Problem
The HR company, a mid-sized recruitment agency placing candidates in technical and creative roles, was losing good candidates to faster competitors. Their recruiters were spending the majority of their time on the lowest-value part of the pipeline: initial screening. For every open role, a recruiter had to manually review dozens of resumes, send screening questionnaires, wait for responses, evaluate answers, and shortlist candidates — a process that routinely took five to seven days before a candidate ever spoke to a human.
The problem was compounded by volume. The company was handling a significant number of open roles at any given time, each generating many applications. The screening backlog meant that high-quality candidates — who are typically also interviewing with multiple companies — often accepted offers elsewhere before the company could even schedule a first interview. The agency’s time-to-interview metric lagged behind the industry average, and client companies were beginning to express frustration with the pace of candidate delivery.
The screening itself was also inconsistent. Different recruiters evaluated candidates against different implicit criteria, which led to variable quality in the shortlists delivered to clients. The agency needed a system that could screen faster, more consistently, and free recruiters to do what they do best: build relationships with candidates and clients and close offers.
Solution
We built an AI-powered screening agent that automates the initial evaluation of every applicant. When a candidate applies through the Greenhouse-sourced job posting, a LangChain agent reviews the resume against the job requirements, generates a tailored set of screening questions specific to the role and the candidate’s background, and sends them to the candidate via email — all within minutes of application.
The agent evaluates the candidate’s responses using a structured rubric defined by the recruiting team for each role. It scores the candidate on required skills, experience level, and role fit, and produces a shortlist with a rationale for each score. Recruiters receive the shortlist in Greenhouse, complete with the agent’s analysis and the full conversation transcript, so they can review the reasoning and override the score if they disagree — keeping the human firmly in control of the final decision.
n8n orchestrates the entire flow: it listens for new applications via the Greenhouse API, triggers the screening agent, manages the email conversation with the candidate, writes the evaluation back to Greenhouse, and — for candidates who pass the initial screen — automatically proposes interview times based on the recruiter’s calendar. This end-to-end automation compressed what was a five-day process into a matter of hours.
Architecture
The system is composed of three layers: the screening agent, the orchestration layer, and the ATS integration. The LangChain agent is a stateless service that receives a candidate profile and job description as input and returns screening questions, evaluation scores, and a recommendation as output. Its statelessness makes it easy to scale during high-volume hiring pushes and simple to test against historical hiring decisions.
n8n manages all stateful operations and integrations. It tracks the conversation with each candidate, handles email delivery and response collection, calls the agent service for each evaluation step, and synchronizes everything with Greenhouse through its API. The visual workflow editor allowed the recruiting team to adjust the screening rubric, add new question types, and modify the shortlist criteria without engineering support.
Greenhouse remains the system of record for the entire recruitment pipeline. The n8n workflows create candidate records, log screening conversations as activities, update candidate stages, and schedule interviews — all through the Greenhouse API — so that recruiters never have to leave their existing tool and all data flows into the agency’s standard reporting.
Workflow Diagram
Candidate applies (Greenhouse job posting)
│
▼
n8n Webhook (Greenhouse API)
│
├──► LangChain Agent ──► Review resume vs. job requirements
│ │
│ ▼
│ Tailored screening questions ──► Email to candidate
│ │
│ ▼
│ Candidate responses
│ │
├──► LangChain Agent ──► Score against rubric + rationale
│ │
│ ▼
└──► Greenhouse API ──► Shortlist + transcript + scores
│
▼
Recruiter review (override if needed)
│
▼
Auto-schedule interview (if shortlisted)The diagram shows the complete flow from application to scheduled interview. The key design principle is that the agent recommends and the recruiter decides — the AI does the heavy lifting of evaluation, but a human reviews and approves every shortlist before a candidate moves forward.
Technologies
LangChain powers the screening agent, providing the framework for resume analysis, question generation, and response evaluation. The agent uses a combination of structured prompts — encoding the recruiting team’s rubric — and retrieval over a database of historical successful placements to calibrate its scoring against what has actually worked for the agency in the past.
n8n is the orchestration layer that ties the agent to the agency’s existing tools. Its native Greenhouse integration node eliminated the need for custom API code, and its visual workflows let the recruiting operations lead modify the screening process — adding a new question set for a new role type, for example — without a developer deployment.
The Greenhouse API is the integration point with the agency’s applicant tracking system. OpenAI provides the language understanding capabilities for resume parsing, response evaluation, and the natural-language rationale generation that accompanies each candidate score, making the agent’s reasoning transparent and reviewable.
Results
Screening time was reduced substantially, from an average of five to seven days per role down to under 24 hours. The agent handles the entire initial evaluation — resume review, question generation, response collection, and scoring — autonomously, and recruiters now spend their time reviewing shortlists and conducting interviews rather than doing first-pass screening. For a large pipeline, this recovered the equivalent of significant screening labor.
Time-to-interview improved dramatically, which was the single most impactful metric for the agency’s competitive positioning. High-quality candidates who previously accepted offers elsewhere while waiting for the company’s first response now receive a screening conversation within hours and an interview invitation within a day. The agency’s client companies reported a noticeable improvement in the pace of candidate delivery and a reduction in the number of top candidates lost to competitors.
Candidate quality, as measured by client companies’ post-placement satisfaction surveys, improved noticeably. The consistent, rubric-based screening eliminated the variability that came from different recruiters applying different implicit criteria, and the agent’s calibration against historical successful placements meant the shortlists were both more relevant and more defensible to clients.
Lessons Learned
The most important lesson was that AI should recommend, not decide. The first version of the system auto-rejected candidates below a score threshold, and the recruiting team was uncomfortable with candidates being eliminated without human review. Once we changed the design so that the agent scores and recommends but a recruiter reviews and approves every shortlist, adoption was immediate and the team’s trust in the system grew quickly. The lesson: in high-stakes people decisions, keep the human in the loop.
A second lesson was the value of calibrating the agent against historical data. The agent’s scoring became dramatically more accurate once we trained it on the agency’s past successful and unsuccessful placements, rather than relying on generic job-description matching. This domain-specific calibration was the difference between a generic tool and one that felt like it understood the agency’s specific standards.
Finally, we learned that integrating with the existing ATS — rather than building a parallel screening system — was critical for adoption. Because every score, transcript, and shortlist appeared in Greenhouse where recruiters already worked, there was no new tool to learn and no data silo. The system enhanced the existing workflow rather than disrupting it, which is why the recruiting team embraced it rather than resisting it.