Enterprise Voice Agents by Concretio
We are building voice agents that work with enterprise data from established SaaS providers. Our agents are not simple file-powered agents. They talk to CRM and HRMS systems to bring real intel into the conversation. We first did this in April 2025, in the Speech-to-Text and Text-to-Speech era of voice agents. Today we are ready with the latest voice-to-voice agents, with even lower latency than the previous era.
We strongly believe voice agents can help across business functions, with the most obvious ones being Sales and Support, leading on to Hiring (interviews), driving Standup Calls, and more.
Our first goal is hire.concret.io: a simple, easy-to-plug platform that lets companies bring their HRMS, plug it in, and start screening candidates at scale through automated, AI-based voice screenings.
Healthcare credentialing traditionally means slow, manual primary source verification across multiple registries and licensing boards. We work on Verifiable, a no-code Salesforce app that runs real-time Primary Source Verifications (PSV) in seconds, so credentialing teams can manage their entire provider network from inside Salesforce.
The Problem We Set Out to Solve in Hiring
- First-round screening eats recruiter hours, follows the same script every time, and is hard to schedule across time zones.
- Our agent runs that first round on its own, around the clock, with no recruiter on the call. The recruiter only sees the outcome: a score, a recommendation, and the evidence behind it.
Type of AI Voice Agents Implemented
We are currently optimising for CRM and HRMS consumption and will later go deep into project management. We think the key part of any voice agent is tuning and guardrails. That takes time to master, and it is common across industries.
Agents Built So Far
Autonomous Screening Interviewer
Candidate-facing, and it runs the whole call by itself.
In-meeting Assistant (Aria)
Joins Google Meet or Zoom, responds when called by name, and pulls live data from connected systems during the meeting.
Multiple Personas Per Agent
Each with its own voice, tone, and pacing, tested to behave and score the same way whichever voice stack is behind them.
Key Functionality & Workflow
A typical hiring workflow with our agent.
Connect
Open positions are pulled straight from the company's HRMS (Zoho People today).
Shortlist
Resumes are bulk-uploaded, parsed, and ranked against the job description by AI before anyone is invited.
Invite
The recruiter sends screening invites in one click seamlessly.
Verify
The candidate joins from a secure link with email verification and a mic and camera check. No account is needed.
Interview
The voice agent runs a personalised interview built from that candidate's resume and the job description.
Report
The recruiter gets a scored report, the full transcript, and a recommendation and makes the call.
Technology & Architecture
We work with voice-to-voice models from OpenAI, Grok, Gemini, and Deepgram. We are most invested in OpenAI Realtime and are going live with it.
On architecture, as with other cloud solutions, we have mastered scaling voice agents and making sure they survive noise, interruptions, and network drops, and scaling on the server for voice-heavy payloads. Doing a POC is one thing. Making sure the system runs smoothly in production with voice agents is another.
Provider-Agnostic Design
The voice model sits behind an abstraction layer, so we can pick the best model per use case on latency, cost, or voice quality. We run both a native speech-to-speech stack and a modular speech-to-text, LLM, text-to-speech pipeline in production.
Real-Time Media
Hosted on LiveKit and deployed on Fly.io.
Noise Rejection
Ensures the agent doesn't respond to background sound or silence.
Natural Barge-In
The agent stops talking the moment the candidate interrupts.
Silence & Drop-Off Handling
The agent checks in if the candidate goes quiet and ends gracefully if they've left.
Durable Workflows (Inngest)
Scoring, notifications, and deliveries survive restarts and retry on failure.
Full Observability
OpenTelemetry and Grafana provide tracing, alerting, and per-call AI cost tracking.
Reasoning, Decision-Making & Actions
During the Conversation
Context-aware conversational intelligence with safety guardrails
Live Enterprise Grounding
Briefed with real-time JD and resume data to generate tailored inquiries and follow-ups instead of static scripts.
Checklist Guardrails
Mandatory coverage of progression, gaps, notice periods, and technical competency before wrapping up.
Autonomous Tool Calling
Invokes tools to flag misconduct, manage off-script incidents, or gracefully terminate the session.
After the Conversation
Evidence-backed evaluation with human-in-the-loop oversight
Decoupled AI Judge
A separate impartial judge scores weighted dimensions (technical, communication, behavioural) without self-grading bias.
Acoustic & Accent Fairness
Calibrated to account for regional accents, latency, and transcription artifacts so candidates are never penalised.
Evidence Over Assertion
Zero unsupported claims. Every evaluation criteria and flag must link directly to a verbatim transcript quotation.
Human in the Loop
Uncertain ratings are held for recruiter review rather than auto-rejected. The AI recommends; people decide.
Ready to Automate Your Screening with AI Voice Agents?
Connect your HRMS and start conducting automated, high-quality candidate voice screenings at scale. Schedule a call with our team to discuss your project goals and see a live demo.