AI Integration with Lever Video Interview Analysis

AI integrates into Lever's video interview workflow by connecting to the Lever Video Interview feature (or integrated third-party tools like Willo or Spark Hire) via webhook or API. When a candidate completes a recorded interview, an event payload containing the interview id, candidate id, and secure media URL is sent to an AI processing queue. The AI system then performs three core functions: transcription of the audio/video, sentiment and tone analysis across responses, and key response extraction based on the predefined question set from the Lever requisition. The results are structured into a summary object.

The processed insights are written back to Lever via the Candidate Notes or Custom Candidate Fields API, attaching the AI-generated summary directly to the candidate's profile. This allows recruiters to review a concise, searchable transcript alongside sentiment flags (e.g., high confidence, uncertain clarification) before watching the full recording. For high-volume roles, the system can implement a scoring or triage layer, where responses are evaluated against role-specific rubrics, and candidates are automatically tagged or advanced in the pipeline based on configurable thresholds, all while maintaining a full audit log in your AI platform.

Rollout should begin with a human-in-the-loop pilot, where AI summaries are provided as a recruiter aid without automated scoring. Governance must address data residency (ensuring video processing complies with regional data laws), bias auditing of the analysis models, and candidate transparency—clearly communicating how AI is used in the process. This integration turns a manual, time-consuming review step into a structured, data-informed evaluation, helping recruiters prioritize their time and maintain consistency across a large candidate pool.

The integration is triggered when a candidate completes a one-way video interview via a platform like HireVue, Spark Hire, or Willo that is configured to send a webhook to your orchestration layer. This event payload contains the candidate ID, Lever opportunity ID, and a secure URL to the video file. An AI workflow service then processes the video: first, it extracts audio for transcription via a service like OpenAI Whisper or Google Speech-to-Text; second, it runs the transcript through an LLM for sentiment analysis, key response extraction, and competency scoring based on predefined job criteria. The structured output—transcript, sentiment scores, extracted answers, and a summary—is then posted back to Lever via the REST API, typically writing to a custom candidate field (e.g., video_interview_summary) or appending a private note to the candidate's profile for recruiter review.

For governance and scale, this workflow should be built with idempotency and audit trails. Each processing job should generate a unique correlation ID, log processing steps, and store the raw transcript and AI outputs in a secure object store (like S3) linked to the candidate record. This ensures you can explain scoring decisions and reprocess if models improve. A human-in-the-loop checkpoint can be configured where low-confidence analyses or extreme sentiment flags are routed to a recruiter dashboard for manual review before the data populates Lever, maintaining control over the candidate experience.

Rollout typically starts with a pilot for a specific job family. Recruiters are trained to use the new video interview summary panel in Lever's candidate profile, which highlights key quotes and sentiment trends. The architecture is designed to handle batch processing during off-peak hours to manage API rate limits and cost. For teams using Lever's E-Mail Dropbox or other sources, the same pattern can be extended to analyze candidate-submitted video links, creating a unified video assessment layer across all hiring channels. For a deeper look at orchestrating multi-step AI workflows, see our guide on AI Agent Builder and Workflow Platforms.

The integration architecture is built around Lever's webhooks and REST API. When a candidate completes a recorded video interview (e.g., via a platform like Willo or Spark Hire), a webhook payload is sent to a secure, queued ingestion endpoint. This triggers an asynchronous workflow: the video file is retrieved via a secure, temporary URL, processed for transcription and analysis by the AI model, and the results (transcript, sentiment markers, extracted key responses) are written back to Lever as private notes on the candidate profile or to custom fields on the application. All PII, including video files and transcripts, is encrypted in transit and at rest, with processing logs kept for audit trails to meet GDPR/CCPA and internal compliance requirements.

Rollout should follow a phased, metrics-driven approach. Phase 1 (Pilot): Select a single recruiting team and 2-3 non-critical roles. Enable analysis on a 100% sample but keep results visible only to a core group of super-users for feedback. Phase 2 (Controlled Expansion): Expand to more teams and roles, implementing a human-in-the-loop review where the AI-generated summary is presented alongside the raw video for recruiter validation. Phase 3 (Full Scale & Automation): After validating accuracy and user adoption, automate the writing of summaries to candidate profiles and trigger workflows—like moving high-sentiment candidates to a "Ready for Review" stage or flagging incomplete responses for follow-up.

Governance is critical. Establish clear rules: AI analysis should augment, not replace, human judgment. Define which roles and stages are eligible. Implement regular quality audits by sampling analyzed interviews against human reviews to monitor for model drift or bias. Use Lever's permission sets (e.g., Interviewer, Recruiter, Hiring Manager) to control who sees AI-generated insights, ensuring sensitive analysis isn't prematurely exposed. This controlled, phased approach de-risks the integration, builds organizational trust, and allows you to measure concrete impact—like reduction in average review time per video or improved consistency in feedback—before scaling company-wide.