Building a Landing AI System to Qualify Inbound Leads

The Core Stack: Forms, Webhooks, and an LLM

You don’t need a team of engineers to build a functional landing AI system. You just need to connect a few off-the-shelf tools. This is plumbing, not rocket science. The entire system has three parts.

First, the form. This is where the lead comes from. I use Tally.so for this because its free tier is incredibly generous. You can also use Webflow for landing pages Forms, Typeform, or whatever else you have on your site. The only requirement is that it can send the submission data out via a webhook.

Second, the orchestrator. This is the glue that catches the form data and sends it to the AI. Zapier automations is the standard here, but Make.com is a solid alternative. This tool will receive the webhook and execute a series of steps, which we’ll define later. This is the piece you’ll likely have to pay for.

Third, the brain. This is the Large Language Model (LLM) that does the actual qualification. We’ll use OpenAI’s API, specifically the GPT-4o model. It’s fast, cheap, and excellent at following instructions for structured data output, which is the key to making this whole thing work reliably.

What Most Guides Get Wrong About AI Lead Scoring

If you search for how to do this, you’ll find a lot of bad advice. Most tutorials make one of two critical mistakes that render their systems useless in a real business context.

The first mistake is focusing on sentiment analysis. They’ll show you a prompt that asks the AI, “Is this lead positive or negative?” This is a complete waste of an API call. A happy, enthusiastic student who fills out your enterprise software form is a positive-sentiment, zero-value lead. You don’t care if they’re happy; you care if they can pay you.

The second mistake is using vague prompts that return unstructured text. A prompt like, “Summarize this lead and tell me if they are a good fit,” will get you a paragraph of text. Now you’ve just replaced the work of reading a form submission with the work of reading an AI’s summary. You haven’t automated anything. You’ve just added a step.

Qualification is not about sentiment. It’s about matching a lead against your Ideal Customer Profile (ICP).

The goal is to get a clear, machine-readable output. A score. A boolean (true/false). A reason that can be routed directly into a Slack message or CRM field. Anything else is a hobby project, not a business automation.

Crafting a Prompt That Actually Works

The prompt is the most important part of this entire landing AI system. This is where you encode your business logic. A well-crafted prompt gets you consistent, structured data. A lazy prompt gets you garbage.

Here is a template that works. You’ll need to customize the ICP section for your own business.

<system>
You are an expert B2B sales development representative named 'QualBot'. Your sole job is to analyze inbound leads from a website form and determine if they match our Ideal Customer Profile (ICP). You must be strict and objective. Respond ONLY with a valid JSON object. Do not include any explanatory text, markdown formatting, or apologies. Just the JSON.
</system>

<user>
Here is our Ideal Customer Profile (ICP) for qualification:
- **Company Size:** 10-100 employees.
- **Industry:** B2B Software-as-a-Service (SaaS) or Marketing Technology.
- **Role of Submitter:** Must be a founder, C-level executive, VP, or Director. Managers are acceptable if in a marketing or sales role. Analysts and students are an immediate disqualification.
- **Budget:** The project described must imply a budget of at least $10,000/year. If they mention 'small budget' or 'just starting out', they are disqualified.
- **Problem:** The user must describe a clear business problem related to customer acquisition, sales automation, or data integration. Vague requests like 'want to learn more' are low quality.

Now, analyze the following lead data:
- **Name:** {lead_name}
- **Email:** {lead_email}
- **Company:** {company_name}
- **Company Size:** {company_size}
- **Title:** {lead_title}
- **Project Description:** {project_description}

Based on the ICP, perform the following tasks:
1.  Score the lead on a scale of 1 to 10, where 10 is a perfect ICP match.
2.  Provide a one-sentence reason for the score.
3.  Determine if the lead is qualified (true/false).

Output your analysis in the following JSON format:
{
  "is_qualified": <boolean>,
  "score": <integer>,
  "reason": "<string>"
}
</user>

This prompt is effective for a few reasons. It gives the AI a specific role (QualBot). It provides a crystal-clear, multi-point ICP. It defines the inputs using placeholders. And most critically, it specifies the exact JSON structure for the output. This is how you turn an LLM from a text generator into a predictable data processing tool.

Is the Free Tier Actually Usable for This?

This is a common question, and the answer is: for some parts, yes, but for the central automation, no. Let’s break down the costs.

My favorite thing here is Tally.so’s free plan. It’s not a gimmick. You get unlimited forms and submissions, plus webhooks. For a solo operator, you might never need to upgrade. I run several projects off their free plan.

