best AI blueprints for agencies

Last month I needed to pull 500 leads from LinkedIn, enrich them with company size and tech stack, then send personalized cold emails for my agency’s new service. Doing it manually ate up two full days and left me with typos and missed follow‑ups. After reading this you’ll be able to assemble a repeatable AI blueprint that runs the same workflow in under an hour.

How do you handle rate limits when scraping LinkedIn?

LinkedIn throttles aggressive scrapers hard; after about 30 profile views per hour you’ll see a temporary block that can last several hours. I ran into this when I tried a simple Python script with Selenium and got locked out after just 45 profiles. The fix is to throttle requests yourself and spread the work over multiple days or use a service that already respects LinkedIn’s limits.

I now use PhantomBuster’s LinkedIn Sales Navigator scraper. It runs in the cloud, rotates IPs, and lets you set a delay between actions. In my test I set a 2‑second pause and pulled 120 profiles per hour without hitting a block. The free tier gives you 10 hours of execution time a month, which is enough for roughly 200 profiles.

If you need more volume, the Starter plan at $15/mo buys 100 hours of runtime. I think $15/mo is fair for the amount of data you get, especially when you factor in the time saved.

— and good luck finding docs for this — the PhantomBuster interface hides the rate‑limit settings under a “Advanced” tab that’s easy to miss.

I’ve tried the free PhantomBuster plan and ran out after three days of testing.

What most guides get wrong about AI agent hallucinations

Many tutorials tell you to just feed raw LinkedIn data into GPT-4 and expect perfect outreach lines. In practice the model will invent job titles, fake certifications, or claim a prospect works at a company they left two years ago. I once sent an email that congratulated a VP on a product launch that never happened; the reply was a confused “What are you talking about?”

The better approach is to ground the AI in verified facts. I pull the company’s LinkedIn “About” section and the employee’s current title, then ask GPT-4 to write a short intro that only uses those two pieces. My prompt looks like this:


You are a friendly sales assistant. Use ONLY the following facts:
- Prospect name: {{firstName}} {{lastName}}
- Current title: {{title}}
- Company: {{companyName}}
- Company description: {{companyDesc}}
Write a 2‑sentence opener that mentions a specific detail from the company description and shows why you’re reaching out. Keep it under 30 words.


By limiting the model to the supplied fields, hallucinations drop to near zero. I still review each output, but the error rate went from about 18 % to under 2 %.

Concrete example: PhantomBuster, Make.com, and GPT-4 in action

Here’s the full flow I run every Monday morning:

1. PhantomBuster scraper exports a CSV of LinkedIn profiles (name, title, company URL) to a Google Sheet.
2. Make.com watches that sheet for new rows.
3. For each row, Make.com calls the PhantomBuster “Company Info” agent to grab the company description and employee count.
4. Make.com then sends a request to the OpenAI API (GPT‑4‑turbo) with the prompt shown above, using the scraped name, title, company name, and description.
5. The generated opener is written back to the sheet in a new column.
6. Finally, Make.com triggers a Gmail draft with the opener inserted into a templated cold email.

I keep the whole scenario active for about 20 minutes each week. The cost breakdown:

• PhantomBuster Starter: $15/mo
• Make.com Free tier (enough for 10 000 operations/mo)
• OpenAI API: ~ $0.03 per 1 K tokens; my average call uses 600 tokens → ~$0.018 per lead. For 500 leads that’s under $10/mo.

I think the total of ~$25/mo is a steal compared to hiring a VA for $8/hour.

How to debug when this breaks

When something goes wrong, the first place I look is the Make.com execution log. Each module shows its input and output, so you can see if the scraper returned empty fields or if the API call failed with a 429.

A common hiccup is the PhantomBuster agent timing out because the LinkedIn page took too long to load. The log will display a “TimeoutError” after 30 seconds. I fix it by increasing the agent’s timeout setting from 30 s to 60 s in the PhantomBuster dashboard.

If the OpenAI call returns a blank response, I check the prompt length. The API will reject prompts over 8 192 tokens; my prompt stays well under that, but if I accidentally concatenate the whole CSV row the size can balloon. I solve it by using Make.com’s “Set variable” step to pull only the needed fields before the API call.

It’s annoying when the email draft lands in the spam folder because Gmail flags the templated link.

Finally, I always keep a backup CSV of the raw scrape. If the AI step corrupts data, I can rerun the enrichment without re‑scraping LinkedIn.

Pricing and opinion: is the stack worth it?

Let’s be blunt: the free tiers of PhantomBuster and Make.com will get you started, but you’ll hit limits fast if you run more than a couple of hundred leads per month. I tried the free PhantomBuster plan (10 hours) and ran out after three days of testing. The $15/mo Starter plan removed that bottleneck.

I love how Make.com’s visual scenario builder lets you see data flow in real time; it cut my debugging time from hours to minutes. That feature alone justifies the $9/mo I pay for their “Core” plan when I need premium apps.

My gripe is that PhantomBuster’s billing page forces you to enter a credit card before you can even view the usage dashboard. I wish they offered a read‑only preview first.

Overall, for a solo agency owner spending less than $30/mo on automation, the time saved is worth every penny. If you’re handling thousands of leads each month, you’ll probably upgrade to the PhantomBuster Team plan at $45/mo, which still feels reasonable.

Adjacent reading: 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.