I Built a Customer Service AI Using Google NotebookLM — SME Owners Should Take Note

Small and Medium Enterprise owners, take note.

Everyone is talking about using AI for customer service. The obvious approach seems simple: connect your website to ChatGPT, Gemini, Claude, DeepSeek or another Large Language Model (LLM), and let it answer your customers.

But there are problems.

A general-purpose LLM may be incredibly intelligent, but it doesn't know your company. It doesn't necessarily know your latest warranty policy, product specifications, prices, service procedures or return rules.

Worse, when it doesn't know something, it may still give you a very convincing answer.

That's the famous hallucination problem.

If a customer asks:

“My coffee machine broke after eight months. Can I get a refund?”

I don't want an AI telling me what a typical warranty might be.

I want it to tell me exactly what my company's warranty says.

For customer service, a confidently delivered wrong answer can be worse than no answer at all.

And then there is cost.

Commercial LLM, Chinese LLM or Open Source — Someone Still Pays

You can use commercial APIs from OpenAI, Google, Anthropic and others. They are easy to use, but generally you pay according to usage.

Chinese models have made the economics even more interesting. Models from DeepSeek and Alibaba's Qwen, for example, provide increasingly capable alternatives, including models that can be deployed on your own infrastructure.

There are also many excellent open models available today. But open does not mean zero cost.

If you run an LLM yourself, you still need somewhere to run it:GPU or CPU servers, memory, storage, electricity, networking, security, monitoring and somebody to maintain it. You can rent all of this from the cloud, but then you pay the cloud provider. You can buy the hardware yourself, but then you have to operate it.

The cost simply moves from API tokens to infrastructure.

And after all that, your LLM still doesn't automatically know your company's warranty policy.

You may then need to build a RAG system—document processing, embeddings, retrieval, a vector database and all the infrastructure around it.

That can be absolutely appropriate for a serious enterprise application.

But I wondered:

Does an SME really need to build all this just to experiment with AI customer service?

That led me to Google NotebookLM.

My OrdinaryJack Coffee Shop Experiment

I created a fictitious company called OrdinaryJack Coffee Shop.

I gave it a simple warranty policy:

• All coffee machines have a 12-month warranty.

• Within 30 days, customers can request a refund or replacement.

• From day 31 through month 12, customers receive free repair or replacement, but normally no refund.

• After 12 months, the standard warranty expires.

• Refunds above US$500 require human-manager approval.

https://www.youtube.com/watch?v=4O2TcNyz2BQ

I uploaded this document into Google NotebookLM.

Then I built a simple customer-service website in front of it.

Now the customer can ask:

“My machine broke after eight months. Can I get a refund?”

Instead of asking a general LLM to guess, my application asks NotebookLM.

NotebookLM has the actual OrdinaryJack warranty document.

That is important because NotebookLM is designed around source grounding: its answers can be grounded in the sources you provide rather than relying only on the model's general knowledge.

I Didn't Train an AI

This is perhaps the most important point.

• I didn't train anything.

• I didn't fine-tune Gemini.

• I didn't train my own LLM.

• I didn't buy a GPU.

• I didn't build a vector database.

• I didn't create an embedding pipeline.

• I simply gave NotebookLM the information that I wanted the AI to know.

• This distinction is important:

Knowledge is not training.

Businesses often say, “I want to train an AI using my company data.”

Usually, that isn't really what they need.

Google and other AI companies have already trained enormously sophisticated models.

What your company needs to provide is something different:

your knowledge.

Think of it this way:

GOOGLE
provides the intelligence

        +

YOUR COMPANY
provides the knowledge

        +

YOUR APPLICATION
connects the customer

        ↓

YOUR CUSTOMER SERVICE AI

SME Owners: Look at the Knowledge You Already Have

Think about what already exists inside your company:

• Product manuals.

• Warranty policies.

• Price lists.

• FAQs.

• Return procedures.

• Installation instructions.

• Technical specifications.

• Troubleshooting guides.

• Training materials.

• Service procedures.

Your company may already have hundreds or thousands of pages of useful knowledge.

And your employees may be answering questions from those documents every day.

So perhaps the question isn't:

“How do I build an AI?”

It is:

“How do I connect existing AI to the knowledge my company already owns?”

That is a much easier question.

My Architecture Is Almost Embarrassingly Simple

My experiment basically looks like this:

CUSTOMER
    ↓
OrdinaryJack Website
    ↓
Small Python Program
    ↓
GOOGLE NOTEBOOKLM
    ↓
COMPANY KNOWLEDGE
    ├── Warranty Policy
    ├── Product Manuals
    ├── FAQs
    └── Service Procedures

The website doesn't know anything about coffee machines.

