Back to Case Studies
    AI Recruitment · B2B SaaS · 5-Week Project

    From 8-Hour Resume Screening
    to 22 Minutes with AI

    How Akranit built an N8N + GPT-4 recruitment pipeline that eliminated 90% of manual recruiter work, cut time-to-hire from 28 days to 11 days, and reduced cost-per-hire by 64.6% — for a total infrastructure cost of $127/month.

    Screening time: 8 hrs → 22 min· 90% recruiter time saved · $3,273/month recovered
    Industry: B2B SaaS / Tech
    Timeline: 5 Weeks
    Stack: N8N · GPT-4 · Calendly
    Team size: 3 Recruiters
    The Problem

    The Spreadsheet Recruitment Trap

    ScaleHQ, a Series A B2B SaaS company in hypergrowth, had three in-house recruiters managing 6–10 open roles simultaneously. Every application triggered the same exhausting loop: download the resume PDF, manually read it, score it in a shared Google Sheet, copy notes into Slack, email the hiring manager, then chase interview slots back and forth for days. With 280+ applications per month, the spreadsheet became a single point of failure — three people editing simultaneously, version conflicts every week, and a 28-day average time-to-hire that was losing candidates to faster-moving competitors.

    Screening Time / Role
    0 hrs

    reading 30–50 resumes per open position

    Avg Time-to-Hire
    0 days

    vs 21-day industry benchmark

    Qualified Candidates Missed
    0%

    rejected at screening due to keyword bias

    Recruiter Time Cost / Month
    $0

    at $40/hr across 3 recruiters on manual tasks

    280+ applications/month spread across 3 recruiters with no structured scoring system

    24% of qualified candidates rejected at first screening due to inconsistent, keyword-based review

    Interview scheduling required 6–8 email exchanges per candidate — averaging 3.4 days

    Shared Google Sheet had version conflicts weekly; no single source of truth for candidate status

    Rejection emails manually typed, taking 5–8 minutes each and often delayed 3–4 weeks

    Hiring managers received inconsistent Slack updates — no standard format, missing context

    The AI Pipeline

    How the Recruitment Automation Was Built

    Seven N8N nodes. One webhook. Every application — whether from LinkedIn Easy Apply, Workable, or direct email — enters the same pipeline and is parsed, scored, routed, and responded to within 22 minutes, without a recruiter touching it.

    ApplicationJob Board / Email
    GPT-4 ParserResume Analysis
    Score GateIF 7+ / 4–6 / <4
    ATS AdvanceGreenhouse Stage
    CalendlyAuto-Schedule
    GmailDecline & Archive
    Weekly ReportPipeline Health
    N8N (self-hosted)$20/mo VPS

    Orchestration layer — connects job boards, Gmail, ATS, Calendly, and Slack in one visual pipeline with error handling and webhook retry logic

    OpenAI GPT-4~$35/mo

    Resume parsing node — reads raw PDF text, extracts role-specific qualifications anchored to the job description, outputs a 1–10 score with structured JSON candidate profile

    Google Sheets APIFree tier

    Real-time candidate tracking log — every application timestamped with score, status, and stage. Replaced 3 conflicting spreadsheets with a single shared source of truth

    Gmail APIFree tier

    Reads incoming applications via OAuth, triggers the N8N pipeline, and sends personalized offer / next-step / decline emails within minutes of AI scoring

    Calendly API$12/mo

    After ATS stage advance, N8N sends a personalized email with a Calendly booking link. Candidates self-schedule — zero recruiter coordination, zero email chains

    Slack WebhooksFree tier

    Sends structured hiring manager alerts to #recruiting with candidate name, role, AI score, top 3 qualifications, 2 gaps, and one-click ATS link — everything in one message

    Step-by-Step Pipeline Walkthrough

    01

    Application Intake Webhook

    N8N listens on a single HTTPS webhook. Workable and LinkedIn Easy Apply both POST new applications via their webhook APIs. A Gmail parser node catches direct email applications and normalizes them into the same payload structure — so every channel feeds one pipeline.

    02

    GPT-4 Resume Parser Node

    Raw resume text (extracted from the PDF) is sent to a GPT-4 system prompt that includes the full job description. GPT-4 outputs a qualification match score (1–10), top 3 strengths, top 2 gaps, and 3 tailored interview questions — structured JSON, every time.

    03

    Score Gate (IF Router)

    Score ≥7: hot track — ATS advance, Calendly invite, and Slack alert fire in parallel. Score 4–6: a 'human review' Slack message with full candidate profile goes to the recruiter for a quick one-click decision. Score <4: N8N sends a polite, AI-personalized decline email within 24 hours and logs to Google Sheets.

    04

    ATS Stage Advancement

    For hot candidates: N8N calls the Greenhouse API to advance the candidate to 'Recruiter Phone Screen', attaches the GPT-4 score as a custom candidate tag, and appends the full AI assessment to the candidate profile notes — the hiring manager sees a complete picture on open.

    05

    Calendly Auto-Scheduling

    Immediately after ATS advance, N8N sends a personalized email via Gmail with a Calendly booking link pre-configured to the hiring manager's availability. Average time from application received to interview booked: 31 minutes. Previous average: 3.4 days.

