AI Integration for Greenhouse Resume Screening

AI integrates into Greenhouse's resume screening workflow primarily through its webhook and REST API surfaces. The most common pattern is a candidate.created or candidate.stage_change webhook triggering an AI processing pipeline. This pipeline ingests the candidate's resume (often from the Greenhouse attachments object or a linked URL), parses it for skills, experience, and education, and then scores the candidate against the specific job_post or job requisition. The resulting match score and extracted metadata are written back to Greenhouse via the API, typically populating custom fields like ai_match_score, primary_skills, or years_experience on the candidate record. This creates an automated, data-rich first impression before a recruiter even opens the profile.

A production implementation involves several key components: a secure queue (e.g., Amazon SQS or RabbitMQ) to handle webhook bursts, an AI service for document parsing and LLM-based scoring, and a sync service to update Greenhouse. Governance is critical; you must implement a human review loop for low-confidence parses and audit logging for all score adjustments. The impact is operational: reducing manual resume review from hours to minutes for each requisition, ensuring consistent application of screening criteria, and allowing recruiters to prioritize high-potential candidates from day one. For a deeper look at orchestrating these workflows, see our guide on AI Integration for Greenhouse Hiring Workflow Automation.

Rollout should be phased. Start with a pilot on a single, high-volume job family to tune your scoring model and integrate feedback from recruiters. Use Greenhouse's permission sets (job_admin, interviewer) to control who sees the AI-generated scores. It's also advisable to maintain the original, unaltered resume and notes for compliance. The goal isn't full automation but augmentation—giving recruiters a powerful copilot that handles the tedious first pass, so they can focus on human-centric evaluation and candidate experience.

The integration is triggered via a Greenhouse webhook on candidate stage change (e.g., move to "Application Review") or new application submission. The webhook payload contains the candidate ID, job ID, and relevant metadata. A lightweight middleware service—often deployed as a serverless function or containerized microservice—receives this event, fetches the full candidate record (including resume PDF/Word doc and job requisition details) via the Greenhouse Harvest API, and queues the documents for AI processing. This decoupled design ensures the ATS remains responsive, and failed processing attempts can be retried from the queue.

The core AI processing involves two parallel workflows: document parsing & skills extraction and semantic match scoring. The resume is parsed to extract structured fields (name, contact), work history, education, and certifications. Concurrently, a transformer-based model performs a semantic comparison between the candidate's resume text and the job description, generating a match score (e.g., 0-100) and a list of key matching skills and potential gaps. This score, along with extracted entities, is formatted into a JSON payload that conforms to Greenhouse's Custom Field or Scorecard API schema. The middleware then makes an authenticated PATCH request to update the candidate record, populating fields like ai_match_score and ai_skills_summary for recruiter visibility.

Critical guardrails are implemented at multiple layers. A human-in-the-loop approval can be configured for scores below a certain threshold or for specific, high-priority roles, requiring a recruiter to review before the score is posted. All AI-generated content and scores are written to a dedicated audit log (outside Greenhouse) with timestamps, model versions, and input data hashes to support compliance reviews and model drift detection. To prevent bias amplification, the system can be configured to exclude demographic proxies (e.g., names, universities, dates) from the semantic matching input, focusing solely on skills and experience language. The entire data flow is encrypted in transit, and PII is never persisted in the AI service's logs beyond the processing window required for the API call back to Greenhouse.

A production-ready integration for Greenhouse resume screening is built on a secure, event-driven architecture. The core flow is triggered by a Greenhouse application.created or application.updated webhook, which pushes the candidate's resume file URL and application metadata to a secure processing queue. An AI agent retrieves the document via Greenhouse's secure API, performs parsing and analysis (skills extraction, experience matching, role fit scoring), and posts the results back to the candidate's Greenhouse profile using custom fields like ai_match_score or top_skills. All PII is processed in-memory or within a VPC; no candidate data is persisted in long-term external storage without explicit governance controls. API calls are authenticated via Greenhouse API keys stored in a secrets manager, and all data movements are logged for a full audit trail.

Rollout should follow a phased, risk-managed approach. Phase 1 (Pilot): Start with a single, non-critical job requisition. The AI scores candidates but does not auto-reject; scores are visible only to a core team of recruiters in a custom field for validation and bias testing. Phase 2 (Controlled Expansion): Expand to 5-10 similar roles, implementing a human-in-the-loop rule—for example, auto-advance candidates scoring above 85, flag candidates below 40 for review, and require manual decision for scores in the middle range. Configure Greenhouse scorecard workflows to surface the AI score alongside human evaluations. Phase 3 (Scale & Optimize): Integrate scores into Greenhouse reporting to measure impact on time-to-fill and quality-of-hire. Introduce more sophisticated workflows, such as triggering automated, personalized status updates for high-scoring candidates or alerting recruiters to potential high-match candidates from past applications.

Governance is critical. Establish a review board that meets quarterly to audit the model's scoring outcomes for adverse impact across gender, ethnicity, and other protected classes using Greenhouse's EEO data (where available). Maintain a versioned prompt library and a log of all scoring inputs and outputs to enable explainability—if a recruiter questions a score, you can reconstruct the AI's reasoning. Finally, ensure your integration respects Greenhouse's permission model; scores should only be written to fields and visible to users based on their existing Greenhouse role-based access control (RBAC). This controlled, transparent approach allows you to capture efficiency gains—reducing manual screen time by 50-70% for inbound applications—while maintaining human oversight and compliance.