself. The other risks and measures are an editorial analysis by AIrassvet.
How to Set Up the System Technically
Minimum AI Interview Architecture:
- Channels: chat on job boards and in messaging apps, and a voice line for calls.
- Speech recognition and synthesis for the voice channel; recognition quality is checked on recordings of your candidates, with their accents and background noise.
- Language model with instructions, the interview script, and a database of answers about open roles. For candidate data, use a model deployed in Russia or within the company’s own infrastructure.
- Scoring module: scores against job criteria with a required quote from the candidate’s answer for each score.
- ATS integration: create and update the candidate card, stages, tags, transcript, and meeting booking.
- Recruiter dashboard: recommendation queue, transcripts, and “agree” and “disagree” buttons with a reason.
- Logs and analytics: every agent action, script version, and funnel metrics.
More on voice agents — in the article on voice AI for call centers, and on HR use cases overall — in the overview of AI in HR and recruiting. For requirements for an agent before development — in the piece on the technical specification for an AI agent.
Six-Week Pilot Plan
| Week | What We Do | Result |
|---|---|---|
| 1 | Choose one high-volume role and establish the baseline: response time, share who make it through, attendance, recruiter time | Baseline metrics |
| 2 | Formalize the job criteria and interview script together with top recruiters; legal reviews consent language and decision flow | Approved script |
| 3 | Build the agent and ATS integration; test recognition on real recordings | Working prototype |
| 4 | Suggestion mode: the agent conducts interviews, and the recruiter reviews every score and every decision | Alignment with HR on live volume |
| 5 | Review discrepancies and false rejections, refine the criteria and script | Improved version |
| 6 | Compare against the baseline and decide whether to expand to other roles | Data-driven decision |
This plan is an editorial recommendation from AIrassvet; the timeline depends on application volume and integration readiness.
Do not start with automatic rejections. In the first few weeks, the agent should only recommend, while a person confirms. You can automate tasks where a mistake costs the candidate nothing: role-related answers, interview booking, and reminders.
When AI Interviews Are Not Needed
- Low volume: if only dozens of applications come in per month, it is easier to speed up the recruiter’s work than to build an agent.
- Unique positions: for executives and rare specialists, the first interview is already a chance to sell the company to the candidate, and a person does that better.
- No formalized criteria: if recruiters themselves cannot describe whom they hire and why, the agent will lock in the chaos instead of removing it.
- No ATS or funnel tracking: without baseline metrics, you will not be able to tell whether things got better or worse.
Frequently Asked Questions
What is an AI interview?
It is an initial interview conducted with a candidate by an AI agent in chat, messenger apps, or by phone. The agent responds to the application, explains the role, asks questions according to the script, transcribes the answers, and sends the recruiter a score based on the job criteria.
Can AI reject a candidate on its own?
Article 16 of Federal Law 152-FZ prohibits decisions that produce legal consequences for a person solely on the basis of automated processing of their personal data, except in cases of written consent or where expressly provided by law. It is safer for the agent to recommend, with a person confirming the rejection decision.
What quality metrics should be set for an AI hiring agent?
Time to first response, share of candidates who move to the next stage, attendance, share of abandoned conversations, agreement between the agent’s ratings and HR ratings on a blind sample, and the share of false rejections. According to CNews, Beeline required 99% speech recognition accuracy and 85% alignment with the HR expert’s assessment in its tender.
Do candidates need to be told they are speaking with AI?
Yes, that should be done. Honest notice and the ability to reach a human reduce candidate frustration and help meet the obligation to explain the decision-making process if the decision is made solely on the basis of automated data processing.
Which roles are suitable for AI interviews?
For high-volume roles with a large flow of applications and clear criteria: call center operators, sales associates, cashiers, couriers, warehouse workers. For executives and rare specialists, a person should conduct the initial interview.
Where should you start with AI interviews?
Start with one high-volume role: record the current funnel metrics, formalize the criteria with top recruiters, review consent language and decision flow with legal, then launch the agent in suggestion mode, where a person confirms every decision.
Primary Sources and Verification Method
Checked on October 8, 2026:
- CNews, October 5, 2026 — Beeline’s request for proposals for an AI agent platform for high-volume recruiting, requirements and SLA, comments from Beeline, T2, MegaFon, and experts, data from Skolkovo and Get Experts.
- ConsultantPlus, Federal Law 152-FZ, Article 16 — rights of the data subject in decisions based solely on automated processing.
- ConsultantPlus, the Labor Code of the Russian Federation, Article 64 — guarantees when concluding an employment contract.
- Habr, October 2, 2026 — California laws on AI in employment decisions.
The tender requirements are presented as a paraphrase of CNews, not directly from the procurement documentation. The stage and risk tables, architecture, and pilot plan are AIrassvet’s editorial analysis; the article does not include its own testing or project results.
How AIrassvet helps automate the initial screening process
If recruiters cannot keep up with responses and candidates get lost between stages, the task can be broken down step by step. On the AIrassvet website, which also uses the name “AI Studio of Vladimir Lomtev,” the following relevant directions are described:
- An AI agent that communicates with candidates in messaging apps and on the website and works with your recruiting system.
- Integration of voice and text models for phone interviews with transcription and evaluation against job requirements.
- Deployment of a language model on the company’s servers or in a secure cloud so candidate data does not go to uncontrolled services.
- Audit of the hiring funnel and data preparation: job criteria, response database, and baseline metrics for comparison.
The first step is to choose one high-volume role, describe the current funnel metrics, data sources, constraints, and the criterion by which you will consider the pilot successful.
Bottom line: the agent runs the funnel, a person makes the decision
AI interviewing has already moved from experiments to tenders with measurable requirements. Its strength lies in response speed, consistent questions, and bringing back candidates who dropped out; its weakness is the risk of scaling flawed criteria too quickly.
A safe setup for a Russian company: the agent responds, asks follow-up questions, transcribes, and recommends, while a person confirms decisions and reviews discrepancies. Start with one high-volume role in copilot mode — this way, you get the data needed to make a decision without risking candidates or legal compliance.