Most outreach fails not because the offer is weak — it's because the process is either too manual to scale or too automated to feel human. The sweet spot is a pipeline that thinks before it sends. That's exactly what you get when you combine n8n's workflow orchestration with Claude's language reasoning.
This isn't a theoretical walkthrough. The architecture below is a close approximation of what we deploy at Workflow AI Advisors for clients running outbound across SaaS, professional services, and e-commerce verticals. We've used variations of this stack to eliminate 40+ hours per week of manual SDR work while keeping reply rates well above industry benchmarks.
Let's build it properly.
Why n8n + Claude Is the Right Stack for This
Before we get into the architecture, it's worth being clear about why these two tools specifically.
n8n is a self-hostable workflow automation platform with a visual node editor. Unlike Zapier or Make, it gives you genuine programmatic control — you can write JavaScript directly inside nodes, branch logic conditionally, loop through arrays, and connect to virtually any API. Crucially, you can self-host it on a VPS, which matters for data privacy when you're handling prospect data at scale.
Claude (by Anthropic) consistently outperforms GPT-4 on tasks requiring nuanced tone calibration, instruction-following, and long-context reasoning. For email personalisation, this matters enormously. Claude is less likely to hallucinate sycophantic fluff and more likely to produce copy that reads like a human wrote it after doing actual research.
Together, they give you a pipeline where n8n handles the data orchestration and Claude handles the thinking.
The Pipeline Architecture at a Glance
Here's the high-level flow before we drill into each component:
- Lead Ingestion — Prospects enter from a Google Sheet, Airtable, CRM webhook, or Apollo export
- Enrichment — n8n calls enrichment APIs (Clearbit, Hunter, or LinkedIn scrape via Apify) to fill gaps
- Research Synthesis — Claude reads enrichment data + any scraped company intel and produces a structured prospect summary
- Email Generation — Claude generates a personalised first-touch email based on the summary and a prompt template you control
- Quality Gate — A second Claude call scores the email against your criteria before it's approved for sending
- Send or Queue — Approved emails go to Instantly, Smartlead, or directly via SMTP; rejected ones flag for human review
- Reply Handling — Inbound replies trigger a new branch: classification, CRM logging, and optional AI-drafted response
This isn't a linear chain — it's a conditional graph. Each node can branch, fail gracefully, or escalate to a human. That's what makes it genuinely autonomous rather than just automated.
Step 1: Lead Ingestion and Deduplication
Start with a Google Sheets trigger node (or an Airtable trigger if that's your source of truth). Set it to poll every 15 minutes or trigger on row append via webhook.
The first thing your workflow should do is check for duplicates. Add a Postgres or Supabase node that queries your outreach log table:
SELECT id FROM outreach_log WHERE email = '{{ $json.email }}' LIMIT 1;If a row is returned, route to a No Operation node and stop. This single step prevents the embarrassing double-send problem that plagues simpler automation setups.
Step 2: Data Enrichment
Raw lead lists are almost always incomplete. Claude needs context to write a good email — you can't personalise off a first name and a company domain.
Set up a HTTP Request node calling your enrichment provider of choice. We typically use a combination:
- Hunter.io for email verification and basic company data
- Clearbit Reveal for firmographic data (headcount, industry, funding stage)
- Apify's LinkedIn scraper for the prospect's recent activity or job title history
Merge the enrichment responses back into the main data object using n8n's Merge node in "Merge By Key" mode. Your enriched object should now include: full name, verified email, job title, company name, industry, headcount, recent LinkedIn post or activity (if available), and any relevant news about their company from the past 90 days.
For company news, add a Serper.dev HTTP node querying site:techcrunch.com OR site:businesswire.com "{company_name}" and pull the top result's snippet. This gives Claude something real to reference.
Step 3: Research Synthesis with Claude
Now you make your first Claude call. This is a critical step that most "AI outreach" tutorials skip entirely — and it's why those tutorials produce generic emails.
Rather than sending raw enrichment data straight into an email prompt, you ask Claude to synthesise first.
In an HTTP Request node pointed at the Anthropic API (https://api.anthropic.com/v1/messages), set your system prompt to something like:
You are a senior B2B research analyst. Given the following data about a prospect, produce a structured JSON summary with these fields: pain_points (array of 2-3 likely business challenges based on their role and company stage), conversation_hook (one specific, non-generic observation about their company or role that could open a natural conversation), avoid_topics (anything that might make this outreach feel intrusive or poorly timed). Be precise. Do not invent facts not supported by the data.
Pass the enriched prospect object as the user message. Parse Claude's JSON response using n8n's Code node with JSON.parse($json.content[0].text).
This intermediate step is what separates a pipeline that produces 3% reply rates from one that produces 11%.
Step 4: Email Generation
Now make your second Claude call — this time to write the actual email.
Your system prompt should include your core outreach framework. We use a variant of the PPPP framework (Promise, Picture, Proof, Push) compressed into 4-6 sentences for cold email, but adapt this to your own proven structure.
The user message should inject:
{{prospect.first_name}}{{prospect.company}}{{synthesis.conversation_hook}}{{synthesis.pain_points}}- Your offer/value proposition (hardcoded in the prompt or pulled from a config table)
- A firm instruction: "Do not use generic phrases like 'I hope this email finds you well.' Do not mention AI. Write in first person. Maximum 120 words."
