Automate Trial-to-Paid Lead Scoring for SaaS with n8n + Claude Step-by-Step Guide

2. System Architecture

Six components, each replaceable. The orchestration layer is self-hosted n8n so security review can audit every external API call that touches prospect data. Postgres holds both the event store and the score history.

The stack

Cost estimate (5,000 signups/month)

At an ACV of $9k–$30k, a single incremental closed deal from this pipeline pays for the entire stack for a year. The same orchestration layer slots into our broader AI automation services.

The pipeline needs every signup, regardless of which surface it came in through. In a typical SaaS that means three sources: the product itself (your auth provider’s webhook), your billing provider when a card is added (Stripe), and your marketing forms (HubSpot Forms, Default.com, or a custom landing page).

Webhook sources

• Auth0 / Cognito / Clerk: emits a user.created event with email, signup IP, optional company name from the form.
• Stripe: customer.created for self-serve trial-with-card flows. Fires roughly 2-5 minutes after auth signup.
• App POST: for product-qualified events the auth webhook can’t capture (workspace_created, integration_connected). Use a signed HMAC payload.
• Form fills: demo requests, content downloads, pricing-page submissions. Treat these as first-class signups.

n8n webhook node config

A single n8n Webhook node accepts the union schema; a Switch node downstream branches on source. HMAC verification happens before the workflow does anything else.

{
  "source": "auth0" | "stripe" | "app" | "form",
  "event_id": "evt_01HXYZ...",
  "received_at": "2026-05-03T09:14:22Z",
  "signature": "sha256=...",
  "user": {
    "email": "[email protected]",
    "first_name": "Alex",
    "last_name": "Chen",
    "company_name_provided": "Acme Corp",
    "ip": "203.0.113.42",
    "user_agent": "Mozilla/5.0 ...",
    "utm": {
      "source": "google",
      "campaign": "pql-q2-2026",
      "term": "workflow automation"
    }
  },
  "workspace_id": "ws_92ad3c"
}

Raw signup data is almost useless for scoring. You need company size, industry, tech stack, role seniority, location and recent funding signals. Run enrichment in parallel with a short timeout (4 seconds) so a slow vendor doesn’t block the whole pipeline. The same enrichment pattern shows up in our work for AI automations for digital marketing agencies.

Enrichment cascade (fail fast, fail cheap)

1. Domain MX check — if the email domain has no MX record, mark invalid and stop.
2. Free-email filter — gmail.com, yahoo.com, qq.com etc. flag consumer_email. Don’t enrich (waste of credits) but still score.
3. Clearbit Reveal on the signup IP — gives company name + domain even if they signed up with a personal email.
4. Clearbit Person + Company on the work email — title, seniority, company size, industry, funding, tech stack.
5. Apollo as fallback when Clearbit returns <30% field coverage.
6. Bright Data LinkedIn scraper as last resort for missing seniority on small companies.

SQL: enrichment cache to avoid double billing

-- Reuse company-level enrichment for any user from the same domain
-- within 30 days. Saves ~40% of Clearbit credits at scale.

WITH cached AS (
  SELECT enrichment_payload
  FROM   enrichment_cache
  WHERE  domain = $1
    AND  enriched_at > now() - interval '30 days'
  ORDER  BY enriched_at DESC
  LIMIT  1
)
SELECT
  CASE
    WHEN EXISTS (SELECT 1 FROM cached)
    THEN (SELECT enrichment_payload FROM cached)
    ELSE NULL
  END AS payload;

A traditional rules-based scorer has to be hand-tuned every time you launch a new feature, target a new vertical, or learn a new buying signal. Claude reads the enriched profile plus the early product signals, scores against a written ICP, and returns a structured JSON with reasoning — which makes the score auditable.

The 5 scoring dimensions

ICP scoring system prompt

Tested on 6 months of historical signups labeled by closed-won vs churned. Tunable per company by editing the ICP block at the top.

You are an ICP scoring engine for a B2B SaaS workflow automation tool.

ICP definition (highest fit):
- Industry: SaaS, fintech, e-commerce ops, agencies
- Size: 50-2000 employees, >$5M ARR
- Tech stack: uses Slack, HubSpot/Salesforce, AWS or GCP
- Role: ops, RevOps, marketing ops, growth eng (IC level OK
  if they're the documented buyer for tooling)

DISQUALIFY signals:
- Personal Gmail/Yahoo with no company match
- Competitor employee (check enriched company against our
  competitor list in context)
- Student / agency-doing-research signals
- <5 employees (free tier suffices)

OUTPUT — strict JSON, no prose:
{
  "score": 0-100,
  "tier": "hot" | "warm" | "cold" | "disqualify",
  "dimensions": {
    "industry_match": 0-20,
    "company_size_band": 0-20,
    "tech_stack_signals": 0-20,
    "role_seniority": 0-20,
    "intent_signals": 0-20
  },
  "reasoning": "2-3 sentences citing the strongest signals",
  "recommended_action": "ae_dm" | "sdr_queue" | "drip_only" | "discard"
}

