Build a Voice AI Agent to Schedule Interviews
The top voice AI solution is a custom agent that calls candidates to confirm availability. It connects directly to your ATS and calendars, booking interviews without human intervention.
The scope is defined by your Applicant Tracking System's (ATS) API and your scheduling rules. A firm booking simple one-on-one screenings with a modern ATS like Greenhouse is a straightforward build. A firm that needs to coordinate multi-person panel interviews across different time zones requires more complex logic.
We built a voice agent for a 7-person IT staffing firm that places 30 contractors a month. Recruiters were losing hours to phone tag trying to book initial screens. The system went live in 4 weeks and now schedules 90% of their screening calls automatically, often within 10 minutes of a candidate applying.
What Problem Does This Solve?
Most small staffing agencies start by sending candidates Calendly links. This is passive and puts the burden on the candidate. For high-demand roles, less than 50% of candidates will book a time, and the ones who do not are often the most valuable. This approach cannot proactively follow up or escalate when a top candidate goes dark.
A few agencies try off-the-shelf call center AI tools from providers like Dialpad. These systems are designed for sales or support, not recruiting. They are expensive, with per-seat fees often exceeding $75/month for AI features. Worse, they cannot read real-time availability from multiple recruiter calendars or connect deeply with an ATS like Bullhorn. The agent can take a message, but it cannot complete the primary job: booking a confirmed time on the calendar.
This forces a broken workflow. A recruiter has their ATS open, their calendar in another tab, and their phone dialer in a third. They spend their day cross-referencing information and leaving voicemails instead of talking to qualified candidates. The core bottleneck, the manual act of scheduling, remains untouched.
How Does It Work?
We start by building a direct, read-write API connection to your ATS and every recruiter's Google or Microsoft calendar. We use Python with the httpx library to make asynchronous calls, allowing the system to pull a list of 100 candidates to call and their associated recruiter's availability in under 2 seconds.
The agent's conversational logic is powered by the Claude API, managed by a state machine that tracks the conversation from introduction to confirmation. We use a specialized text-to-speech API that provides low-latency, natural-sounding audio, keeping the round-trip response time under 800ms. The system is designed to handle ambiguous responses like "sometime next Thursday afternoon" by offering specific slots like "Does 2:00 PM or 4:00 PM work?"
The entire application is built with FastAPI and deployed on AWS Lambda. This serverless function is triggered by a webhook from your ATS whenever a candidate is moved to the 'Phone Screen' stage. This architecture is efficient, keeping infrastructure costs under $50 per month for a volume of over 1,000 outbound calls. There are no servers to manage.
Every call attempt, transcript, and outcome (booked, voicemail, no answer, follow-up required) is logged to a Supabase database. We provide a simple dashboard to review performance, showing the booking success rate and call durations. This data allows for tuning the agent's script in the first 30 days to improve performance, reducing the average time-to-schedule from 2 days to under 15 minutes.
What Are the Key Benefits?
Schedule First-Round Calls in Minutes
The agent calls candidates within 60 seconds of them entering the screening stage, engaging them while interest is highest.
Pay for Calls, Not for Recruiter Seats
A single fixed-price build with low monthly hosting costs replaces expensive per-user subscriptions on call center platforms.
You Own the Agent, Code, and Data
You receive the full Python source code in your GitHub repository and a Supabase instance with all call data. No vendor lock-in.
Alerts When Conversations Go Wrong
We use structlog for structured logging. If a call fails to book after 3 attempts, it triggers a Slack alert with the transcript for human review.
Connects Natively to Your ATS
Direct API integration with Bullhorn, Greenhouse, or Lever. The agent reads candidate data and writes back interview details automatically.
What Does the Process Look Like?
Week 1: Scoping and API Access
You provide read/write access to your ATS and recruiter calendars. We map your exact scheduling workflow and define the agent's conversation logic.
Weeks 2-3: Core Agent Development
We build the voice agent, integrate the APIs, and test against 20+ common scheduling scenarios. You receive call recordings for review.
Week 4: Deployment and Live Testing
We deploy the system on your AWS infrastructure and run it on a small batch of 50 live candidates while you monitor the results.
Post-Launch: Monitoring and Handoff
For 30 days, we monitor every call and tune the prompts for performance. You receive a runbook detailing how to update scripts and monitor costs.
Frequently Asked Questions
- How much does a custom voice agent cost?
- The build is a fixed price, determined by the complexity of your ATS integration and scheduling rules. A system for one-on-one interviews with a modern ATS API takes about 4 weeks. Panel interviews or integrations with legacy systems take longer. We provide a fixed quote after a 30-minute discovery call.
- What happens if the AI can't understand a candidate?
- If the agent fails to understand the candidate twice, it gracefully ends the call. It says, 'I'm having trouble understanding. A recruiter will reach out to you shortly.' It then logs the call transcript and sends an alert to a designated Slack channel for a human to follow up. This prevents frustrating loops for the candidate.
- How is this different from using a Virtual Assistant (VA) service?
- A VA is a human who works limited hours and can handle one call at a time. Our AI agent works 24/7, can make 100 concurrent calls, and costs a fraction of a VA's hourly rate after the initial build. The agent's performance is consistent and logged, unlike the variable quality of different VAs.
- Does it work for non-English speaking candidates?
- Yes. The core logic is language-agnostic. We can configure the agent with prompts and voice synthesis in other languages like Spanish or French. Each additional language is scoped as a small add-on to the main project, typically adding 2-3 days to the build for translation and testing of the new scripts.
- What about call recording consent and compliance?
- The agent is scripted to handle compliance. It begins every call with a disclosure like, 'This is an automated call from [Agency Name] on a recorded line.' This is standard practice for TCPA compliance. We work with you to ensure the script meets the legal requirements for your specific region.
- Can we change the agent's voice or personality?
- Yes. We can select from a library of dozens of pre-trained voices to match your brand. We can also adjust the Claude API prompts to make the agent more formal, casual, or enthusiastic. You review and approve the voice and a sample script before we begin the main build.
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