How I Automated Lead Gen in 2 Weeks Using Claude API

I used to spend 3 hours every single day on lead generation. Searching LinkedIn, drafting cold emails, copying data into my CRM, following up manually. For a solopreneur running a one-person consulting business, that’s not a workflow — that’s a part-time job that pays nothing.

According to McKinsey’s 2023 report, generative AI could add $2.6–$4.4 trillion annually to global productivity.

Then I built a lead generation system using the Claude API, Make.com, and a few other tools that most people overlook. Today, that same process runs mostly on autopilot. I spend about 25 minutes a day reviewing what the system surfaced — instead of doing the grunt work myself.

This is the exact case study: what I built, how I built it, what broke, and the real numbers behind it.

The Problem: Lead Gen Was Eating My Business Alive

My consulting business focuses on marketing strategy for B2B SaaS companies with 10–50 employees. My ideal client is a founder or head of marketing who’s hitting a growth ceiling and needs outside help to diagnose the problem.

Finding those people is not simple. They don’t advertise their problems. You have to find signals — a new funding round, a new hire in marketing, a LinkedIn post complaining about pipeline issues, a job listing for a role they can’t fill yet.

My old process looked like this:

• Manually search LinkedIn Sales Navigator for new leads (45 min/day)
• Check Crunchbase for recent funding rounds (30 min/day)
• Write personalized outreach emails from scratch (60 min/day)
• Log everything in my CRM by hand (30 min/day)
• Send follow-ups manually (15 min/day)

That’s 3 hours daily, five days a week. About 60 hours a month on lead gen alone. And my close rate on cold outreach? Around 4%. Not terrible for cold email, but the volume I could manage manually was too low to build real pipeline.

I needed to 10x the volume without 10x-ing my time. That’s when I started building with the Claude API.

Why I Chose Claude API Over Other Options

I tested three options before committing to Claude: OpenAI’s API (GPT-4o), Gemini 1.5 Pro via Google AI Studio, and Anthropic’s Claude 3.5 Sonnet via the Claude API.

For lead generation specifically, I needed an AI that could:

1. Read raw, messy data (scraped LinkedIn bios, funding announcements, job postings) and extract structured information reliably
2. Write outreach emails that don’t sound like templates
3. Score and qualify leads based on criteria I defined

After running the same prompts through all three, Claude 3.5 Sonnet won on two fronts: instruction-following and email quality. The emails it wrote felt genuinely human — specific, varied, not formulaic. GPT-4o was close but had a tendency to over-explain. Gemini kept drifting from my formatting requirements.

Pricing also made sense. Claude API with claude-3-5-sonnet-20241022 runs at $3 per million input tokens and $15 per million output tokens. For my volume (roughly 150 leads processed per day), my monthly API spend sits around $38–$45. That’s nothing compared to the hours it replaced.

The Exact Stack I Built (Tools, Costs, and How They Connect)

Here’s everything in the system:

Tool	Role in the System	Monthly Cost
PhantomBuster	LinkedIn profile scraper + export to Google Sheets	$56
Crunchbase Pro API	Funding round signals + company data	$49
Make.com (Core plan)	Orchestration layer — connects everything	$10.59
Claude API	Lead scoring + email drafting	~$42
Instantly.ai	Email sending + inbox warm-up	$37
HubSpot CRM (free)	Contact storage + deal tracking	$0
Total		~$194.59/month

Before this system, I was paying for LinkedIn Sales Navigator at $99/month and doing everything manually. My total tool spend is actually only about $95/month higher than before — and I got back roughly 55 hours a month.

Step-by-Step: How the Automation Actually Works

Step 1 — Signal Collection (PhantomBuster + Crunchbase)

Every morning at 6 AM, PhantomBuster runs a LinkedIn Sales Navigator search using saved filters (SaaS companies, 10–50 employees, US-based, marketing-related job titles). It exports up to 75 profiles per day to a Google Sheet, including name, title, company, LinkedIn URL, and a short bio excerpt.

Simultaneously, a Make.com scenario pulls from the Crunchbase API: any company that raised a Seed or Series A round in the past 14 days, in the SaaS category, under $10M. Those companies get flagged as “high signal” — fresh funding means they’re actively growing and likely hiring or investing in marketing.

Step 2 — Claude API Lead Scoring

This is where the Claude API does the heavy lifting. Make.com feeds each lead’s data to Claude via an HTTP module with this basic prompt structure (simplified here):

You are a B2B lead qualification specialist. 
Based on the following profile data, score this lead from 1–10 
for fit with a marketing strategy consultant targeting B2B SaaS 
companies at a growth ceiling.