    06

    Hiring Manager Slack Alert

    A rich Slack message posts to #recruiting with candidate name, applied role, AI score, top 3 qualifications, 2 concerns, and a direct deep link to the Greenhouse record. The hiring manager has full context in one 30-second Slack read — no email chains.

    07

    Weekly Pipeline Health Report

    A cron node fires every Monday at 8 AM. It queries Greenhouse for all active candidates by stage, flags anyone stuck >5 days in the same stage, calculates average days-to-offer, and posts a structured pipeline summary to #recruiting-ops. Zero manual reporting.

    Candidate Pipeline (Per Month)

    How 280 applications become 6 hires — fully automated until the interview stage

    Applications received
    280
    100%
    AI-qualified (score ≥ 7)
    84
    30%
    Sent to hiring manager
    47
    16.8%
    Interviews auto-scheduled
    22
    7.9%
    Offers extended
    8
    2.9%
    Hires made
    6
    2.1%

    Recruiters touch zero resumes manually. Human judgment enters only at the interview stage.

    After 30 Days

    The Results After Month One

    Measured at the 30-day mark across 280 applications and 8 positions filled. All metrics compared to the 30-day pre-deployment baseline.

    Screening Time / Role
    0 min

    ↓ from 8 hours (↓ 95.4%)

    Time-to-Hire
    0 days

    ↓ from 28 days (↓ 60.7%)

    Manual Work Eliminated
    0%

    35 hrs/wk → 3.5 hrs/wk

    Monthly Infra Cost
    $0

    vs $3,400 in recruiter time

    MetricBeforeAfter (30 days)Change
    Resume screening time per role8 hours22 minutes↓ 95.4%
    Time-to-hire (average)28 days11 days↓ 60.7%
    Qualified candidates missed24%2.1%↓ 91.3%
    Interview scheduling (email rounds)6–8 emails0 (automated)↓ 100%
    Manual recruiter hours / week35 hrs3.5 hrs↓ 90%
    Rejection email time per candidate5–8 min each0 (automated)↓ 100%
    Monthly cost (infra vs recruiter time)$3,400$127↓ 96.3%
    Cost-per-hire$2,400$850↓ 64.6%
    Lessons Learned

    What Nobody Tells You About AI Recruitment

    Five weeks of deployment, 280 applications processed, and several surprises that no AI recruitment tutorial covers.

    01Prompt Engineering

    GPT-4 must be anchored to the job description — not general 'good candidate' criteria

    Our first prompt asked GPT-4 to score for 'software engineering experience.' It rewarded long resumes and penalized career-changers with shorter but highly relevant backgrounds. After anchoring the prompt to the exact job requirements and adding hiring manager culture-fit signals, accuracy jumped from 71% to 96% match with human judgment during our shadow testing week.

    02Infrastructure

    PDF text extraction is the silent dependency that breaks everything

    The pipeline worked perfectly until we received image-based PDFs — scanned documents with no extractable text. GPT-4 received an empty string and scored them all 3/10, sending qualified candidates to the decline track. Fix: a PDF-to-text OCR preprocessing node added before the GPT-4 node at $8/month. Always test your pipeline with edge-case file types before go-live.

    03Workflow Design

    Mid-score candidates (4–6) need a 'hold queue', not auto-rejection

    In week one, every 4–6 score auto-declined. By week two, the lead recruiter noticed 3 strong candidates for a different open role had been declined. We rebuilt the mid-range branch: N8N creates a 'talent pool' record in Airtable and notifies the recruiter with a one-click Keep/Decline Slack button. That pool has since generated 4 fills for future roles.

    04Product Insight

    The scheduling bottleneck is as painful as the screening bottleneck — automate both

    We originally scoped the project as screening-only. After the first week, the team flagged they were still spending 2–3 hours per day on interview scheduling emails. Adding the Calendly API integration took 4 hours of N8N work. It eliminated the remaining manual bottleneck entirely and delivered more measurable ROI than the screening automation by itself. Scope both from day one.

    05Brand Impact

    Rejection emails sent within 24 hours improved Glassdoor employer reviews

    The old process: candidates waited 3–4 weeks with no response. The new process: GPT-4-personalized rejection emails sent within 24 hours of application. After 60 days, ScaleHQ received 12 unsolicited Glassdoor reviews mentioning 'fast communication' and 'respectful process.' Automated rejection improved employer brand more than the team expected. Candidate experience is a product, not an afterthought.

    06Deployment Strategy

    Run 2 weeks of 'shadow mode' before going live — it's change management, not overhead

    Before switching live, we ran the pipeline for 2 weeks in shadow mode: it processed every real application but posted results to a private Slack channel instead of acting on them. Recruiters compared AI decisions to their own and flagged disagreements. This caught 4 edge cases, built recruiter confidence, and produced zero pushback when we went live. Shadow mode is the difference between adoption and rejection.

    Built by Akranit

    Want This Built for Your Recruiting Team?

    We build the same AI recruitment pipelines — tailored to your ATS, job sources, and hiring criteria — typically live in 4–5 weeks.