AI Lead Scoring Agent: n8n + Claude Sonnet Setup (2026)

AI Lead Scoring Agent: n8n + Claude Sonnet Setup (2026)

The n8n + Claude 3.5 Sonnet lead scoring agent captures new leads from web forms, enriches them with Apollo.io firmographic data, evaluates each lead against a 4-criteria ICP rubric, and writes the score plus rationale back to HubSpot — all in under 10 seconds per lead. An SDR who previously spent 10-12 hours per week on manual enrichment recovers 8 of those hours in the first week. Setup takes 90 minutes.

THE REAL PROBLEM

An SDR at a 50-person B2B SaaS company generates 40-60 inbound leads per week from form fills and content downloads. Every lead needs enrichment: company size, industry, funding stage, tech stack. This means opening Apollo.io, searching each company, copying 5-6 data points into HubSpot, and making a subjective judgment: hot, warm, or cold? At 10 minutes per lead, that is 7-10 hours of copy-paste work per week. Not selling. Not strategizing. Just data entry.

[ STAT ] Sales reps spend only 28% of their week actually selling. The remaining 72% disappears into admin work, data entry, and manual research. — Salesforce State of Sales, 2026

At a fully loaded cost of $60K/year per SDR, that 72% represents $43,200/year per person in non-selling activity. For a team of 5 SDRs, that is $216,000/year spent on data entry and research that an automated workflow can do in 10 seconds per lead.

Rule-based tools like Zapier fail here because they cannot evaluate nuance. An IF node can check if headcount > 50, but it cannot assess whether a VP of Engineering at a Series A fintech is a better fit than a Director of IT at an enterprise. Claude evaluates the full context.

WHAT THIS WORKFLOW ACTUALLY DOES

The workflow transforms a raw form submission into a scored, routed, and documented HubSpot lead in under 10 seconds end-to-end. Here is what each tool contributes:

[TOOL: n8n] Orchestrator connecting 5 API calls across 4 services. Handles webhook ingestion, data transformation, conditional routing (score >= 80 triggers Slack), and error handling with retry logic. Avg execution time per lead: 8 seconds.

[TOOL: Apollo.io API] B2B data enrichment source. Accepts an email or company name, returns industry, headcount range, revenue range, funding stage, tech stack tags, and LinkedIn URL from a database of 275M+ contacts and 73M+ companies. Avg response time: 1-3 seconds per enrichment call.

[TOOL: Claude 3.5 Sonnet] The ICP scoring engine. Receives enriched firmographic data, evaluates against a weighted rubric — industry alignment (40 pts), company size fit (25 pts), job title seniority (20 pts), tech stack relevance (15 pts) — and returns a structured JSON with total score, grade, and per-criterion rationale. Avg inference time: 2-4 seconds.

[TOOL: HubSpot CRM] Record destination. Receives enriched data and Claude's scoring output written to custom contact properties: icp_total_score, icp_grade, icp_rationale, enrichment_status.

The agentic reasoning step that no scripted automation can replicate: Claude evaluates the combination of signals rather than checking individual conditions. A lead from a non-target industry with a perfect title score might still score 65 if the tech stack shows they use a complementary tool — a signal pattern a rule-based system would miss entirely.

WHO THIS IS BUILT FOR

FOR SDRs at B2B SaaS companies (10-200 employees) SITUATION: You spend 2+ hours daily on lead research. Your CRM has 500+ leads with missing enrichment data. You guess which leads to call first. PAYOFF: Every lead arrives pre-enriched and scored with auditable rationale. You review 20 scored leads in 10 minutes instead of 3 hours.

FOR sales ops managers at growth-stage B2B companies SITUATION: Your 5 SDRs each score leads differently. Lead quality is inconsistent. You cannot prove which SDR's scoring is accurate. PAYOFF: A single Claude ICP rubric scores every lead uniformly. Weekly reports show score distribution and conversion by tier.

