Build a AI SDR: Hyper-Personalized Sales Outreach at Scale — n8n Workflow Workflow with n8n
Your reps spend 40% of their day not selling. They're staring at LinkedIn tabs, skimming company blogs, and trying to write a cold email that doesn't read like a template. A good SDR sends maybe 30 ge
Your reps spend 40% of their day not selling. They're staring at LinkedIn tabs, skimming company blogs, and trying to write a cold email that doesn't read like a template. A good SDR sends maybe 30 genuinely personalized emails a day. A mediocre one sends 200 identical ones that go straight to spam. Neither scales, and both burn money. The bottleneck isn't your offer — it's the research-and-write loop that has to happen before every single touch.
Why manual personalization doesn't scale
Personalization works. Emails that reference a prospect's actual business get 2-3x the reply rate of generic blasts. Everyone knows this. The problem is the math: real personalization takes 8-12 minutes per lead — open their site, figure out what they do, find a hook, write an opener that connects your product to their reality. Multiply that by a 500-lead list and you've committed 80+ hours before you've had a single conversation.
So teams cheat. They fake personalization with mail-merge tags — {{first_name}}, {{company}} — which every buyer now recognizes on sight. Or they hire more SDRs, which pushes cost-per-meeting up and quality down. The result is a false choice: personalized-but-slow, or fast-but-ignored. What you actually want is the research done for you, at machine speed, with a human only in the loop for the final judgment call.
The solution: an AI SDR that reads and writes for you
This is exactly what an n8n-based AI SDR does. You hand it a lead list — name, company, website URL. For each lead, the workflow visits the company website, extracts what the business actually does, feeds that context to GPT-4o, and generates a personalized opening line grounded in the prospect's real situation. It writes the draft. You review, tweak if needed, and hit send. The 10-minute research task collapses into a 20-second review.
The key distinction: this isn't a spam cannon. The AI does the reading and the first draft — the parts that are tedious and mechanical. You keep the judgment: is this hook actually good, is this the right prospect, is the ask right. That division is what keeps deliverability high and your domain reputation intact. You're scaling the research, not the send button.
Step-by-step: building it in n8n
Here's the architecture. Every node below maps to a real step in the workflow, and the whole thing runs on a standard n8n instance (self-hosted or cloud).
1. Trigger and lead source. Start with a Google Sheets Trigger node (or a manual Google Sheets read node) pointed at your lead list. Columns: first_name, company, website, email. Set the trigger to poll on new rows, or run it on-demand when you drop a fresh list in. If your leads live in a CRM, swap this for a HubSpot or Pipedrive node filtering on a "needs outreach" stage.
2. Loop through leads. Add a Loop Over Items (Split in Batches) node set to a batch size of 1. This processes one lead at a time, which keeps you under rate limits on both the scraping and the OpenAI side, and makes debugging a single lead trivial.
3. Fetch the prospect's website. Use an HTTP Request node with the method set to GET and the URL mapped to {{ $json.website }}. Set a response timeout of ~15 seconds and enable "Continue On Fail" so one dead domain doesn't halt the whole run. Follow this with a Code node (or an HTML Extract node) that strips tags and pulls the readable text — grab the homepage headline, the meta description, and the first few paragraphs. Truncate to ~2,000 characters so you don't blow your token budget on navigation menus.
4. Generate the opener with GPT-4o. Drop in the OpenAI node (Message a Model), model set to gpt-4o. Your system prompt does the heavy lifting: "You are a senior SDR. Given a company's website text, write a single one-sentence opening line for a cold email that proves you understand their business. Be specific, reference something concrete from their site, no flattery, no buzzwords." Pass the scraped text in the user message. Set temperature around 0.6 — high enough to avoid robotic sameness, low enough to stay on-topic. The output is your personalized hook.
5. Assemble the email. A Set node (Edit Fields) combines the GPT opener with your fixed value proposition and CTA into a full email body. Keep the AI portion to the opener only — the body and ask stay consistent and human-written, which is what keeps your messaging on-brand and testable.
6. Human review before send. This is the non-negotiable step. Route drafts to a Google Sheets "drafts" tab or a Slack channel via the Slack node, with the generated email and a "approve" column. Only approved rows flow to the final Gmail (or SMTP) send node. You can gate this with an IF node checking {{ $json.approved === "yes" }}. Now you review 50 drafts in the time it used to take to write two.
What you actually gain
Throughput without the headcount. One person reviewing AI-drafted, genuinely-researched emails can cover the ground of a three-person SDR team — without the ramp time, the salaries, or the quality drift.
Personalization that survives a skeptical read. Because the opener is built from the prospect's actual website, it references things a mail-merge never could — a product they just launched, the market they serve, the language they use about themselves. Reply rates follow.
Deliverability protection. Human-in-the-loop review plus per-lead sending (not a 500-recipient blast) keeps you off spam blocklists and preserves your sending domain's reputation. The AI accelerates the work; it never removes the human checkpoint that keeps you safe.
A system you can tune. Every part is a node you can edit. Bad openers? Rewrite the system prompt. Wrong tone? Drop the temperature. New segment? Point the trigger at a different sheet. You own the logic, not a black-box SaaS.
Common pitfalls (and how to avoid them)
Scraping failures cascading. Websites time out, block bots, or return JavaScript-rendered pages with no text. Always set "Continue On Fail" on the HTTP Request node and add an IF node that routes leads with empty scraped text to a "manual" pile instead of feeding an empty prompt to GPT-4o (which will then hallucinate a generic opener — worse than no opener).
Trusting the AI blind. GPT-4o occasionally invents details or misreads a vague homepage. That's precisely why the review step exists. Never auto-send. The moment you remove the human gate, one hallucinated claim about a prospect's business torches your credibility with them.
Token and cost blowout. Feeding entire raw HTML pages into the model is expensive and noisy. Truncate to clean text, cap at ~2,000 characters, and you'll keep per-lead cost to fractions of a cent while improving output quality.
Rate limits. Sending too fast from a fresh domain looks like spam to inbox providers. Add a Wait node inside the loop (30-90 seconds between sends) and warm your domain before scaling volume. Speed on research, patience on sends.
Stale lead data. Dead URLs and closed companies waste tokens. A quick validation node — or filtering your source list first — keeps the workflow pointed at real, reachable prospects.
Set this up once and your outreach stops being a headcount problem. The research runs while you sleep; you show up to a queue of grounded, personalized drafts and spend your time on the one thing a machine can't do — deciding who's worth talking to and closing them.
Ja construimos isso pra voce
Nao comece do zero. O AI SDR: Hyper-Personalized Sales Outreach at Scale — n8n Workflow e um workflow n8n pronto para instalar — conecta suas ferramentas em minutos, sem codigo.
Instalar por $99.0 →
