Build a Custom Voice AI for Automated Reference Checks
The best voice AI reference checking solution is a custom system using a high-quality transcription model and a large language model like Claude. It calls references, asks structured questions, analyzes responses for key traits, and generates a summary report.
The system's complexity depends on the integration with your Applicant Tracking System (ATS) and the depth of the required analysis. A standard build connects to a modern ATS like Lever or Greenhouse, asks a set of 8-10 questions, and scores candidates on 5 core competencies. More complex builds might involve multiple languages or custom sentiment analysis models.
We built a voice AI reference checker for a 12-person recruiting firm processing 400 applicants a month. Their team spent over 25 hours monthly on manual calls. Our system went live in 3 weeks, completely automated the process, and cut the time spent on reference checks by 95%.
What Problem Does This Solve?
Recruiting teams often start with email-based reference tools like Checkster. These are essentially automated surveys. The problem is that references provide brief, low-effort written answers, and response rates are often below 50%. There is no way to ask a follow-up question or gauge the tone and conviction behind an answer, which is where the real signal is.
A more significant issue is the manual process itself. A recruiter at a small firm trying to fill 15 roles a quarter must conduct over 100 individual reference calls. Each 20-minute call is preceded by 10 minutes of email tag for scheduling. This consumes more than 50 hours per quarter, time that could be spent sourcing new candidates. The process is impossible to scale without hiring more recruiters.
Some firms try to solve this with virtual assistants, but that introduces new problems of consistency and quality control. One VA may be a great interviewer while another just reads the script. This introduces significant bias and noise into the hiring process. The data collected is unstructured and difficult to compare across candidates, making the entire effort subjective.
How Does It Work?
Our process starts by integrating directly with your ATS API, whether it is Lever, Greenhouse, or Ashby. We pull candidate and reference contact information automatically when a candidate moves to the reference check stage. The question sets, tailored to each specific job role, are stored and managed in a Supabase database.
The core of the system is a Python application built with FastAPI. It uses Twilio to place outbound calls to references. We use Deepgram for real-time speech-to-text transcription, achieving over 95% accuracy on clear phone lines. The conversational flow is managed by the Claude 3 Sonnet API, which processes the transcript and formulates the next question in under 800ms, ensuring a natural conversational pace.
After a call completes, the full transcript is sent to the Claude 3 Opus API for a multi-layered analysis. The model extracts key information, scores the reference's feedback against 5 predefined competencies from the job description, and generates a concise summary. This entire analysis and report generation process takes less than 20 seconds. The final output is a structured JSON object and a PDF summary.
The entire service is deployed on AWS Lambda, ensuring it only runs when needed. This keeps hosting costs under $50 per month for a team processing 500 references. The final PDF report and structured data are pushed back into the candidate's profile in your ATS via its API. Recruiters get a Slack notification and can view the results in less than 3 minutes after the call concludes.
What Are the Key Benefits?
From 30-Minute Calls to 3-Minute Reports
Our system completes three reference checks in the time it takes a recruiter to schedule one call. Final reports are in your ATS in under 180 seconds.
Pay Once, Not Per-Reference
A one-time fixed-price build with minimal monthly API fees. No per-seat license or per-reference charge that penalizes you for growing your pipeline.
You Own the System and the Code
You receive the complete Python source code in your company's GitHub repository. There is no vendor lock-in. Extend or modify it with any developer.
Consistent Questions, Unbiased Analysis
Every reference gets the exact same questions in the same neutral tone. The Claude API analysis is objective, removing recruiter bias from the summary.
Works Inside Your Existing ATS
We integrate directly with platforms like Lever, Greenhouse, or Ashby. Recruiters trigger checks and see reports without leaving their primary tool.
What Does the Process Look Like?
Scoping & ATS Integration (Week 1)
You provide read-only access to your ATS and the job descriptions for 2-3 key roles. We map the data flow and build the API connection.
Voice Agent Build (Week 2)
We build the core Python application using Twilio and the Claude API. You receive a deliverable test number to try the voice agent yourself.
Reporting & Deployment (Week 3)
We build the PDF report generation and push data back to your ATS. The system is deployed to AWS Lambda and we run 10 live test cases.
Handoff & Monitoring (Week 4+)
We monitor system performance for 30 days post-launch. You receive the full source code, deployment scripts, and a runbook for managing the system.
Frequently Asked Questions
- What factors determine the cost and timeline for this build?
- The primary factor is your ATS integration. Connecting to a system with a modern REST API like Greenhouse is faster than a legacy platform. The second is analysis complexity. Scoring against 5 traits is standard; scoring against 20 adds scope. A typical build is completed in 3-4 weeks. Book a discovery call at cal.com/syntora/discover for a detailed quote.
- What happens if a reference hangs up or the call drops?
- The system is designed for interruptions. If a call drops, it logs the partial transcript and flags the check as incomplete in your ATS. It can be configured to automatically retry once after a 60-minute delay. Your recruiter is notified via Slack of any failed or incomplete checks, so nothing gets missed.
- How is this better than an off-the-shelf tool like Checkster?
- Checkster uses email surveys, which have low response rates and yield generic, written answers. Our voice AI system creates a real conversation, allowing for more detailed, nuanced feedback. It can ask dynamic follow-up questions based on the reference's previous answer, unlike a static form, providing much deeper insight into a candidate's past performance.
- What about consent and legal compliance for recording calls?
- The system begins every call with a clear disclosure: 'This is an automated reference check call for [Candidate Name] which will be recorded. Do you consent to proceed?'. It will not continue without verbal consent. We ensure the script and process align with one-party or two-party consent laws based on your jurisdiction.
- Can we customize the questions for different roles?
- Yes. The questions are not hardcoded. They are stored in a Supabase database and mapped to specific job IDs from your ATS. You can have different question sets for engineering, sales, and operations. We provide a simple web interface for your team to add or edit questions for any role without needing to change code.
- Can we change the AI's voice or accent?
- Yes, the voice is generated by a text-to-speech API. We can select from dozens of voices, accents, and languages to match your brand or the geography of your references. During the build, we provide 3-5 voice options for you to choose from. The default is a standard American English neutral voice for broad compatibility.
Related Solutions
Ready to Automate Your Small Business Operations?
Book a call to discuss how we can implement ai automation for your small business business.
Book a Call