From Voice Note to Quote: Building AI-Powered Workflows

A voice note can contain enough information to prepare a quote.

The problem is that a voice note is not structured data.

It first has to be understood.

Let’s take an example:

“At Dupont’s, replace the 200-litre water heater, allow two hours of labour and call the customer back to confirm access to the room.”

For a human, the information is fairly clear.

For business management software, it has to be transformed.

Step 1: transcribe

The first step is to turn the voice into text.

Audio
↓
Speech-to-text
↓
Text

This step can be handled by a speech recognition model.

But transcription alone is not enough.

Step 2: extract the information

The text then needs to be analysed.

For example, we look for:

{
  "customer": "Dupont",
  "product": "Water heater 200L",
  "quantity": 1,
  "labor_hours": 2,
  "follow_up": "Confirm access to the room"
}

This is where an LLM becomes interesting.

It can turn free text into a usable structure.

Step 3: apply the business logic

This step is often forgotten in AI demos.

The LLM should not calculate the final price on its own.

The system must apply its own rules.

For example:

Product identified
↓
Catalogue lookup
↓
Catalogue price
↓
Pricing rules
↓
Labour
↓
Taxes
↓
Total

AI understands.

The business engine calculates.

This separation is essential.

Step 4: generate the quote

Once the data has been validated, the system can prepare the quote.

But I would keep a human approval step:

Voice note
↓
AI
↓
Structured data
↓
Business engine
↓
Draft quote
↓
Approval
↓
Sending

This architecture is far more robust than a system in which AI directly generates a final document.

Why this approach is interesting

The gain doesn’t only come from speed.

It comes from eliminating double data entry.

The tradesperson can speak naturally.

The software then turns that information into structured data.

So we move from:

interface → form → fields → data entry

to:

conversation → understanding → data → action

That is a major evolution in human-computer interaction.

The developer’s role

Building this kind of system is not simply a matter of plugging in an LLM API.

You have to design:

• data schemas;
• validations;
• business rules;
• permissions;
• recovery mechanisms;
• logs;
• human confirmations;
• tool calls;
• checks before any action.

AI is one building block.

The architecture does the rest.

A simple workflow can already deliver a lot of value

There is no need to build a fully autonomous agent.

A first workflow can simply do:

🎙️ Voice note
        ↓
📝 Transcription
        ↓
🧠 Extraction
        ↓
🔎 Customer lookup
        ↓
📦 Product lookup
        ↓
🧮 Business calculation
        ↓
📄 Draft quote
        ↓
👤 Human approval

That is already a concrete automation.

And above all, it solves a real problem.

The goal is not to put AI everywhere

This is probably the most important lesson.

You shouldn’t start with:

“Where can we put an LLM?”

You should start with:

“Which task wastes the user’s time?”

Then see whether AI can genuinely reduce that friction.

In business software, that is often where the best use cases appear.

Further reading :

• AI Won’t Replace Tradespeople, Just Their Paperwork
• AI Automation with Python: What Actually Works in Production