AI Lead Desk Build Kit (Make/Zapier/n8n)

You’re wiring an enrich → triage → score → route loop that protects your calendar without adding busywork.

High‑level flow:

• Capture: Site form (or inbox/DM) → webhook.
• Enrich: Firmographics/contacts (Clearbit/Dropcontact or your pick).
• Triage: Tiny LLM prompt returns pass | near_miss | fail + reasons.
• Score: Deterministic pass/fail via your ICP JSON thresholds.
• Route:
  • PASS → show Calendly (or send booking link) and assign owner.
  • NEAR‑MISS → add to value lane (email sequence + useful asset).
  • FAIL → graceful “not a fit” reply with alternatives.
• Log: Emit a run‑log event per step with failure classes.

Wiring blueprints:

• Make (scenario): Webhooks → HTTP (Clearbit/Dropcontact) → AI (OpenAI/HTTP) → Tools > JSON → Router (3 paths) → HTTP (Calendly/email/CRM) → Data store (run‑logs).
• Zapier (Zap): Webhooks by Zapier → Webhooks (Clearbit/Dropcontact) → OpenAI (or Webhooks) → Filter + Paths → Gmail/Outlook + CRM → Webhooks (run‑log).
• n8n (workflow): HTTP Trigger → HTTP Request (enrichment) → OpenAI/HTTP → Function (score JSON) → Switch (paths) → HTTP Request (Calendly/email/CRM) → HTTP Request (log).

Use this as your single source of truth for qualification. Keep it short, explicit, and testable.

{
  "version": "1.0.0",
  "hard_disqualifiers": [
    { "field": "email_domain", "operator": "in", "value": ["gmail.com", "yahoo.com", "outlook.com"], "reason": "Free email domain" },
    { "field": "country", "operator": "not_in", "value": ["US", "CA", "UK", "EU"], "reason": "Out of service region" },
    { "field": "industry", "operator": "in", "value": ["B2C e-commerce dropshipping", "crypto", "adult"], "reason": "Outside target industries" }
  ],
  "weights": {
    "employee_count": 0.25,
    "industry_fit": 0.20,
    "title_seniority": 0.20,
    "tech_stack_fit": 0.15,
    "geo_fit": 0.10,
    "inbound_intent": 0.10
  },
  "scoring_rules": [
    { "field": "employee_count", "operator": "between", "value": [10, 500], "score": 1 },
    { "field": "industry", "operator": "in", "value": ["B2B SaaS", "Professional Services", "Healthcare B2B"], "score": 1 },
    { "field": "title_seniority", "operator": "in", "value": ["Founder", "Owner", "VP", "Head"], "score": 1 },
    { "field": "tech_stack", "operator": "contains_any", "value": ["HubSpot", "Salesforce", "Webflow", "Gmail", "Office365"], "score": 1 },
    { "field": "country", "operator": "in", "value": ["US", "CA", "UK", "DE", "FR", "AU"], "score": 1 },
    { "field": "intent", "operator": "in", "value": ["Implementation now", "Book an audit"], "score": 1 }
  ],
  "thresholds": {
    "auto_book_min_score": 0.70,
    "near_miss_floor": 0.50
  },
  "owner_lookup": {
    "precedence": ["domain", "company_name", "crm_account_id"],
    "fallback_owner_id": "OWNER_DEFAULT"
  },
  "notes": "Tune thresholds monthly. Treat hard disqualifiers as final unless overridden via manual review queue."
}

Implementation tips:

• Keep 3–4 scoring signals you trust; prune anything you can’t enrich reliably.
• Near‑miss should be generous (keeps optionality via nurture).

This prompt is deliberately short. Require JSON only. Budget tokens so it’s fast and cheap.

Recommended caps (practitioner defaults):

• Input cap: 900 characters of lead context (≈ 225 tokens).
• Output cap: 120 tokens.
• Temperature: 0.0–0.2 for consistency.

Prompt:

You are a lead triage classifier. Using only the provided lead context and ICP rules, decide PASS, NEAR_MISS, or FAIL and explain briefly.

Return JSON with keys: decision ("pass"|"near_miss"|"fail"), reasons (array), confidence (0–1), must_followups (array of 0–2 short questions if NEAR_MISS else []).

