47 Seconds: The Slack Referral Agent That Skips Recruiters

By Lazar Milicevic · Published July 29, 2026 · 7 min read

Your best hire this quarter is sitting in a #referrals channel right now. Somebody dropped a resume three weeks ago, tagged nobody, and Sarah — a warm intro, senior at Stripe — has already accepted an offer somewhere else. This isn't a people problem. It's a routing problem, and n8n fixes it in about twenty minutes.

Why referrals rot in Slack while recruiters chase cold InMails

Referrals convert 4–5x better than cold sourcing and referred hires stay roughly 2x longer, yet most SMB pipelines lose them because the channel has no owner. Your ATS doesn't listen to Slack. Your recruiter doesn't live in Slack. The highest-quality source in your funnel has zero automation attached to it.

Meanwhile every big HR-tech vendor is selling the opposite: AI sourcing agents that scrape LinkedIn and blast cold outreach. Reply rates hover around 9%, your domain reputation tanks, and you still ignored the warm candidate your engineer vouched for. The fix isn't more sourcing. It's a listener on the channel you already have.

Here's what the actual latency looks like on the manual path vs. the automated one, from a client build I shipped last month:

Step 1–2: Slack trigger and a single parsing prompt

The trigger is a Slack node listening to the Events API on one channel. That's it. You need a small Slack app with two scopes: channels:history and files:read. The trigger fires on every new message and hands you the text, the referrer's user ID, and any file attachments (usually a PDF).

Don't overbuild the parser. One LLM node with a strict JSON schema handles both the message text and the PDF content:

{
  "candidate_name": "string",
  "current_company": "string | null",
  "current_role": "string | null",
  "referred_for": "string | null",
  "years_experience": "number | null",
  "email": "string | null",
  "linkedin_url": "string | null",
  "raw_notes": "string"
}

Prompt in one line: "Extract the following fields from the Slack message and attached resume. If a field is not present, return null. Do not guess." That's the whole parsing layer. No dedicated resume parser, no OCR pipeline, no regex hell.

What to keep out of this node

• Scoring logic (separate node, separate prompt)
• Enrichment fields like tenure or seniority (comes from Apollo/Clearbit)
• Any writing back to Airtable (never mix extraction with side effects)

Step 3–4: Dedupe against your ATS, then enrich

Before you spend a cent on enrichment or send anyone an email, check if the candidate already exists. An Airtable "Search records" node (or your ATS's REST endpoint) against email and linkedin_url covers 95% of cases. If there's a hit, branch to a soft path:

IF existing_record:
  → Slack DM to referrer:
    "Thanks for sending Sarah — she's already in our pipeline
     for the backend role. We'll keep you posted."
  → END

That single auto-DM saves a recruiter roughly 30 minutes a week of awkward "hey did you get my referral" back-and-forth, and it saves the referrer from feeling ignored.

If it's a new candidate, hit an enrichment API. Apollo, Clearbit, People Data Labs — pick one you already pay for. Pass candidate_name + current_company and pull back LinkedIn URL, current title, tenure at company, location, and (if available) work email. You now have a full profile from a two-line Slack message and a PDF.

Step 5: The scoring node that decides everything

This is the node that separates a toy workflow from a real one. You pass the enriched profile plus the current job description into an LLM node with an explicit rubric. Not vibes. A rubric with weights, returning a score and evidence.

You are scoring a candidate against a job description.
Return JSON:
{
  "score": 0-100,
  "breakdown": {
    "must_have_skills": 0-40,
    "nice_to_haves": 0-20,
    "seniority_match": 0-20,
    "location_match": 0-20
  },
  "justification": "2 sentences citing specific evidence
                    from the profile. No hedging."
}

Rules:
- If a must-have skill is absent, cap must_have_skills at 15.
- If seniority is off by more than one level, cap seniority_match at 10.
- Never invent experience not present in the input.

The justification field is what makes this trustworthy. When the hiring manager sees "Score 84 — 6 years Go at fintech scale, EU timezone match, missing formal team-lead title but ran a 4-person squad at Klarna," they can act on it. When they see a raw "84/100" they can't.

Branch on threshold. Above 70, right branch (act). Below 70, left branch (polite hold + referrer feedback). Pick your own threshold after 20–30 real referrals — I've seen clients land anywhere from 62 to 78.

Right branch — high score

• Create the ATS row with all enriched fields, score, and justification
• Send the candidate an email from the hiring manager's address (Gmail node with delegated send), not noreply@
• Include a Calendly link and one specific sentence about why they were flagged
• Post to the hiring manager's private Slack: "New referral, score 84, contacted, calendar sent"

Left branch — low score

• Send a polite hold email to the candidate ("we'll keep your profile on file")
• DM the referrer with a specific reason: "Strong profile but the role needs 5 years of Go and she's coming from Python. We'll keep her warm for future roles."

That second DM is the entire reason referrals keep coming. Most teams stopped referring because they sent three people and heard nothing. Reply thoughtfully in under a minute, every time, and the channel wakes up.

Step 6: Close the loop with a monthly referrer summary

Every referral gets logged against the referrer's Slack user ID in an Airtable table: date, candidate name, score, outcome, bonus status. Once a month, an n8n cron node loops through the table grouped by referrer and sends each one a summary DM:

Hey Marko — your Q3 referrals:
  • Ana Petrovic  — final round (data eng)
  • Luka M.       — passed, kept warm
  • Jelena K.     — hired ✅ ($2,000 bonus pending payroll)

Total this quarter: 3 referrals, 1 hire, 1 in pipeline.

Two things happen when you ship this loop. First, referrers see their impact quantified and refer more. Second, finance gets a clean source of truth for referral bonuses, which was probably a spreadsheet that nobody trusted. In the client build I mentioned, referral volume roughly doubled in month two — not because we asked people to refer, but because they stopped feeling like they were shouting into a void.

What breaks in production and how to handle it

Three failure modes to plan for before you turn this on for a real channel:

• PDF parsing garbage. Some resumes are image-only scans. Add a text-length check after PDF extraction; if under 200 chars, route to a manual review Slack channel instead of failing silently.
• Referrer confusion. People post questions in #referrals too ("anyone know a good ATS?"). Add a preliminary classifier prompt: "Is this message a candidate referral? Return true/false." Cheap, prevents noise.
• Enrichment misses. Apollo won't find every candidate, especially early-career. If enrichment returns null, don't skip scoring — score on what you have and flag the confidence lower in the justification.

Cost per referral on the live setup: about $0.04 in LLM calls (parse + score), plus your enrichment credit — usually $0.10–$0.30 depending on vendor. Under 50 cents to route a candidate that a recruiter would spend 15 minutes on.

Where bizflowai.io fits in

We build these Slack-to-ATS routing agents for clients as part of the operations layer — not a chatbot, not a sourcing tool, but the plumbing that makes the tools you already pay for talk to each other. If you already run n8n and want to copy the pattern above, everything in this post is enough. If you'd rather have it built, scored against your actual job descriptions, and wired into your ATS, that's the shape of work we ship on bizflowai.io.

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.