Rules:
- score = sum of dimensions
- tier mapping: >=80 hot, 50-79 warm, <50 cold
- if disqualify_reason present, tier = "disqualify" regardless
- never invent enrichment fields not in the input

n8n HTTP Request to Claude

{
  "method": "POST",
  "url": "https://api.anthropic.com/v1/messages",
  "headers": {
    "x-api-key": "{{ $credentials.anthropic.apiKey }}",
    "anthropic-version": "2023-06-01",
    "content-type": "application/json"
  },
  "body": {
    "model": "claude-sonnet-4-5",
    "max_tokens": 600,
    "system": "{{ $json.icpSystemPrompt }}",
    "messages": [
      {
        "role": "user",
        "content": "Enriched signup:n{{ JSON.stringify($json.enriched) }}nnEarly product events (last 1h):n{{ JSON.stringify($json.events) }}"
      }
    ]
  }
}

A trial user who hits first-value, invites a teammate, and connects an integration in the first 48 hours converts at roughly 7x the rate of a passive signup. Capture those events into the same Postgres event store and feed them back into Claude on a 6-hour and 24-hour re-score cadence.

The activation events that move scores

• first_value_reached — user completed the core “aha” action (varies per product).
• team_invited — strongest single signal of buying intent. Multi-seat = paid.
• integration_connected — they brought their data in. Switching cost just went up.
• pricing_page_visited — late-trial intent. Re-score within 30 minutes.
• api_token_created — engineering champion is real, evaluation is serious.

Event store schema

CREATE TABLE product_events (
  id            uuid PRIMARY KEY DEFAULT gen_random_uuid(),
  workspace_id  text NOT NULL,
  user_email    text NOT NULL,
  event_name    text NOT NULL,
  event_value   jsonb,
  occurred_at   timestamptz NOT NULL,
  ingested_at   timestamptz DEFAULT now()
) PARTITION BY RANGE (occurred_at);

CREATE INDEX idx_events_workspace_event
  ON product_events (workspace_id, event_name, occurred_at DESC);

-- Activation feature view used by the re-score job
CREATE MATERIALIZED VIEW activation_features AS
SELECT
  workspace_id,
  user_email,
  bool_or(event_name = 'first_value_reached')    AS hit_first_value,
  bool_or(event_name = 'team_invited')           AS invited_team,
  bool_or(event_name = 'integration_connected')  AS connected_integration,
  count(*) FILTER (WHERE event_name = 'pricing_page_visited') AS pricing_views,
  max(occurred_at)                                AS last_active_at
FROM product_events
WHERE occurred_at > now() - interval '14 days'
GROUP BY workspace_id, user_email;

Once a score lands, routing is deterministic. Hot tier needs a human in front of the prospect inside 5 minutes; warm tier feeds the SDR queue with a personalized opener Claude already drafted; cold tier never touches a human until they re-score upward.

Routing matrix

Slack alert payload (hot tier)

{
  "channel": "#hot-trials",
  "blocks": [
    {
      "type": "header",
      "text": { "type": "plain_text",
                "text": ":fire: Hot trial — score 87" }
    },
    {
      "type": "section",
      "fields": [
        { "type": "mrkdwn",
          "text": "*Alex Chen*nDirector RevOps, Acme Corp" },
        { "type": "mrkdwn",
          "text": "*Stack*nSlack, Salesforce, Snowflake" }
      ]
    },
    {
      "type": "section",
      "text": { "type": "mrkdwn",
                "text": "*Why hot:* came from /integrations/salesforce, invited 2 teammates, connected Slack within 11 min of signup." }
    },
    {
      "type": "actions",
      "elements": [
        { "type": "button", "text": { "type": "plain_text",
          "text": "Claim" }, "style": "primary", "action_id": "claim_lead" },
        { "type": "button", "text": { "type": "plain_text",
          "text": "Open in HubSpot" }, "url": "https://app.hubspot.com/contacts/.../{{contactId}}" }
      ]
    }
  ]
}

The Claim button hits a webhook back into n8n that locks the lead to that AE for 24 hours and drops them out of the SDR round-robin. This is the pattern we re-use for our consulting industry automations where partners need first-touch on enterprise leads.

HubSpot or Salesforce remains the source of truth for everything sales sees. The n8n workflow upserts the contact, attaches it to a company record (matching on domain), updates lifecycle stage, and pushes the score plus reasoning into custom fields. Dedupe runs on email + domain to handle the multi-stakeholder evaluation case cleanly.