Lead context (<=900 chars): {{{LEAD_CONTEXT}}}
ICP summary: {{{ICP_SUMMARY}}}

Confidence rubric:

• 0.80–1.00: All critical signals present and align with ICP; reasons cite concrete fields.
• 0.50–0.79: Some ambiguity or missing enrichment; at least one reason references a gap.
• <0.50: Conflicting or insufficient signals.

Retry/backoff on 429s:

• Retries: up to 3 with random exponential backoff (base 1s, cap 20s, full jitter).
• On repeated 429: reduce max_tokens by ~20% and drop non‑essential context.

Output example:

{
  "decision": "near_miss",
  "reasons": ["employee_count below floor (7)", "title=Founder matches ICP"],
  "confidence": 0.62,
  "must_followups": ["Timeline to start?", "Budget owner?"]
}

Goal: show booking only to qualified visitors while assigning the right owner.

Routing Form (keep to 3–4 fields):

• Work email (required)
• Company website or domain (required)
• Team size (select: 1–9, 10–49, 50–249, 250–500, 500+)
• What do you need help with? (select: Audit, Implementation, Support)

Rules (examples — mirror your ICP JSON):

• If team size ≥ 10 AND need = Implementation → show event type “Discovery – Qualified (30m)”.
• Else if team size 1–9 OR need = Support → show message + link to resource and “Join waitlist”.

CRM owner lookup:

• Enable Salesforce or HubSpot lookup to assign based on domain; fallback to default owner if no match.

Embed (example — styling optional):

<div id="calendly-inline-widget" style="min-width:320px;height:780px" 
  data-url="https://calendly.com/yourteam/routing-form?hide_event_type_details=1&background_color=0D1321&text_color=ffffff&primary_color=2FD4EA">
</div>
<script type="text/javascript" src="https://assets.calendly.com/assets/external/widget.js" async></script>

Friction control:

• Cap total visible fields at 3–4; use enrichment to infer the rest.
• Pre‑fill email/domain from URL params where possible.
• Measure: form starts → qualified shows → bookings; watch false negatives via the run‑log.

Two short sequences you can paste today.

1. No‑fit (1 email)

• Subject: We’re not the best fit — here’s something better
• Body:

Hi [FIRST_NAME] — thanks for reaching out.
Based on what you shared, we’re not the best fit for [NEED/USE CASE]. Rather than waste your time, here are resources that help right away:
• [LINK_1: Practical guide]
• [LINK_2: Template or checklist]
If your situation changes (team ≥ [MIN_EMPLOYEES] and [TECH_STACK] in place), reply to this email and we’ll take another look.
— [SENDER]

2. Near‑miss (2 emails over 7 days)

• Email 1 (immediate)
  • Subject: Quick win now + 15‑min audit when ready
  • Body:

Hi [FIRST_NAME], good news: you’re close. Two quick wins to make you a fit fast:
1) Implement [TINY STEP] — tutorial: [LINK]
2) Add [MISSING STACK ITEM] — checklist: [LINK]
When that’s in place, use this link to book: [QUALIFIED_CALENDLY_URL].
— [SENDER]

• Email 2 (day 7)
  • Subject: Still want the audit?
  • Body: Short reminder + the same booking link.

Assets to prep once:

• 1-page “Starter Stack” PDF or Notion page.
• 5–7 minute loom showing the qualified path.
• Public checklist link you can safely share.

Log every meaningful step so you can debug and tune thresholds. Emit one event per step to a single webhook (or data store).