FOR RevOps leaders at Series A-C SaaS companies SITUATION: You manage leads from web forms, events, chatbots, and content downloads — each with different data quality. PAYOFF: A single workflow handles all sources. HubSpot custom properties provide full audit trail: who scored it, when, and why.

HOW IT RUNS: STEP BY STEP

1. Lead Capture (n8n Webhook — instant) Input: HubSpot form POST to n8n webhook URL. Fields: email, name, company, title, phone. Action: n8n normalizes the payload: lowercases email, formats phone to E.164, strips HTML from text fields. Output: validated lead object ready for enrichment.

2. Apollo.io Enrichment (n8n HTTP Request Node — 2-5 seconds) Input: lead email + company name sent via POST to https://api.apollo.io/api/v1/people/match. Header: X-Api-Key. Action: Apollo searches its 275M+ contact database for the best matching profile. Returns firmographic data: industry, headcount, revenue band, funding stage, tech stack, LinkedIn URL. Output: enriched lead object with person + company data merged.

3. ICP Scoring via Claude 3.5 Sonnet (n8n HTTP Request to Anthropic — 3-5 seconds) Input: enriched lead data injected into Claude system prompt with the 4-criteria ICP rubric. Action: Claude evaluates: industry alignment (40 pts) — is this a target vertical? company size (25 pts) — is headcount in range? title seniority (20 pts) — decision-maker or influencer? tech stack (15 pts) — complementary or competing tools? Output: JSON — { total_score, grade, industry_score, size_score, title_score, tech_score, rationale }

4. HubSpot Update (n8n HubSpot Node — 1-2 seconds) Input: original contact ID + enriched data + Claude scoring JSON. Action: n8n sends PATCH to HubSpot CRM API writing to custom properties: icp_total_score, icp_grade, icp_rationale, enrichment_status, enrichment_timestamp. Output: updated HubSpot contact with full enrichment and scoring audit trail.

5. Slack Alert (n8n Slack Node — < 1 second) Input: leads where total_score >= 80 (checked by IF node). Action: Slack webhook posts formatted message to #hot-leads: lead name, company, score, grade, top rationale, HubSpot link. Output: Slack message with actionable lead card.

6. Weekly Summary (n8n Schedule Trigger — cron, Sunday 8 AM) Input: HubSpot contacts with enrichment_timestamp in last 7 days. Action: n8n aggregates by grade, computes averages, formats report. Output: Slack or email summary: leads processed, score distribution, top industries, conversion by grade.

SETUP AND TOOLS

Total setup: approximately 90 minutes if all API access is already provisioned. Add 1-2 business days for HubSpot custom property creation approval if your org requires it.

n8n (n8n.io) — Workflow orchestrator (free self-hosted / $24/mo cloud) Apollo.io API — Contact and company data enrichment (free tier: 50 req/min, 600 req/day) Claude 3.5 Sonnet (Anthropic) — ICP scoring engine ($3/M input tokens) HubSpot CRM — Lead record storage (free CRM tier works) Slack — Hot lead alerts (free workspace)

Gotcha: Apollo's free tier limits you to 50 API requests per minute and 600 per day. At 10 seconds per lead, you process 60 leads per hour, well within the per-minute limit, but the daily limit of 600 means you cannot enrich more than 600 leads in a single day. If you anticipate spikes above 600 leads/day (e.g., after a conference), upgrade to the Professional plan (2,000 req/day) or add a rate-limiting queue in n8n that distributes enrichment across hours.

THE NUMBERS

The single most impactful number from deployment data: teams using AI lead scoring report a 3x improvement in lead-to-opportunity conversion vs manual qualification.

▸ Lead enrichment time 10-15 min/lead -> 8-10 sec/lead (SyncGTM AI Benchmarks, 2026) ▸ SDR research hours/week 8-12 hrs -> <1 hr (Apollo.io Insights, 2026) ▸ Lead response time 24-48 hrs -> 30-60 sec (HubSpot State of Marketing, 2026) ▸ API cost per lead N/A (manual) -> ~$0.004/lead (Anthropic pricing, 2026)

Source: SyncGTM, How Much Time Can AI Actually Save Sales Reps, 2026. A team of 5 SDRs recovering 8 hours each per week saves 40 hours/week total.

