94% Get Rejected. I Automated The Part HR Won't Touch

By Lazar Milicevic · Published September 8, 2026 · 8 min read

Every ATS on the market automates the "yes." The offer letter, the onboarding pack, the DocuSign flow. Nobody ships the "no." That's why 94% of your applicants get ghosted — and it's why your Glassdoor page slowly rots while your hiring manager pretends the inbox doesn't exist. Here's the exact n8n workflow I run for clients, node by node, prompt included.

The math nobody in HR wants to run

A typical small business hiring cycle looks like this: 650 applications for one open role over two weeks. 38 shortlisted. 12 interviewed. 1 hired. That leaves 612 humans sitting in an inbox waiting for a reply that, industry-wide, arrives less than half the time. In small business hiring specifically, ghost rates run closer to 80%.

Those 612 people are not a rounding error. They write reviews. They talk to their network. They remember. In one client's audit, 4 of the previous quarter's Glassdoor complaints traced directly to "never heard back after interview" — not the rejection itself, the silence.

The reason no HR platform ships real automated rejections is legal exposure. One phrase adjacent to age, gender, disability, national origin, family status, or health, and you've handed a plaintiff's attorney a template. So the big platforms give you a canned merge-field email — "Dear {FirstName}, we've decided to move forward with other candidates" — and call the problem solved. It isn't. That template is exactly why candidates hate rejection emails.

The build below sends emails that reference a specific detail from the candidate's resume, pass an EEOC-style guardrail check, and land in the inbox from the hiring manager's own address. Real numbers from the last run: 612 emails, 8 minutes 41 seconds, 94 cents in OpenAI cost, zero complaints.

The n8n workflow: five nodes, no ATS required

The whole thing runs on n8n because it's self-hostable, the JSON exports cleanly, and clients can own it forever without paying a per-seat SaaS tax. No ATS integration — it hooks straight into Gmail.

The five nodes

• Gmail Trigger — watches a label called candidates-rejected
• Classifier (GPT-4o-mini) — extracts role and interview stage
• Personalizer (GPT-4o-mini) — writes the email, references one real detail
• Guardrail (GPT-4o-mini) — reviews draft against a protected-class blocklist
• Send + Log (Gmail + Airtable) — sends via hiring manager, logs to Silver Pool

The trigger is deliberately dumb. The hiring manager reads a candidate's application, decides no, drags the email to a label called candidates-rejected. That's the entire UI. No new tool to learn, no browser tab to keep open, no ATS to log into. Every hiring manager already knows how to drag an email into a label.

{
  "node": "Gmail Trigger",
  "event": "labelAdded",
  "label": "candidates-rejected",
  "pollInterval": "1m",
  "includeAttachments": true
}

Node 2 — the classifier that decides what kind of "no" this is

Someone who did three interviews deserves a very different email than someone whose resume didn't match the job title. The classifier's job is to figure out which bucket the candidate is in and pull the role they applied for, using only the email thread and the attached resume.

GPT-4o-mini handles this at roughly $0.15 per million input tokens. A typical application email plus a two-page resume runs about 1,800 tokens in, 80 tokens out. That works out to around $0.0003 per classification — a hundred candidates costs three cents.

system: |
  You are a hiring pipeline classifier. Return strict JSON.
user: |
  Read the email thread and resume below. Return:
  {
    "role": "<the job title they applied for>",
    "stage": "resume | phone_screen | interview | final_round",
    "strongest_skill": "<one skill, 1-3 words>"
  }
  Email: {{ $json.thread }}
  Resume: {{ $json.resume_text }}
temperature: 0
max_tokens: 120

Stage matters because it dictates tone. A resume-stage rejection is short and encouraging. A final_round rejection is longer, warmer, and explicitly leaves the door open for future roles. Hard-coding these tone shifts in the next node makes the outputs feel human without any randomness in the structure.

Node 3 — the personalization prompt that changes everything

This is the node that makes the difference between an email candidates delete and an email candidates reply to. The instruction is narrow: read the resume, find one concrete, verifiable detail — a project, a skill, a company, a certification — and reference it naturally in the second paragraph. Not "we loved your background." Something like:

Your work migrating the billing system at your previous role stood out — we're keeping your profile in mind for a similar opening we expect to have in Q2.