Event schema (suggested):

{
  "run_id": "uuid-v4",
  "lead_id": "uuid-or-email-hash",
  "source": "webform|inbox|dm|ads",
  "stack": "make|zapier|n8n",
  "timestamp": "2026-06-12T19:20:30Z",
  "event": "enrichment_success|enrichment_miss|llm_triage|score_computed|route_auto_book|route_near_miss|route_disqualified|email_sent|rate_limit_hit|retry_scheduled|error",
  "payload": { "redacted_fields": ["email"], "summary": "..." },
  "attempt": 1,
  "duration_ms": 1420,
  "http_status": 200,
  "provider": "clearbit|dropcontact|openai|gmail|graph|calendly",
  "cost_usd_estimate": 0.09,
  "failure_class": null,
  "trace": { "request_id": "...", "retry_after_s": null }
}

Failure classes (standardize these):

• ENRICH.NO_MATCH — no firmographic hit.
• ENRICH.TIMEOUT — enrichment SLA exceeded.
• LLM.RATE_LIMIT — 429 during triage.
• LLM.BAD_JSON — non‑parseable output.
• SCORE.MISSING_FIELD — required key absent.
• ROUTE.CALENDLY_UNAVAILABLE — embed/API error.
• EMAIL.SEND_FAILED — provider error.

Storage:

• Lightweight: Post to a serverless endpoint → append to a log table (SQLite/Postgres/Airtable/Notion DB).
• Heavyweight: Stream to BigQuery or Snowflake if you’re volume‑heavy.

Tip: Make logs idempotent via run_id + event + attempt.

Use exponential backoff with jitter and honor Retry-After headers. Keep retries idempotent.

Node/JS (axios):

async function withBackoff(fn, {max=3, base=1000, cap=20000}={}) {
  for (let i=0; i<=max; i++) {
    try { return await fn(); } 
    catch (e) {
      const status = e?.response?.status;
      const retryAfter = Number(e?.response?.headers['retry-after']);
      if (i===max || !(status===429 || status===503)) throw e;
      const sleep = retryAfter ? retryAfter*1000 : Math.min(cap, (base * 2**i) * (0.5 + Math.random()));
      await new Promise(r => setTimeout(r, sleep));
    }
  }
}

Python (requests):

import time, random, requests

def with_backoff(call, max_retries=3, base=1.0, cap=20.0):
    for i in range(max_retries+1):
        try:
            return call()
        except requests.HTTPError as e:
            status = e.response.status_code if e.response else None
            retry_after = e.response.headers.get('Retry-After') if e.response else None
            if i==max_retries or status not in (429, 503):
                raise
            sleep = float(retry_after) if retry_after else min(cap, (base*(2**i)) * (0.5 + random.random()))
            time.sleep(sleep)

Make/Zapier patterns:

• Keep an attempt counter in a data store.
• On 429/503, Delay (randomized) → re‑queue the same step.
• Use idempotency keys when calling mail/cal providers.

Provider notes:

• Gmail API and Microsoft Graph both return 429/503 under load — back off and retry with jitter.
• For calendars/booking, expect occasional 429s; avoid fan‑out bursts (queue requests).

Start with a small, consistent dataset and a simple cost sheet. Replace bracketed values with your numbers.

Sample CSV (3 rows):

email,domain,company,employee_count,industry,country,intent
alex@acme.io,acme.io,Acme,42,B2B SaaS,US,Implementation now
sara@solo.dev,solo.dev,Solo Dev,1,Freelance,UK,General inquiry
lee@contoso.com,contoso.com,Contoso,320,Professional Services,CA,Book an audit

Ops cost estimator (Notion/Sheets columns):

• Enrichment provider, Enriched? (Y/N), Cost/lead
• LLM prompt tokens, LLM completion tokens, Model $/1K tokens, LLM cost/lead
• Sends (email/calendar API ops), $/op, Ops cost/lead
• Total cost/lead

Formulas:

• Enrichment (Clearbit Batch): =IF([Enriched?]="Y", 0.30, 0)
• Enrichment (Dropcontact): =[CREDITS_USED] * [CREDIT_PRICE_USD] (Pay‑On‑Success — use your plan’s credit value)
• LLM: =((PromptTokens + CompletionTokens)/1000) * ModelPricePer1k
• Total: =Enrichment + LLM + Ops

Guardrails to track weekly:

• % Qualified show rate (qualified shows ÷ form starts)
• Booking rate (bookings ÷ qualified shows)
• False negative rate (manual approvals ÷ near‑miss + fail)
• Median latency (capture → booking shown)

Tuning loop:

• If false negatives > 5%, lower auto_book_min_score by 0.05 and/or relax one scoring rule.
• If calendar quality drops, tighten 1 high‑signal rule (e.g., title_seniority).