I Built an AI WhatsApp Lead Qualification Platform: Here’s How It Works

Most businesses don’t have a lead generation problem in isolation.

They have a problem with what happens after the lead arrives.

Someone submits a form, sends a WhatsApp message, clicks an ad or asks about a property.

Then a human has to respond.

They ask the same questions.

They work out whether the person is serious.

They collect contact information.

They understand what the lead wants.

They decide what should happen next.

Kadensio is a product I built around that workflow: using AI inside WhatsApp conversations to collect and structure lead information before the business has to manually handle every interaction.

This isn’t a client case study and I don’t claim revenue or conversion improvements that haven’t been demonstrated. It’s one of my own software projects and a useful example of how I approach building AI products beyond a simple chatbot.

If you’re interested in the development process behind projects like this, start here:

The problem

Imagine a business receives dozens of inquiries.

Not every inquiry is equal.

A real-estate agency might need to know:

What type of property does the buyer want?
Where?
What budget?
Are they buying or renting?
When do they want to move?
Do they need financing?

A hospitality business may need completely different information.

A service business might need:

What service?
What location?
What timeline?
What budget?
What’s the best time to speak?

Humans can collect all of this.

The issue is that doing it manually for every incoming lead is repetitive.

A basic contact form can collect structured information, but forms have their own limitations. They force the business to predict every question in advance and don’t behave like a conversation.

That’s where the idea for Kadensio becomes interesting.

The basic workflow

At a high level:

Lead arrives
WhatsApp conversation begins
AI understands the conversation
Relevant qualification questions are asked
Information is extracted into structured fields
The lead can be scored or categorized
The appropriate next action can happen

That next action could eventually mean:

notify a salesperson
assign an agent
schedule a call
book an appointment
update a CRM
trigger another workflow

The AI conversation isn’t the product by itself.

The product is the system around the conversation.

Why WhatsApp?

For many businesses, WhatsApp is already part of how customers communicate.

That makes it interesting because the goal isn’t necessarily to persuade someone to install another application or learn another interface.

The customer can communicate in a familiar channel.

Behind that conversation, however, the business needs structure.

A salesperson doesn’t want to reread a 30-message conversation every time.

They want something closer to:

Name: Sarah
Interest: 2-bedroom apartment
Area: Eixample
Budget: €650,000
Timeline: 3 months
Financing: Approved
Priority: High

That’s where conversational AI becomes much more useful.

It can turn unstructured conversation into structured business data.

Why not just use a chatbot?

Because “chatbot” describes the interface, not the complete system.

A useful qualification product has to think about:

Conversation state

What has already been asked?

What information is still missing?

Structured extraction

How does natural language become reliable fields?

Business rules

Which questions matter for this company?

Lead status

When is someone qualified?

Escalation

When should a human take over?

Integrations

Where does the resulting data go?

Failure handling

What happens when the AI misunderstands something?

Those are software problems, not merely prompt-writing problems.

Building the conversation layer

The conversation needs to feel natural without becoming aimless.

A rigid form asks:

Question 1
Question 2
Question 3

A completely unconstrained AI conversation can wander.

The useful middle ground is allowing natural language while maintaining a structured objective.

The system knows which information needs to be collected but doesn’t necessarily need to ask every question in exactly the same wording or order.

For example, if the user says:

“I’m looking for a two-bedroom in Eixample around €600k and I’d like to buy before January.”

The system shouldn’t respond:

“What area are you interested in?”

It already knows.

The conversation state needs to reflect that.

Turning conversation into data

This is one of the parts I find most interesting.

Humans communicate messily.

A lead might say:

“Probably somewhere around 500 to 600, but I could maybe go higher for the right place.”

The application needs to preserve the meaning while turning it into something the business can use.

That might become structured information such as:

Budget minimum: €500,000
Budget target: €600,000
Flexibility: Yes

That structured data can then drive other software.

Search
Scoring
Routing
CRM updates
Analytics
Follow-up

The conversation becomes an input layer for a larger business system.

InteractiveFrom a WhatsApp message to structured data

Two example lead messages and the fields extracted from them.

Lead message: “I’m looking for a two-bedroom in Eixample around €600k and I’d like to buy before January.”

Field	Extracted value
Interest	2-bedroom
Area	Eixample
Budget	€600,000
Timeline	Before January

The area is already known, so the system does not ask for it. Next it can ask about financing.