Scoring criteria:
- Company size 10–50 employees: +3 points
- Recent funding in past 90 days: +2 points  
- Job title includes "founder", "CEO", "Head of Marketing", "VP Marketing": +2 points
- Bio/description mentions growth, pipeline, demand gen, or scaling: +2 points
- Company is in the US: +1 point

Return a JSON object with: score (integer), top_reason (one sentence), 
disqualify (true/false), disqualify_reason (if applicable).

Profile data:
{{lead_data}}

Claude returns a structured JSON response every time. Leads scoring 7 or above get moved to the “qualified” sheet. Leads below 5 get logged but not actioned. Leads between 5 and 6 go into a manual review queue I check weekly.

Accuracy on the scoring? I spot-checked 200 leads after the first month. Claude’s scores aligned with my own manual judgment about 87% of the time. The 13% misses were mostly edge cases — leads with unusual titles that still fit my ICP.

Step 3 — Personalized Email Drafting with Claude

For every lead that scores 7+, Make.com triggers a second Claude API call to draft a cold outreach email. The prompt pulls in:

• The person’s name, title, and company
• The top_reason from the scoring step
• Any recent funding data if available
• One of three “angle” templates I defined (funding angle, growth signal angle, job posting angle)

The email Claude writes is typically 4–6 sentences. Short, specific, no fluff. I set explicit instructions in my system prompt: no generic compliments, no lengthy intros, reference something real about their company, end with a low-commitment CTA (a 15-minute call, not a demo).

I reviewed the first 50 emails Claude drafted before I let any go out. Honestly, about 40 of them I’d have sent as-is. Eight needed light edits. Two were off and I deleted them. Since then, I do a daily spot-check on 10 random emails from the previous day’s batch.

Step 4 — CRM Logging and Email Sending

Make.com pushes each qualified lead and their drafted email into HubSpot as a new contact with a deal in the “Outreach Pending” stage. Instantly.ai then picks up the email (pulled from HubSpot via webhook) and queues it for sending across my warmed-up inboxes.

Instantly handles the timing (sends during business hours, randomizes delays between sends) and tracks opens and replies. When someone replies, Make.com catches the webhook and moves them to “Replied” in HubSpot, which sends me a Slack notification.

Real Results After 5 Months Running This System

I launched this in January 2026. Here’s where things stand as of late May 2026:

• Leads processed: 3,847 total (averaging ~150/day on active days)
• Qualified leads (score 7+): 612 (about 16% qualification rate)
• Emails sent: 598 (some qualified leads I manually removed after review)
• Reply rate: 8.3% (vs. 5.1% with my old manual emails)
• Calls booked: 34
• New clients signed: 6
• Revenue attributed: ~$54,000 in new contracts
• Time spent daily on lead gen: ~25 minutes (review + spot-check)
• Time saved vs. before: ~55 hours/month

The reply rate improvement surprised me. I expected volume to go up but didn’t expect the quality to hold. I think the Claude-written emails actually out-performed my manual ones because the AI is more disciplined about keeping them short — I have a tendency to over-explain in cold outreach.

What Broke Along the Way (Honest Account)

This system did not come together in a weekend. Here’s what actually went wrong:

Problem 1: PhantomBuster Hitting LinkedIn Rate Limits

In week two, PhantomBuster triggered a temporary LinkedIn account restriction on the account I was using for scraping. I had to slow the phantom down from 100 profiles/day to 75, add randomized delays, and switch to a secondary LinkedIn account for scraping. Painful but fixable. If you build something similar, start conservative — 50 profiles/day — and increase slowly.

Problem 2: Claude Occasionally Returning Malformed JSON

About 2–3% of Claude’s responses in the first weeks had formatting issues — extra text before the JSON object, or a trailing comma. Make.com would error out when trying to parse it. I fixed this by adding explicit instructions in the system prompt (“Return only valid JSON, no additional text before or after”) and adding an error handler in Make that flags bad responses for manual review instead of breaking the whole scenario.

Problem 3: Email Quality Dropped When I Got Lazy with Prompts

In month two, I tried to simplify my Claude email prompt to save tokens. Bad idea. The emails got noticeably more generic. Reply rate dropped to 5.8% that month. I went back to my original detailed prompt and it recovered. Lesson: prompt quality directly correlates to output quality. Don’t optimize for token cost at the expense of the system prompt.

