The Simple Stack: What You Actually Need
Let’s get one thing straight: you don’t need a complex, multi-headed AI agent framework to solve 95% of your automation problems. Most of the time, you need three things: a trigger, a brain, and an action. That’s it.
- The Trigger: This is where the data comes from. It’s your website contact form on Webflow, a new row in a Google Sheet, or an email parsed by Mailgun. It’s the starting pistol.
- The Brain: This is the Large Language Model (LLM). Your options are typically OpenAI’s GPT models or Anthropic’s Claude models, accessed via their APIs. This is where the ‘thinking’ happens.
- The Action: This is where the result goes. A message in a Slack channel, a new card in Trello, an entry in your CRM, or an email sent to the lead.
And the glue holding it all together? A no-code automation platform. The two big players are Make (formerly Integromat) and Zapier automations. I build almost exclusively on Make. Its visual, flowchart-style interface makes it far easier to see how data moves and where things might break. Honestly, I think Zapier’s task-based pricing is hostile to experimentation; you get scared to run tests because you’re burning through your plan’s allowance. Make’s pricing is based on operations, which is generally more forgiving for the kind of multi-step AI workflows we’re building.
What Most Guides on AI Automation Get Wrong
Most of the tutorials you’ll find online are written by content marketers, not operators. They make a few critical mistakes that lead to brittle, unreliable automations that break the second they face real-world data.
AI Side Hustles
Practical setups for building real income streams with AI tools. No coding needed. 12 tested models with real numbers.
Get the Guide → $14
First, they obsess over the wrong tools. They’ll talk about exotic AI agent frameworks that require you to write Python, when a simple, deterministic workflow in Make would solve the problem faster and more reliably. The goal isn’t to build the most technically impressive system; it’s to build the most reliable one.
Second, they treat the prompt as an afterthought. The prompt isn’t just a detail; it’s the entire application. A lazy, poorly-structured prompt will give you inconsistent output, often breaking the downstream steps of your automation. A good prompt is specific, gives examples, and strictly defines the output format.
Finally, they don’t plan for failure. APIs go down. LLMs hallucinate and return garbage. A user types an emoji into the ‘budget’ field of your form. A real production system anticipates these failures. Most tutorials build a happy-path-only workflow that shatters at the first sign of trouble, often failing silently so you don’t even know it’s broken for days.
Building the Lead Qualifier: Step-by-Step
Let’s build something real. We’ll create an automation that takes a submission from a website contact form, uses an LLM to classify and score it, and then routes it accordingly. We’ll use Make for the logic and Anthropic’s Claude 3 Sonnet for the brain, because it’s fast, capable, and inexpensive.
Step 1: The Trigger (Webhook)
In Make, create a new scenario. The trigger module will be ‘Webhooks’ > ‘Custom webhook’. Make will generate a unique URL. You’ll paste this URL into the webhook settings of your form builder (like Webflow or Carrd). Now, whenever someone submits the form, the data is sent directly to your Make scenario.
Step 2: The First LLM Call (Triage)
Add an ‘Anthropic’ > ‘Create a Completion’ module. Connect it to the webhook. Our first goal is just to triage the message. Is it a real lead, a support question, or just spam? We’ll feed the text from the form’s ‘message’ field into the LLM with a clear, simple prompt.
Here’s a prompt that works well:
You are a helpful assistant responsible for classifying incoming messages from a website contact form. Your job is to determine the message's primary intent. Please classify the following message into one of three categories: SALES, SUPPORT, or SPAM.
Message: "{{1.message}}"
Respond with ONLY one of the three categories and nothing else.
This prompt is short, gives the AI a role, defines the exact outputs, and tells it to be concise. The `{{1.message}}` is how you map the data from the webhook into the prompt.
Step 3: The Router
Next, add a ‘Router’ module. This splits your workflow into different paths. We’ll create three paths branching off the router. The first path has a filter: `[Triage Output] = SALES`. The second: `[Triage Output] = SUPPORT`. The third: `[Triage Output] = SPAM`. Now your automation can handle each type of message differently.
Step 4: The Second LLM Call (Enrichment)
On the ‘SALES’ path, add another Anthropic module. This one is for extracting structured data. We want to pull out key details and get a lead score so we can prioritize our follow-up. This is where forcing JSON output is critical.
The prompt for this step is more complex:
You are a data extraction assistant. Your task is to analyze the following sales lead message and extract key information into a JSON object. The message is from a potential client.
Message: "{{1.message}}"
Extract the following fields:
- contact_name: The name of the person. If not present, use null.
- company_name: The company they work for. If not present, use null.
- project_details: A brief, one-sentence summary of what they want.
- budget: The estimated budget. If not mentioned, use null.
- lead_score: An integer score from 1 to 10 based on how qualified this lead seems. A high score means a clear project, a mentioned budget, and a sense of urgency.
Respond with ONLY the valid JSON object. Do not include any explanatory text, markdown formatting, or greetings.
The output of this module will be a clean JSON string that you can use in the next steps, like adding a new row to a Google Sheet or creating a deal in your CRM.