Lead message: “Probably somewhere around 500 to 600, but I could maybe go higher for the right place.”

Field	Extracted value
Budget minimum	€500,000
Budget target	€600,000
Flexibility	Yes

The meaning is kept, and the business gets fields it can search, score and route on.

Both messages are the examples from this article. Hover or focus a phrase or a field to see the link.

Qualification isn’t the same for every business

A major design requirement is that different businesses care about different things.

A real-estate agency may care heavily about budget, location and timeline.

A hotel may care about dates, group size and special requirements.

A service company may care about project scope and budget.

That means the qualification logic needs to be configurable rather than hard-coded around one universal sales conversation.

This is also why custom software can be valuable.

The closer software gets to the actual business process, the more useful it can become.

InteractiveWhat each business needs to know

Qualification questions differ for a real-estate agency, a hotel and a service company.

Business	What to find out	Emphasis
Real estate	Property type; Location; Budget; Buying or renting; When they want to move; Need financing?	Budget, location and timeline matter most.
Hospitality	Dates; Group size; Special requirements	Dates, group size and special requirements.
Services	Which service; Location; Timeline; Budget; Best time to speak	Project scope and budget.

The question sets are the examples from this article. Qualification logic has to be configurable per business.

The dashboard matters too

The customer sees WhatsApp.

The business needs something different.

It needs visibility.

That can include:

incoming leads
qualification status
conversation history
extracted information
lead scores
next actions
appointments
assigned team members

This is a recurring pattern in software development.

The interface the customer sees may be only a small part of the complete product.

A lot of engineering exists behind it.

Where AI helps

AI is particularly useful when the input is unstructured.

People don’t speak in database fields.

They say things like:

“I’m just looking for now, probably sometime next summer, somewhere near the beach but not too touristy.”

Traditional software struggles with that unless the user fills in predefined fields.

A language model can interpret it.

But the AI output then needs to become reliable enough for the surrounding software to use.

That’s where structured outputs, validation, application logic and fallback behavior become important.

Where AI should NOT be trusted blindly

Adding AI creates uncertainty.

Models can misunderstand.

They can return unexpected output.

They can make assumptions.

So I don’t think good AI software should simply send everything to a model and hope for the best.

The surrounding application needs constraints.

Important actions may need deterministic rules.

Data should be validated.

Sensitive or high-impact actions may require human confirmation.

The model should do the work it’s good at while conventional software handles the things conventional software is better at.

What makes a product like this technically interesting?

It’s the combination
Messaging
AI

Conversation state.

Structured data.

Business rules.

Dashboards.

Integrations.

Potential scheduling.

Potential CRM synchronization.

None of those individually is the entire product.

The value comes from connecting them into one workflow.

That’s also why I enjoy building this kind of software more than simply producing marketing websites.

The software actually does something.

Could the same architecture work outside real estate?

Yes.

The underlying pattern is broader:

Customer starts conversation.
System understands intent.
System gathers missing information.
Information becomes structured data.
Business workflow continues.

That can apply to:

hospitality
property management
recruitment
home services
professional services
automotive businesses
event companies
other lead-driven businesses

The questions and workflows change.

The underlying engineering pattern remains similar.

What I learned building it

One lesson is that the AI itself is rarely the whole product.

It’s easy to build a demo where an LLM responds intelligently.

The harder and more valuable work is everything surrounding it:

How does the system know what the business wants?
How does it retain state?
How does it turn conversation into useful data?
How does a human take over?
Where does the data go?
What happens next?

That’s where an AI demo becomes an application.

How long does software like this take to build?

That depends entirely on what is included.

A proof of concept showing AI qualification could be built quickly.

A complete multi-business platform with configurable workflows, messaging infrastructure, analytics, billing, integrations, permissions and production reliability is a substantially larger product.

That’s why I scope software around workflows rather than simply counting pages.

I’ve explained timelines here:

And pricing here:

Coming soonHow much does it cost to build an MVP in 2026?

Would I build every AI product this way?

No.

The architecture should follow the problem.

Some products need conversational AI.

Others need document processing.

Some need retrieval
Some need classification
Some barely need AI at all

I don’t think adding AI automatically makes software better.

It makes sense when it solves a part of the workflow that traditional deterministic software handles poorly.

What Kadensio demonstrates

For me, Kadensio demonstrates the kind of project I like building:

A real workflow.

Several systems interacting.

AI used for a specific purpose.