n8n Tutorial: Apollo + Instantly SDR — Full Cold Email Pipeline with AI Personalization Automation
Your SDRs spend 70% of their day on work that doesn't require a human: pulling lists, checking who a company is, guessing at a first line, copy-pasting into a sender. The actual selling — the reply, t
Your SDRs spend 70% of their day on work that doesn't require a human: pulling lists, checking who a company is, guessing at a first line, copy-pasting into a sender. The actual selling — the reply, the objection, the call — is maybe an hour of real work buried under six hours of manual plumbing. That plumbing is the exact shape of a problem n8n solves. This tutorial walks through a full cold email pipeline that runs itself: Apollo sources the leads, Firecrawl researches each company, GPT-4o writes a personalized opener, and Instantly sends the sequence. You define the ICP once. The machine does the rest.
The problem: personalization doesn't scale by hand
Everyone knows generic cold email is dead. "Hi {{firstName}}, I wanted to reach out about..." lands in spam or gets deleted in half a second. The advice is always "personalize" — but real personalization means reading each prospect's website, understanding what they actually do, and referencing something specific. Done manually, that's 4–6 minutes per lead. At 200 leads a week, that's a full workday of research before a single email is written.
So teams cheat. They spin up "personalization" that's really just {{companyName}} jammed into a template, and reply rates stay under 1%. The founders who win cold email in 2026 aren't sending more — they're sending emails that prove a human (or a well-instructed model) actually looked at the account. The bottleneck was never writing. It was research at volume. That's automatable.
The solution: a five-stage pipeline in n8n
The workflow chains five services, each doing one job well:
- Apollo — pulls leads matching your ICP filters (title, headcount, industry, tech stack, geography).
- Firecrawl — scrapes each lead's company site and returns clean, LLM-ready markdown.
- GPT-4o — reads that research and writes a personalized first line plus a full opener.
- Instantly — drops the lead into a warmed sending sequence with proper throttling.
- n8n — orchestrates all of it, handles retries, and keeps state so you never double-contact.
The design principle: keep the model dumb about logic and smart about language. n8n handles branching, dedup, and rate limits deterministically. GPT-4o does only what it's good at — turning research into a sentence a human would actually write. That separation is why the pipeline stays reliable at 500+ leads a day.
Step-by-step: building it in n8n
1. Trigger and ICP definition. Start with a Schedule Trigger (e.g. daily at 07:00) or a Manual Trigger for testing. Immediately after, use a Set node to hard-code your ICP parameters — titles, industries, employee_count_min/max, locations. Keeping the ICP in one Set node means you tune targeting in a single place instead of hunting through the flow.
2. Source leads from Apollo. Use an HTTP Request node pointed at Apollo's POST /api/v1/mixed_people/search endpoint. Pass your Api-Key in the header and map the Set node's ICP fields into the JSON body. Set per_page to 25–50 and paginate with a loop if you need more. The response gives you names, titles, company domains, and email status. Feed the array into an Item Lists / Split Out node so each lead becomes its own item flowing down the pipeline.
3. Dedup before you spend money. Insert a NocoDB / Google Sheets / Postgres node (whichever you use for state) to check whether the email already exists in your contacted list. Route through an IF node — already-seen leads exit here. This one check protects your sender reputation and stops you paying Firecrawl/OpenAI credits on duplicates.
4. Research each company with Firecrawl. For every surviving lead, call Firecrawl's /v1/scrape endpoint via HTTP Request, passing the lead's company domain and formats: ["markdown"]. Firecrawl returns clean text stripped of nav and boilerplate. Trim it with a Code node to the first ~2,000 characters — enough signal for the model, cheap on tokens. Wrap this node with Continue On Fail enabled so a dead site doesn't kill the whole batch; a fallback line ("no site data") lets the lead proceed with lighter personalization instead of erroring out.
5. Write the email with GPT-4o. Use the OpenAI node (or an HTTP Request to /v1/chat/completions) with model gpt-4o. Your system prompt sets the voice and hard rules: one specific observation from the research, one line connecting it to your offer, under 90 words, no emojis, no "I hope this finds you well." Pass the Firecrawl markdown and the lead's title in the user message. Ask for a structured JSON response — { "first_line": "...", "subject": "...", "body": "..." } — and set response_format to JSON so parsing downstream never breaks. A Set node maps those fields onto the lead item.
6. Push into Instantly. Finish with an HTTP Request to Instantly's POST /api/v2/leads endpoint, adding the lead to a specific campaign ID with the generated subject and body mapped into custom variables. Instantly handles warmup, throttling, and the multi-step follow-up sequence. Close the loop by writing the contacted email back to your state table so tomorrow's run skips it.
Benefits: what actually changes
Time. The manual version is 4–6 minutes of research per lead. This runs unattended overnight and hands you a fully personalized, queued campaign by morning. A single operator now covers what took a three-person SDR team.
Reply rates. Because every email references something real from the prospect's own site, opens and replies climb into ranges generic blasts never touch. You're not sending more email — you're sending email that earns a response.
Consistency and cost control. The pipeline never has a bad day, never skips the research step, never forgets the dedup check. And because n8n handles the orchestration, your only variable cost is API usage — Apollo, Firecrawl, and OpenAI credits — which for most teams runs a few cents per fully-researched, personalized lead. Compare that to an SDR's hourly rate doing the same work.
Full ownership. No SaaS "AI SDR" black box charging per seat. You own the workflow, the prompts, and the data. Swap Apollo for another source, GPT-4o for a cheaper model on low-value tiers, or Instantly for Smartlead — the graph stays the same.
Common pitfalls (and how to avoid them)
Sending unverified emails. Apollo returns email status flags. Route through an IF node and only push verified emails to Instantly — sending to guessed addresses tanks your domain reputation fast. When in doubt, add a verification step (e.g. an HTTP call to a validator) before the sender.
Firing everything in parallel. n8n will happily hammer all three APIs at once and trip rate limits or get you soft-blocked. Use a Loop Over Items (batch size 5–10) plus a Wait node to pace the run. Slower and complete beats fast and rate-limited.
Trusting the model blindly. GPT-4o occasionally hallucinates a detail or writes something off-brand. Add a Filter or lightweight validation Code node that rejects outputs over your word limit or containing banned phrases, and log a sample daily. For high-value tiers, route to a manual review sheet before sending.
No state = double-sends. If you skip the dedup table, a re-run will contact everyone again — the fastest way to burn a domain. Persist contacted leads on every run, and key on email, not name.
Warmup neglect. The best pipeline in the world dies if your sending domain is cold. Let Instantly warm the mailbox for 2–3 weeks before volume, keep daily sends conservative early, and use a dedicated domain — never your primary company domain.
Build it once and cold outreach stops being a headcount problem. You define the ICP, and the pipeline turns strangers into a queue of researched, personalized conversations while you focus on the only step that still needs you: closing.
Ja construimos isso pra voce
Nao comece do zero. O Apollo + Instantly SDR — Full Cold Email Pipeline with AI Personalization e um workflow n8n pronto para instalar — conecta suas ferramentas em minutos, sem codigo.
Instalar por $99.0 →