That one line is why the whole system works. Candidates reply. They thank the hiring manager. They ask to stay in touch. In one client's 60-day window, 41 out of 612 rejected candidates opted into the "keep me in mind" pool, and one of them got hired six weeks later for a role that would have cost roughly $4,000 in agency fees.

system: |
  You write rejection emails for a hiring manager.
  Tone: warm, direct, human. No corporate hedging.
  Never mention age, gender, family status, health,
  disability, religion, national origin, or appearance.
  Never invent facts. Only reference details present in
  the resume text.

user: |
  Write a rejection email for this candidate.

  Role: {{ $json.role }}
  Stage reached: {{ $json.stage }}
  Hiring manager: {{ $json.manager_name }}
  Resume: {{ $json.resume_text }}

  Requirements:
  - 3 short paragraphs
  - Paragraph 2 must reference ONE specific,
    verifiable detail from their resume
  - Close with the Silver Pool opt-in line:
    "If you'd like us to keep your profile on file
    for future roles, just reply YES to this email."
  - Sign as {{ $json.manager_name }}

temperature: 0.4
max_tokens: 220

Temperature 0.4 is the sweet spot — enough variation that 612 emails don't sound identical, not enough for the model to hallucinate a job the candidate never held. Token cap at 220 keeps each email under $0.002. Times 612, that's roughly $1.10 — the actual bill for the run in the headline came in at 94 cents because most emails ran shorter.

Node 4 — the guardrail that keeps you out of court

This is non-negotiable. Before any email goes out, a second LLM call reviews the draft against a blocklist of protected-class language. If anything flags, the email is rewritten and re-checked. If it flags twice, it goes to a human review queue instead of sending.

The EEOC's guidance on employment discrimination is the reference point for what to block — see the EEOC's overview of prohibited employment policies for the full list of protected categories. The blocklist below is a starting point, not legal advice; anything you deploy in production should be reviewed by an employment attorney in your jurisdiction.

system: |
  You are a compliance reviewer for hiring communications.
  Flag any language that references, implies, or is
  adjacent to:
  - age (including "fresh", "energetic", "seasoned",
    "digital native", "overqualified")
  - gender or gender presentation
  - family status, pregnancy, childcare
  - health, disability, mental health, stamina
  - religion, national origin, accent, name pronunciation
  - physical appearance or characteristics
  - arrest or conviction history

  Return JSON:
  {
    "pass": true | false,
    "flagged_phrases": ["..."],
    "rewritten_email": "<safe version if pass=false>"
  }

user: |
  Review this email:
  {{ $json.draft_email }}
temperature: 0

In the 612-email batch, the guardrail flagged 9 drafts. One of them said the candidate would "bring fresh energy to the team" — "fresh" and "energy" are both age-adjacent per plaintiff-side employment attorneys. Rewrite, re-check, send. That single guardrail node is what makes the difference between a workflow you can hand a client and one that puts them at legal risk.

Node 5 — send from the hiring manager and log to the Silver Pool

The email sends via Gmail using the hiring manager's own OAuth credential, so it lands in the candidate's inbox from a real person with a real signature. No no-reply@. No careers@. That alone doubles reply rates in my testing.

Simultaneously, the workflow writes a row to an Airtable base I call the Silver Pool — every rejected candidate, tagged with role, stage, strongest skill, source email, and a checkbox for "open to future roles." The opt-in is a one-click mailto link in the email that pre-fills a reply with "YES" in the body. No form, no login, no friction.

The compounding effect is the part that surprised me. Over 60 days after that first run: 41 candidates in the Silver Pool, 3 re-applied for later openings, 1 hired. That one hire alone paid for the entire build ten times over versus using an agency, and it was someone who genuinely wanted to work there because the rejection email had been human.

Why bizflowai.io helps with this

The rejection workflow above is one of half a dozen "inbox jobs that nobody wants to do" I build for small business clients — the others usually being invoice chase-ups, lead qualification, and post-meeting follow-ups. At bizflowai.io I ship these as self-hosted n8n workflows the client owns outright, including the guardrail prompts, the Airtable schema, and a two-page runbook so any team member can maintain it after handoff. The point isn't to sell you a subscription — it's to give a hiring manager her Fridays back.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.