AI agents for lead generation

Most founders waste hours copying names from LinkedIn or buying stale lists that never convert. You can fix that by building an AI agent that finds fresh prospects, checks their fit, and sends a personalized first touch — all without manual work. After reading this, you’ll have a working agent you can run on a schedule and a clear path to a pre‑built blueprint if you’d rather skip the setup.

Why manual lead lists fail at scale

Buying a list feels fast, but the data decays quickly. People change jobs, companies pivot, and email addresses bounce. When you rely on a static file you spend more time cleaning than selling. The real cost isn’t the price of the list; it’s the missed replies from prospects who never see your message because the contact is outdated.

Manual scraping is no better. You open LinkedIn, copy a name, paste it into a spreadsheet, then repeat. After fifty entries your eyes glaze over and mistakes creep in. Worse, you hit LinkedIn’s commercial use limit and get a temporary block that halts your outreach for a day.

An AI agent removes that grind. It runs on a schedule, pulls fresh profiles, applies your filters, and writes a first line that references a recent post or shared group. You still review the output, but the heavy lifting is done by code, not by you.

How do you prompt an AI agent to find qualified leads?

Prompt design is where most people get stuck.

Start with a clear role: “You are a lead‑generation researcher for a B2B SaaS company that sells workflow automation to mid‑size marketing teams.” Then give the agent a concrete task: “Find 10 LinkedIn profiles of marketing managers at companies with 50‑200 employees that posted about lead‑nurturing in the last 30 days.” Finally, specify the output format you need: a JSON array with fields name, title, company, LinkedIn URL, and a one‑sentence icebreaker based on the latest post.

Here’s a real prompt that works with GPT‑4‑turbo via the OpenAI API:

You are a lead‑generation researcher for a B2B SaaS company that sells workflow automation to mid‑size marketing teams.
Find 10 LinkedIn profiles of marketing managers at companies with 50‑200 employees that posted about lead‑nurturing in the last 30 days.
Return a JSON array where each object has: name, title, company, linkedin_url, icebreaker.
The icebreaker must reference the specific post you found (e.g., "Loved your recent post about nurturing leads with webinars — how’s the attendance trending?").
Do not include any extra text.

Run that prompt through your agent’s LLM call, parse the JSON, and you have a clean list ready for enrichment.

What most guides get wrong about AI agent tools

Many tutorials treat the agent as a magic black box. They tell you to “connect your data source” and “let the AI work its magic.” In reality the agent is only as good as the data you feed it and the constraints you set. If you give it a vague goal like “find good leads” it will return random profiles that match none of your criteria.

Another common mistake is skipping the validation step. Agents can hallucinate company sizes or invent recent posts. Without a quick check — maybe a call to a company‑info API or a manual glance at the LinkedIn page — you risk sending outreach that feels spammy because the personalization is based on false information.

Finally, guides often ignore rate limits. They show a demo that pulls fifty profiles in a minute, then never mention that the same call will get you blocked after a few runs. Building in pauses, retry logic, and fallback sources is what turns a demo into a reliable pipeline.

A concrete named example: using PhantomBuster + GPT-4 for LinkedIn scraping

Let’s walk through a stack I’ve run in production for the last six months.

First, PhantomBuster handles the LinkedIn scrape. Its “LinkedIn Sales Navigator Export” phantom can pull profiles based on a saved search URL. You set the search to filter by company size, seniority, and recent activity. The phantom returns a CSV with name, title, company, and profile URL.

Second, a short Python script reads that CSV, extracts the LinkedIn URL, and calls the GPT-4‑turbo endpoint with the prompt shown earlier. The script adds a 1.2‑second pause between requests to stay under PhantomBuster’s usage limits and OpenAI’s rate‑limit of 3,500 requests per hour for the tier I use.

Third, the script writes the enriched JSON to a Google Sheet via the gspread library. A simple Apps Script then triggers a Gmail draft for each lead, inserting the icebreaker into the template.

Here’s the numbered step‑list as a code block you can copy:

1. Create a PhantomBuster agent: LinkedIn Sales Navigator Export.
2. Define a Sales Navigator search: marketing manager, 50‑200 employees, posted in last 30 days.
3. Run the phantom, download the CSV.
4. For each row:
   a. Send the LinkedIn URL to GPT‑4‑turbo with the lead‑gen prompt.
   b. Parse the returned JSON.
   c. Write name, title, company, linkedin_url, icebreaker to Google Sheet.
5. Use Apps Script to send a Gmail draft per row, using the icebreaker as the opening line.
6. Schedule the whole workflow to run daily at 8 am via PhantomBuster’s scheduler and a cron job for the script.

What I love about this combo is how the agent enriches each profile with a genuine icebreaker — no generic “I noticed we’re both in tech” fluff. The icebreaker references a real post, which lifts reply rates from roughly 8 % to over 22 % in my tests.

My gripe? PhantomBuster’s UI hides the usage‑meter behind a few clicks, and I once burned through my monthly credit limit because I forgot to check the dashboard after a weekend test. A simple email warning at 80 % would have saved me a scramble.

(Which, yes, is annoying — but it’s a solvable annoyance once you know where to look.)

How to debug when this breaks

When the agent stops delivering leads, start at the data source. Open the PhantomBuster execution log; look for errors like “Session expired” or “Captcha required.” If you see those, regenerate the session cookie in the phantom settings and rerun.

If the logs show a successful scrape but the Google Sheet stays empty, check the script’s console output. A common failure is the JSON parse step — GPT sometimes returns a trailing apology or extra text before the array. Add a regex that extracts the first ‘[‘ and last ‘]’ before parsing, or ask the model to output only JSON in the system message.

If enrichment works but the Gmail drafts never appear, verify the Apps Script trigger. Sometimes a silent failure occurs because the script exceeds its runtime limit; break the loop into batches of twenty leads and let the trigger fire again.

Finally, watch your API quotas. PhantomBuster credits and OpenAI tokens both have hard ceilings. Set up a simple monitoring script that pings their usage endpoints each morning and sends you a Slack warning when you’re at 70 % of the monthly limit.

Pricing opinion and when to buy the blueprint

PhantomBuster’s “Advanced” plan runs $29 per month and gives you 10 hours of execution time — enough for a daily LinkedIn scrape of 200 profiles. OpenAI’s GPT‑4‑turbo costs about $0.03 per 1k tokens; my prompt‑response pair averages 800 tokens, so roughly $0.024 per lead. At 200 leads a day the LLM bill stays under $15 monthly.

I think paying $199 per month for an all‑in‑one lead‑gen AI platform is overpriced for most solopreneurs; you’re mostly paying for a polished UI that hides the same two‑piece stack I just described. The free tier of PhantomBuster is useless for anything beyond a test — it caps you at 10 minutes of runtime, which won’t finish a single search.

If you want the deep cut on this, deeper coverage of AI agent platforms.

If you’d rather skip the build and deploy a working version in an afternoon, we’ve packaged this workflow as a blueprint at deepusecase.com/vault/ai-agent-builder-kit.