My Python program doesn't know anything about warranties.

NotebookLM provides the knowledge engine behind the application. I rely on a "hack" as provided by https://github.com/teng-lin/notebooklm-py?utm_source=chatgpt.com

I Even Show What the AI Knows

I deliberately added a box to my website called:

WHAT I KNOW

It displays the documents currently available to the AI.

Customers can even click VIEW and see the actual information that has been ingested.

I did this because I wanted to demonstrate something fundamental:

Intelligence and knowledge are different things.

The foundation model provides intelligence.

Your company provides knowledge.

Updating the AI Is Surprisingly Easy

Suppose OrdinaryJack launches a new coffee machine.

I don't retrain the AI.

I don't change my Python program.

I simply add the new product manual to NotebookLM.

Refresh my customer-service website, and the new knowledge appears.

[INSERT VIDEO: ADD DOCUMENT TO NOTEBOOKLM → REFRESH WEBSITE]

That suggests an interesting business model:

IT builds the pipe. The business maintains the knowledge.

The product team maintains product information.

The service team maintains service information.

The warranty team maintains warranty information.

Your programmer doesn't need to become the company's product expert.

Now Think About the Cost

This is where SME owners should pay attention.

If I use a commercial LLM API, I normally have API usage costs.

If I self-host an open model such as DeepSeek, Qwen or another model, I need computing infrastructure.

If I build a conventional RAG system, I also need to build and maintain the retrieval infrastructure.

For this experiment, I am instead leveraging the AI and source-grounding capabilities already available through Google NotebookLM.

My application is not separately paying a commercial LLM API token charge for every customer question.

That does not mean NotebookLM is “free AI.” Google has its own plans, usage limits and policies, which can change.

But for an SME wanting to experiment, this is potentially a very inexpensive way to find out whether the idea is useful before building a much larger AI infrastructure.

The important question isn't:

“Which AI is free?”

It is:

“What is the simplest and lowest-cost architecture that solves my business problem?”

And I Can Leverage the Rest of NotebookLM

This is another reason I find the experiment interesting.

Once company knowledge is inside NotebookLM, it isn't useful only for the customer chatbot.

The same knowledge can potentially help employees research products, understand documents, summarize information and generate other NotebookLM outputs.

So I haven't created a knowledge system that serves only one chatbot.

I am leveraging a much broader AI knowledge environment that Google is already building.

For an SME, that is powerful.

But Knowing Is Not the Same as Doing

My OrdinaryJack AI can say:

“According to our warranty policy, your machine qualifies for a replacement.”

But it hasn't actually replaced anything.

This is the next important distinction:

Knowing is not the same as doing.

NotebookLM gives my application knowledge.

The next step could be connecting it to tools such as email, CRM, inventory, order management, APIs, Zapier or MCP.

Then the system might eventually say:

“Your machine qualifies for replacement. I verified your purchase, created a service ticket and sent you confirmation.”

Now we have crossed an important boundary:

A chatbot answers. An agent can act.

But Don't Start With an AI Agent

Everyone wants an AI agent today.

I wouldn't start there.

Start with the 20 questions your employees answer repeatedly.

Find the documents containing the answers.

Put them into your knowledge system.

Build a simple chatbot.

See whether customers find it useful.

Then add one action.

Send an email.

Create a service ticket.

Check an order.

Check inventory.

Eventually allow carefully controlled transactions, with human approval where appropriate.

Build the staircase. Don't jump to the top floor.

Intelligence, Knowledge and Action Are Different Things

This is how I increasingly think about business AI:

INTELLIGENCE
Foundation models
        +
KNOWLEDGE
Company documents and data
        +
INSTRUCTIONS
Company policies and rules
        +
TOOLS
APIs / Zapier / MCP
        +
MEMORY
CRM / databases
        +
AUTHORITY
What the AI is allowed to do
        +
HUMAN OVERSIGHT
Where people remain in control

Most SMEs don't need to build the intelligence.

They need to organize their knowledge, define the rules, connect the right tools and decide how much authority the AI should have.

And today, getting started may be much easier — and much less expensive — than you think.

Disclaimer

This project was created solely for educational, demonstration and experimental purposes. It is not a production customer-service product. The demonstration uses an unofficial open-source interface to Google NotebookLM rather than an official Google-supported customer-service API. Any commercial deployment should independently assess security, privacy, reliability, regulatory requirements, service terms, usage limits and appropriate human oversight.