WHAT IT CANNOT DO

1. Apollo enrichment gaps in niche industries (moderate risk). Apollo's coverage is strongest for tech/SaaS companies. For manufacturing, healthcare, or government leads, match rates can drop below 50%. Mitigation: add Clearbit as a secondary enrichment source in the workflow for industries where Apollo returns low-confidence matches.

2. Stale data in enrichment results (moderate risk). Apollo's contact database may show a person's previous role or an outdated company headcount. If Claude scores based on stale data, the lead may be misclassified. Mitigation: add an enrichment_date field and flag leads where enrichment data is 30+ days old for SDR review.

3. Claude scoring bias toward verbose profiles (minor risk). Leads with richer firmographic data tend to score higher because Claude has more signals to evaluate. A well-funded SaaS startup with a detailed Apollo profile will score higher than an equally qualified bootstrapped company with sparse data. Mitigation: Claude's prompt includes: 'If firmographic data is sparse, score conservatively and note data gaps in the rationale.'

START IN 10 MINUTES

1. (3 min) Sign up at n8n.io and create a new workflow. Add a Webhook trigger node. Copy the webhook URL.

2. (3 min) Go to apollo.io > Settings > Integrations > API. Generate an API key. Add it as an HTTP Header Auth credential in n8n with header name X-Api-Key.

3. (2 min) Go to console.anthropic.com > API Keys. Generate a key. Add it as an HTTP Header Auth credential in n8n with header name x-api-key. Set model to claude-sonnet-4-20250514.

4. (2 min) Create custom HubSpot contact properties: icp_total_score (number), icp_grade (text), icp_rationale (textarea), enrichment_status (text). Test by clicking Execute Workflow with a sample lead payload.

FREQUENTLY ASKED QUESTIONS

Q: How much does the n8n + Apollo + Claude lead scoring workflow cost per month? A: Approximately $125-150/month for 500 leads. n8n is free self-hosted or $24/month cloud. Apollo Professional is $99/month (free tier works for testing). Claude API costs ~$2/month for 500 scoring calls at ~750 tokens per call. HubSpot free CRM tier covers the basics.

Q: Can I use GPT-4o instead of Claude for the ICP scoring step? A: Yes. Replace the Anthropic HTTP Request node with an OpenAI node and point it to gpt-4o. GPT-4o is slightly cheaper ($2.50/M input vs $3/M for Claude Sonnet) but produces longer rationale text, increasing output token costs. Both models support structured JSON output. Test both with your ICP rubric on 50 leads and compare scoring consistency.

Q: What happens when Apollo.io returns a 429 rate limit error? A: The workflow should catch the 429 status and retry with exponential backoff. On the free plan (50 req/min), space enrichment calls using n8n's Wait node with a 1.5-second delay between calls. For bursts above 50 req/min, upgrade to the Professional plan (200 req/min). Never retry more than 3 times — if Apollo is under load, retries compound the problem. (Source: Apollo.io Rate Limits documentation, 2026)

Q: How do I handle data privacy when sending lead data to Claude for scoring? A: Claude 3.5 Sonnet is SOC 2 Type II certified and Anthropic does not train on API inputs. However, if your compliance team requires zero external data transfer, replace Claude with a local LLM via Ollama running on the same server as n8n. Expect 3-5 seconds per scoring call instead of 2-4 seconds with the API.

Q: How long does it take to set up this workflow from scratch? A: 90 minutes if you have API keys ready. Breakout: 15 minutes for n8n workflow scaffolding and webhook setup, 20 minutes for Apollo enrichment node configuration, 15 minutes for the Claude scoring prompt with ICP rubric, 20 minutes for HubSpot custom property creation and CRM node wiring, 10 minutes for Slack webhook and alert formatting, 10 minutes for end-to-end testing with a sample lead.