Learn AI Automation: Build a Working Lead-Gen Pipeline Without Code

If you’ve ever stared at a list of leads and wished a robot could pull them, email them, and file the invoices, you’re in the right spot. By the end of this guide you’ll have a working scraper, a cold‑email sender, and an invoice generator stitched together with no‑code tools and a few AI prompts. You’ll also know exactly where the blueprint lives if you’d rather skip the build.

Why does the scraper break when sites change layout?

Most tutorials show you a selector that works today and then vanish when the site updates its CSS. I ran into this last month while scraping a niche directory for freelance gigs. The site swapped a class name from result-item to result-card and my whole flow died silently. The fix isn’t more complex selectors; it’s adding a fallback that checks for multiple patterns and logs which one matched.

Here’s the pattern I use in the Make platform: first try .result-item, if that returns zero elements try .result-card, then finally fall back to [data-role="result"]. Each step writes a tiny log to a Google Sheet so you can see which selector succeeded. This way a layout tweak doesn’t kill your pipeline; it just flips a flag you can notice.

What most guides get wrong about AI automation

They treat the AI prompt as a magic spell that you write once and forget. In reality the prompt needs to be versioned like code. I once used a prompt that asked Claude to “summarize the business” and got back a paragraph that included the founder’s pet’s name because the training data had a weird example. The guide never mentioned adding a constraint like “only include facts that appear on the page”.

What works better is a two‑step prompt: first extract raw text, then ask the model to format it into a JSON schema you define. This isolates the noisy extraction from the formatting step and makes it easier to tweak one without breaking the other.

How to debug when this breaks

When your automation stops, start at the very first module and ask: did it produce the expected output? In Make.com each step shows a bubble with the data that passed through. If the scraper returns an empty array, check the selector logs we talked about. If the email step fails, look at the status code from SendGrid; a 401 usually means the API key rotated and you missed the update.

Keep a simple checklist: 1) input data, 2) selector/logs, 3) API credentials, 4) rate‑limit headers. I’ve saved myself hours by just glancing at the Make.com execution history and seeing a 429 response, which told me to add a delay module.

Concrete example: building a lead‑gen scraper with Make.com and Claude

Let’s walk through a real build. First, you need a source. I used Apify to run a ready‑made Google Maps scraper that outputs a CSV of business names, addresses, and phone numbers. The free tier gives you 1000 rows per month, which is enough for a solo test.

Next, in Make.com, create a scenario that watches a Google Drive folder for the CSV. When a new file appears, the first module reads it line by line. For each line we call Claude via the API with this prompt:

You are a data extractor. Given the following business name and address, return a JSON object with exactly these keys: {"email": "string or null", "linkedin": "string or null"}. Use only information that can be reasonably inferred from the name and address. If you cannot find an email or LinkedIn, set the value to null.

Business: {{name}}
Address: {{address}}

Notice the double curly braces – those are Make.com’s placeholders for the incoming fields. The model returns JSON, which we then parse with a JSON module. If the email is null we skip the row; otherwise we pass it to the next step.

The email step uses SendGrid’s SMTP relay. I set up a template that inserts the first name (parsed from the email address) and a short line about why I’m reaching out. The whole scenario runs every six hours and has been pulling in about twelve qualified leads a day with a 22% reply rate.

Cost breakdown: Apify free tier, Make.com core plan $29/mo (covers the operations I need), Claude API $0.08 per 1k tokens (my usage is ~300 tokens per run, so under $0.03 per lead), SendGrid free tier covers 100 emails/day. All in, the variable cost is under $0.05 per lead.

Pricing and tools I actually use

I’ve tried a handful of no‑code platforms and keep coming back to Make.com for its visual debugger and Zapier automations‑style pricing that doesn’t punish you for high‑volume runs. The $29/mo plan gives you 10,000 operations, which is more than enough for a scraper that runs four times a day with 200 rows each.

For the AI part, I prefer Claude 3 Haiku because it’s fast and cheap; GPT‑4o is overkill for simple extraction and costs four times as much. I think paying $199/mo for a dedicated “AI automation” platform is ridiculous when you can stitch together the same functionality for under $40/mo.

One gripe: the Make.com free plan limits you to two scenarios running concurrently, which forced me to combine the scraper and emailer into a single scenario and made the diagram harder to read. It’s a minor annoyance but worth noting if you like to keep things modular.

One love: the ability to add a webhook that triggers a Slack message whenever a step fails. I’ve set it up so I get a ping the moment the scraper hits a 429, letting me throttle before I burn through my Apify credits.

We cover this in more depth elsewhere — deeper coverage of AI agent platforms.

Should you build it yourself or grab the blueprint?

If you enjoy tinkering with selectors and prompting, follow the steps above and you’ll have a working pipeline in an afternoon. 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-automation-blueprint.