AI Voice Interview Platform for Employee Insights
An AI voice platform that interviews everyone in an organisation by phone or browser, in their own language, and turns the conversations into counted, quote-backed themes and next steps without rating any individual.
Key Details
| Challenge | Before a reorganisation, a transformation or a post-deal plan, leaders need to know what is really happening across the workforce. Surveys return scores without reasons. Workshops and consultant interviews reach a handful of people and take weeks. Frontline staff without a desk are rarely asked at all. Any listening tool also has to protect anonymity and satisfy employee representatives and data-protection rules. |
|---|---|
| Solution | A voice-first insights platform. The team sets one main question and shares background documents. An AI interviewer then holds short, natural conversations with every invited person by phone or browser, in their own language, and asks follow-ups based on what they say. A group of AI coding agents reads every transcript against a fixed codebook, and a theme only counts when the agents agree. Results show how many people said what, with the quotes behind it, split by group and turned into suggested next steps. |
| Technologies | Real-time voice AI (speech recognition and speech synthesis), phone and browser calling, LLM interview agents, multi-agent coding ensemble with agreement checks, PII redaction, Python and TypeScript services, Next.js and React, PostgreSQL, MCP server, HRIS, document and chat connectors, cloud hosting with data-residency controls |
Technologies used
Voice AISpeech-to-textText-to-speechTelephonyLLM agentsMulti-agent ensemblePII redactionPythonTypeScriptNext.jsReactPostgreSQLMCP serverREST API
Overview
Axtra Labs engineered an AI-native workforce insights platform for an HR tech startup. It lets an organisation talk to everyone it wants to hear from at once, instead of a few people in a workshop. An AI interviewer holds short voice conversations by phone or in the browser, in each person's own language. AI agents then read every conversation and report what people said, how many said it and in their own words. HR, operations and transformation teams, investors' operating teams, consultants and employee representatives use the results to decide what to change. The client is not named.
The challenge
Big decisions about how people work are often made on thin evidence. The client wanted to replace a few weeks of interviews with a broad, fast and fair picture of the whole organisation.
- Engagement surveys give a score but not the reason behind it, and long free-text answers are rarely read.
- Interviews and workshops are slow and expensive, so they reach a small sample, usually managers and office staff.
- Frontline, shift and multilingual teams are the hardest to reach and the most often missed.
- AI analysis has to be trustworthy: themes must be counted consistently and backed by real quotes, not made up by a language model.
- People only speak openly if they trust the process. Small groups must never be identifiable, personal data must be protected, and nobody can be scored.
- Many organisations need employee representatives to agree the purpose and the groups before a project starts.
The goal was a platform that goes from one question to organisation-wide results in days, with privacy and fairness built into the product rather than added as a policy.
Product screenshots
The solution
Axtra Labs designed and built the platform as a pipeline of AI agents with privacy controls at every step, from setting up a project to exporting results.
- Project setup. The team writes one main question, uploads background documents and imports the org chart from the HR system so people can be invited and grouped. The interviewer reads the documents before the first call.
- AI voice interviewer. A real-time voice agent holds short conversations by phone or in the browser, in the person's own language. Its first sentence says it is an AI. It follows a topic guide, chooses follow-up questions from what it hears and can look at a shared screen or camera when the person wants to show something.
- Privacy pipeline. Names and identifying details are removed from transcripts before any language model reads them. Audio is deleted soon after transcription, and retention periods are configurable. The system does not analyse emotion or keep voiceprints.
- Multi-agent coding. Several AI agents code each transcript independently against a fixed codebook. A theme is counted only when a majority of them agree, so results are consistent and repeatable rather than a single model's opinion.
- Counted, quote-backed results. Each theme is reported as how many people said it, with the quotes behind it, split by team, site, role or other group. A minimum group size is enforced in the dashboard, exports and API, so small groups stay hidden.
- From findings to action. The platform suggests next steps and owners for each theme, and repeat waves with the same question show what has changed since last time.
- Employee representative mode. The purpose and the groups are agreed before launch, and only aggregated results are released.
- Integrations. Connectors for document stores, chat tools and HR/ERP systems bring in context and org structure. An MCP server and API let teams query the results from the AI assistants they already use.
The platform describes what groups of people said. It does not rate, rank or profile any individual.
Key features
- AI voice interviews by phone or browser, in each person's own language, with adaptive follow-ups.
- Context-aware interviewer that reads the documents shared before the project starts.
- Optional screen and camera sharing during the interview.
- Multi-agent theme coding with a majority-agreement rule and a fixed codebook.
- "N of M said" results with supporting quotes, group splits and an ask-the-results chat.
- Suggested next steps with owners, and repeat waves to track change.
- Privacy by design: PII removed before AI analysis, a minimum group size, short audio retention, no emotion reading and no individual scores.
- Excel, report and API exports, plus an MCP server for AI assistants.
What was delivered
- A real-time AI voice interviewer for phone and browser, with multilingual conversations and AI disclosure.
- A transcript privacy pipeline and configurable data retention.
- A multi-agent analysis engine that codes themes with agreement checks and links every theme to its quotes.
- A results dashboard with group splits, minimum-group-size protection, next steps and wave comparison.
- Connectors for HR systems, document stores and chat tools, plus a public API and MCP server.
- Project setup, invitation and scheduling flows for HR and transformation teams.