Word count constraints are non-negotiable. Without them, Claude will produce polished prose that's 30% too long for cold outreach.
Step 5: The Quality Gate
This is the step that makes the pipeline trustworthy enough to run autonomously.
Make a third Claude call — a lightweight one using claude-3-haiku to keep costs down — with a strict scoring prompt:
Score this cold email on a scale of 1-10 across three dimensions: Specificity (does it reference something real about this prospect?), Tone (does it sound human and peer-to-peer, not salesy?), Clarity (is the ask unambiguous?). Return JSON: {"specificity": int, "tone": int, "clarity": int, "pass": boolean, "rejection_reason": string or null}. Pass = true only if all three scores are 7 or above.
Route on the pass boolean. Passed emails move to the send queue. Failed emails write to a "human review" Airtable table with the rejection reason attached — so you can audit, improve your prompts, and reprocess.
In practice, a well-tuned prompt template sees 85-92% pass rates. Anything below 70% tells you the enrichment quality is poor or your prompt is conflicting with itself.
Step 6: Sending Infrastructure
Don't send directly from your primary domain. Use a warmed sending domain routed through Instantly or Smartlead. Both have APIs that n8n can call natively via HTTP Request nodes.
A basic Instantly API call looks like:
POST https://api.instantly.ai/api/v1/lead/add
{
"api_key": "YOUR_KEY",
"campaign_id": "YOUR_CAMPAIGN_ID",
"skip_if_in_workspace": true,
"leads": [{
"email": "{{prospect.email}}",
"first_name": "{{prospect.first_name}}",
"custom_variables": {
"email_body": "{{generated_email}}"
}
}]
}Set daily send limits per mailbox (we recommend no more than 30-40 per warmed inbox per day) and stagger sending times using n8n's Wait node with a randomised interval between 45 and 180 seconds.
Step 7: Reply Handling and CRM Sync
A pipeline that can't handle replies isn't autonomous — it's just a fancy mail merge.
Set up an IMAP Email trigger node or use Instantly/Smartlead's reply webhook to catch inbound responses. Feed the reply text into Claude with a classification prompt:
Classify this email reply as one of: INTERESTED, NOT_INTERESTED, OUT_OF_OFFICE, WRONG_PERSON, REFERRAL, or UNSUBSCRIBE. Return JSON with classification and a one-sentence reasoning.
Branch on classification:
- INTERESTED → Create deal in HubSpot or Pipedrive via their n8n nodes, notify the assigned SDR via Slack
- UNSUBSCRIBE → Add to suppression list immediately, update CRM
- OUT_OF_OFFICE → Extract return date using Claude, schedule a follow-up task for that date + 1 day
- INTERESTED / REFERRAL → Draft a response using Claude, send to SDR for one-click approval via a Slack button
This last point — human-in-the-loop approval for warm replies — is a deliberate design choice. Full autonomy is appropriate for prospecting. For live conversations with interested buyers, a human should stay in the loop.
Cost and Performance Benchmarks
Running this pipeline at scale, here's what you can expect:
- API cost per lead processed: $0.008 - $0.015 (using Haiku for QA gate, Sonnet for generation)
- Processing time per lead: 12-25 seconds end-to-end
- Reply rate improvement vs. template-only outreach: 2.8x - 4.1x in our client deployments
- SDR time saved: 35-50 hours per week per rep at 500+ leads/week volume
The enrichment APIs are typically the biggest cost variable. If budget is tight, prioritise Clearbit or a similar firmographic source and drop LinkedIn scraping to a sample-based approach rather than every lead.
Common Failure Points (And How to Avoid Them)
Over-engineering the prompt on day one. Start with a simple two-paragraph email structure. Add complexity only after you've validated reply rates. Prompt bloat is the number one cause of inconsistent output quality.
Skipping the quality gate. Without it, bad emails will go out. It takes 20 minutes to set up and prevents reputation damage that takes months to undo.
Ignoring sending infrastructure hygiene. Claude can write the perfect email. If your domain has a spam score above 3 on MXToolbox, it won't matter. Warm your inboxes for at least 3 weeks before live sends.
No error handling. n8n workflows fail silently if you don't build error branches. Add an Error Trigger node that writes to a Slack channel and a Postgres error log on any workflow failure. You want to know immediately when something breaks, not three days later.
If you're looking to implement this kind of infrastructure without building it from scratch, our AI automation service covers end-to-end pipeline design, including prompt engineering, n8n hosting, and ongoing optimisation. We also integrate outreach pipelines directly with SEO and GEO content strategies for clients running combined inbound/outbound growth programmes.
Scaling Beyond the Basics
Once your core pipeline is stable, there are several high-leverage extensions worth building:
- Multi-touch sequencing: Trigger follow-up variations based on open/click data from Instantly, with each follow-up generated fresh by Claude rather than being a static template
- A/B testing at the prompt level: Route alternating leads through two different system prompt variants and log reply rates per variant in Supabase for statistical comparison
- Intent signal triggers: Connect to a tool like Bombora or G2 to trigger outreach automatically when a prospect's company shows high intent signal for your category
-
RELATED READING