HubSpot upsert (GraphQL via the v3 API)

{
  "method": "PATCH",
  "url": "https://api.hubapi.com/crm/v3/objects/contacts/[email protected]?idProperty=email",
  "headers": {
    "authorization": "Bearer {{ $credentials.hubspot.token }}",
    "content-type": "application/json"
  },
  "body": {
    "properties": {
      "firstname": "Alex",
      "lastname":  "Chen",
      "jobtitle":  "Director, RevOps",
      "lifecyclestage": "salesqualifiedlead",
      "skru_icp_score":      "87",
      "skru_icp_tier":       "hot",
      "skru_icp_reasoning":  "{{ $json.score.reasoning }}",
      "skru_signup_source":  "auth0",
      "skru_first_event":    "team_invited"
    }
  }
}

Dedupe SQL

-- Idempotent contact write keyed on lower(email).
-- Aggregates same-domain stakeholders into one company score.

INSERT INTO contacts (email, domain, first_seen_at, latest_score, latest_tier)
VALUES ($1, $2, now(), $3, $4)
ON CONFLICT (email) DO UPDATE
  SET latest_score = EXCLUDED.latest_score,
      latest_tier  = EXCLUDED.latest_tier,
      updated_at   = now();

-- Promote the company tier to the highest stakeholder tier in the last 30d
UPDATE companies c
SET    rolled_up_tier = sub.max_tier
FROM (
  SELECT domain,
         CASE WHEN max(score_rank) = 4 THEN 'hot'
              WHEN max(score_rank) = 3 THEN 'warm'
              ELSE 'cold' END AS max_tier
  FROM (
    SELECT domain,
           CASE latest_tier WHEN 'hot' THEN 4
                            WHEN 'warm' THEN 3
                            WHEN 'cold' THEN 2
                            ELSE 1 END AS score_rank
    FROM contacts
    WHERE updated_at > now() - interval '30 days'
  ) ranked
  GROUP BY domain
) sub
WHERE c.domain = sub.domain;

Common Failures & Fixes

Three failure modes show up in nearly every deployment. Plan for them on day one — they are cheap to design around and expensive to retrofit.

Failure 1: Score inflation on free-email ICPs

Symptom: A consultant with a Gmail address claims VP-level role on the form, gets enriched against the wrong company (their last employer’s domain on LinkedIn) and lands as hot.

Fix: Cap any consumer-email signup at score 60 in the prompt rules and require a manual SDR confirm before promoting to hot. Document the rule in the system prompt so the reasoning string explains why.

Failure 2: Competitor signups get enriched and routed

Symptom: A PM at a direct competitor signs up, gets routed to an AE who books a discovery call before realizing.

Fix: Maintain a Postgres table competitor_domains and inject it into the Claude prompt context. Auto-classify as disqualify with a separate notification to product marketing instead of sales.

Failure 3: Rate-limit cascades during signup spikes

Symptom: Product Hunt launch sends 3,000 signups in 90 minutes. Clearbit rate-limits, n8n queue backs up, hot leads land in Slack 4 hours late.

Fix: Run n8n in queue mode with at least 4 worker processes. Wrap external calls in retry-with-jitter (3 attempts, 250ms-2s). Pre-warm the enrichment cache from your last 30 days of signup domains before any planned campaign.

Privacy: GDPR, CCPA, CASL & SOC 2

B2B prospect data still falls under GDPR (EU), CCPA (California), and CASL (Canada). PCI is not relevant because no payment data flows through the scoring pipeline — Stripe handles that on its own surface. Treat compliance as a design constraint, not a checklist at the end.

What Claude sees (and doesn’t)

• Sees: business email, name, job title, employer, public LinkedIn-style firmographics, product event names + counts, UTM source.
• Doesn’t see: message text inside the product, file contents, payment data, IP address (only the country derived from it), session tokens.

GDPR specifics

Lawful basis for B2B prospecting in the EU is legitimate interest, but it requires a balancing test on file and instant honor of objection requests. Implement DSR endpoints (access, erasure, portability) that purge the contact across Postgres, the enrichment cache, and HubSpot in one transaction. CASL in Canada is stricter — you need express consent before sending the warm/cold sequences to Canadian addresses; flag and gate by enriched country.

SOC 2 Type 2 controls this satisfies

• Audit trail: every score, prompt hash, and CRM write is logged in the audit table with reviewer (if any), timestamp, and source IP.
• Change management: n8n workflows live in git with a CI deploy, no point-and-click prod edits.
• Access controls: role-based n8n credentials, vault-backed secrets, no shared API keys per human.

Measured Results — 90 Days In

Numbers from a real implementation at a Series A vertical SaaS (1,800 signups/month, 3-person SDR team, 4 AEs) after the first full quarter on the new pipeline. No change in trial volume during the test period — the lift comes entirely from prioritization and speed.

The headline metric inside the SDR team is precision on the hot tier: 91% of leads scored hot became opportunities. That’s the number that lets sales leadership trust the score enough to actually change behavior. Without it, AEs revert to working their old lists by gut.

Implementation Timeline & Cost

FAQ