Build a Voice AI That Prequalifies Recruiting Candidates

Most recruiting teams try to automate screening with basic tools that fail in practice. Standard IVR (Interactive Voice Response) systems are the first attempt. They use rigid "press 1 for yes" menus that cannot handle natural conversation. A candidate who says "I have about five years of experience" instead of just "five" breaks the workflow, leading to a 60% drop-off rate and useless data.

A second approach is using a scheduling tool with screening questions in the form. This just pushes the manual work around. Instead of screening on the phone, the recruiter now sifts through dozens of form submissions filled with unqualified candidates who can book calendar slots. This clogs the pipeline and wastes interview time with people who don't meet basic requirements, like having a valid driver's license for a field position.

These tools fail because they are not conversational. They are glorified forms. They cannot ask clarifying questions, understand context, or interpret nuanced answers. Real qualification requires a system that can have a basic, stateful conversation, which is an engineering problem, not a configuration problem.

Our approach begins with a comprehensive discovery phase. We would start by auditing your existing prequalification scripts and processes, mapping your exact requirements to a detailed conversational flow graph. This involves identifying the critical data points you need to collect and defining how they would map to fields in your Applicant Tracking System (ATS), whether it is Greenhouse, Lever, or a custom-built system. Leveraging our experience with large language models, including building document processing pipelines using Claude API for financial documents, we would then script conversational paths designed to feel natural and handle a range of candidate responses.

The core of the system would be a robust Python application, typically built with FastAPI. It would utilize a Supabase Postgres database to maintain conversational state for each candidate, enabling the AI to reference previous answers in follow-up questions for a coherent interaction. For telephony, we would integrate directly with Twilio's API to manage outbound calls. AI responses, generated by the Claude API, are designed for rapid turnaround to maintain a fluid conversation experience.

Deployment would typically be as a serverless function on AWS Lambda. When a new candidate applies through your ATS, a webhook would trigger the Lambda function, initiating the prequalification call. After the call, the service would generate a structured summary and a full transcript, then use the ATS API to post them directly to the candidate's profile. We would engineer this entire process, from call completion to ATS update, for efficient and timely delivery.

We prioritize system observability and reliability. This would include implementing structured logging with `structlog`, sending JSON logs to AWS CloudWatch. This detailed logging enables us to set up specific, actionable alerts. For example, if the ATS API integration fails repeatedly or the call completion rate falls below predefined thresholds, a notification would be sent to a designated Slack channel for immediate review, ensuring you have full visibility into the system's performance and health.