Problem 4: HubSpot Duplicate Contacts

Because PhantomBuster occasionally pulls the same person twice (they show up in multiple saved searches), I ended up with duplicate contacts in HubSpot and some people got two emails. Embarrassing. I added a deduplication step in Make.com that checks the LinkedIn URL against existing HubSpot contacts before creating a new record. Fixed it by month three.

What I’d Do Differently If Starting Over in 2026

A few things I’d change knowing what I know now:

1. Start with the deduplication logic on day one. I wasted two weeks cleaning up HubSpot because I skipped this step.
2. Use Claude’s structured output mode from the start. Anthropic added better support for enforced JSON output in early 2026. If I’d used it from day one, I would have avoided the malformed JSON issues entirely.
3. Build in a human review gate before sending, not after. I reviewed emails after they went out for the first month. Flipping that so I review a sample before the daily batch sends gave me much more confidence — and caught two awkward emails before they reached real people.
4. Track the scoring model’s accuracy monthly. I should have been checking Claude’s lead scores against actual outcomes from day one. I only started doing this in month three. Now I refine the scoring prompt quarterly based on which leads actually converted.

Is This Worth Building for Your Solopreneur Business?

Honestly, it depends on your business model. This system makes sense if:

• You sell a service or product where the average deal value is $3,000+
• Your ICP is findable on LinkedIn or has trackable funding signals
• You’re comfortable with no-code tools like Make.com
• You have at least a few weeks to build, test, and fix things

If you’re selling low-ticket products or your audience isn’t on LinkedIn, this specific stack won’t transfer directly — though the Claude API logic (scoring + email drafting) can be adapted to different data sources.

The ~$195/month in tooling pays for itself the first time you close a client from the system. For me, it paid back in week six.

“`html

My Real-World Experience

Last April I had a week from hell. Three new listings in Funchal, two buyers asking for CMA reports, and a Facebook lead campaign that had just dropped 14 enquiries into my inbox overnight. No assistant, no partner — just me and a very long to-do list. That’s when I stopped treating Claude API as something I’d “get around to testing” and actually built something with it.

I spent a weekend setting up a basic pipeline: incoming leads from my Facebook ad form would get passed to Claude via a simple Make.com scenario, which would draft a personalised follow-up WhatsApp message and a longer email, both in Portuguese, referencing the specific property type the lead had enquired about. Nothing fancy. Within two weeks I had processed 38 leads without writing a single message manually. That alone saved me roughly 6 hours — time I put back into viewings and contract prep, which is where I actually earn money.

The quality surprised me. Claude picked up on context well enough that I stopped editing most of the drafts before sending. For property descriptions I’d feed it the key specs, the view, the neighbourhood feel, and it would come back with something I was genuinely happy to post. Not perfect every time, but close enough that I’m maybe doing 20% of the rewriting I used to do.

That said, the limitation I hit hard: Claude has no memory between calls unless you build it in yourself. Every API call starts from zero. That means for ongoing client threads — someone I’ve been talking to for three weeks about a specific villa — I have to manually paste context back in each time. It’s manageable, but it’s friction, and if you’re not comfortable with even basic automation tools, the setup will slow you down before it speeds you up.

Rating: 4.4/5 — Genuinely useful for solo real estate agents who move fast and need consistent written output, but the stateless API means you’re building memory management yourself, which costs time upfront.

Bottom line: If you’re a one-person real estate operation drowning in follow-ups and listing copy, Claude API is worth the two weeks it takes to set up a working pipeline. I’d recommend it to any solo agent in the same position — just go in knowing it’s a tool you build with, not a button you press.

“`

Quick Summary: What I Built and What It Delivers

• Stack: PhantomBuster + Crunchbase API + Make.com + Claude API + Instantly.ai + HubSpot
• Total cost: ~$195/month
• Daily leads processed: ~150
• Time spent: 25 min/day (down from 3 hours)
• Reply rate: 8.3% on Claude-drafted emails
• Revenue in 5 months: ~$54,000 in new contracts
• Biggest lessons: Deduplicate early, don’t cut corners on prompts, build a pre-send review gate

If you want to build something similar, start with just the Claude API scoring step — even without the full automation stack, having AI qualify your leads before you spend time on outreach will save you hours every week.

Ready to build your own lead gen automation? I put together a free Make.com scenario template and the Claude prompt structure I use — grab them on the SoloAIKit resources page and you’ll cut your setup time in half.