The OpenAI API is also incredibly cheap. A single lead qualification with GPT-4o, using the prompt above, will cost a fraction of a cent. You could process a thousand leads for less than $10. The cost of the AI itself is not the bottleneck.

The bottleneck is the orchestrator. My big gripe is with Zapier’s free plan. It gives you 100 tasks per month. A single lead qualification might take 3-4 tasks (Trigger, AI call, Filter, Notification). That means you can process maybe 25 leads a month before it shuts off. Honestly, the free plan is a joke for anything beyond a quick test. You’ll need their “Starter” plan at $29.99/month to get 750 tasks, which is a more realistic starting point. I think it’s overpriced for what it is, but it’s the industry standard and it works.

So, your all-in cost to run this system is effectively the price of your Zapier or Make.com subscription. Budget about $30/month to get started.

Step-by-Step: Plumbing It Together in Zapier

Here’s how you connect the pieces. This assumes you have accounts for your form tool, Zapier, and the OpenAI API.

1. Create the Trigger: In Zapier, create a new Zap. The trigger is “Webhook” or, if there’s a native integration (like for Tally), “New Submission in Tally.” This will give you a URL to paste into your form’s settings. Send a test submission to pull in the data fields.
2. Add the AI Action: Add a new action step and search for OpenAI. Select the “Send Prompt” or “Conversation” event. In the prompt field, paste the entire prompt from the section above. Here’s the important part: you need to replace the placeholders like {lead_name} with the actual data fields from your form trigger. Zapier lets you select these from a dropdown. Set the model to gpt-4o and the temperature to 0.2 for more consistent results.
3. Parse the AI’s Response: The AI will return a text string that contains your JSON. It’s not yet usable by other Zapier steps. You need to parse it. You can try Zapier’s built-in “Formatter” tool, but it can be brittle. My preferred method is to add a step called “Code by Zapier” and use a tiny snippet of Javascript: return JSON.parse(inputData.raw_response);. This will turn the text into structured data (which, yes, is annoying that you have to do it manually).
4. Create a Filter or Path: Now you can use the data. Add a “Filter” step. Set it to only continue if the is_qualified field from the previous step is `true`. This stops junk leads from proceeding any further.
5. Route the Qualified Lead: This is the payoff. Add your final action steps. Send a notification to a specific Slack channel with the lead’s details and the AI’s score/reason. Add the lead to a Google Sheet called “Qualified Leads.” Create a new deal in your CRM like Pipedrive or HubSpot. You can do all three.

That’s the basic flow. Trigger -> AI Qualify -> Parse -> Filter -> Route.

How to Debug When Your Landing AI System Breaks

This system will eventually fail. Not because the AI is sentient, but because you’re dealing with unpredictable user input and APIs. Here are the most common failure points and how to fix them.

• Failure Mode 1: The AI output is inconsistent or not valid JSON.
  Check the history for your Zap. Look at the raw output from the OpenAI step. Sometimes the model will add conversational text like “Here is the JSON you requested:”. This breaks the parser. The fix is to make your prompt even stricter. Add a line like: “Your response must begin with { and end with }. Do not include any other text or formatting.” You can also try enabling JSON Mode in the API call if the tool supports it, which forces the model’s output to be a valid JSON object.
• Failure Mode 2: Good leads are being scored as unqualified.
  Your ICP is either too strict or poorly defined in the prompt. Go back and analyze the leads that were incorrectly rejected. Did a CEO with the title “Chief Visionary Officer” get rejected because your prompt only looks for “CEO” or “Founder”? Your ICP definition needs to be more flexible. Add examples of acceptable and unacceptable inputs directly into the prompt (this is called few-shot prompting) to give the model better context.
• Failure Mode 3: The Zapier “Code” or “Parser” step errors out.
  This almost always goes back to Failure Mode 1. The input to the parser was not what it expected. Dig into the run history. Copy the exact text the AI provided and paste it into an online JSON validator. You’ll likely find a small syntax error, like a trailing comma or a missing quote, that you need to coach the AI in your prompt to avoid making.

Debugging this kind of AI agent framework is an iterative process. You run traffic, find the edge cases that break it, and then refine your prompt to handle them. You won’t get it perfect on the first try.

Building this yourself gives you full control over your lead qualification logic and costs just a few dollars and a monthly subscription. It’s a foundational operator playbook for anyone drowning in form submissions.

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 entire workflow, including the tested prompt and debugging guide, as an AI automation blueprint at deepusecase.com/vault.