A small-business owner recently told me something I now hear more and more often: "We use ChatGPT, but it doesn't run the company."
He was right.
His team knew how to open ChatGPT. They knew how to ask it to rewrite an email, summarize a note, sometimes prepare a sales argument. But when it came to creating a real quote, with the right products, the right dimensions, the right discounts, the right customer details, the right CRM update and an internal validation step, everything became manual again.
This case matters because it looks like many companies with 5 to 50 employees. They already have tools. They already have habits. They may have a CRM, quote software, a sales assistant, two sales reps on the road, technicians visiting customers. And in the middle of all that, they wonder how AI can become more than a chat window in a browser.
Here is how we built a concrete AI agent for a B2B joinery and installation company, with field teams, quotes to produce quickly, and one simple need: stop retyping tomorrow what someone had already explained yesterday.
The starting point: a process held together by habit
The company sells and installs joinery products for B2B and residential customers. In the field, two sales reps visit prospects. Installers then handle fitting, adjustments, repairs, and sometimes price additional work.
On paper, the process looked normal. A sales rep visits the customer, captures the need, prepares a proposal. An installer notices extra work, writes down what should be added, then the office turns it into a clean quote.
In practice, each person had their own method.
Sales reps used a tablet with an old tool, something close to a homegrown Salesforce setup. Some information was entered, some stayed in notes, and some was passed verbally. The secretary or sales assistant had to reopen the file the next day, find what was missing, reconstruct the context, then produce a proper quote.
For technicians, it was even more uneven. Sometimes paper. Sometimes handwritten notes. Sometimes a message sent to the office. Sometimes a prefilled Word document on a tablet, depending on the person, the urgency and the place.
No one was being careless. That matters. The team was doing what it could with the tools it had, with field constraints, incoming calls, customers calling back and appointments piling up. The problem was not a lack of goodwill. The problem was a process that asked humans to carry too much information in their heads.
And the more a process depends on memory, the more it creates silent errors.
The hidden cost of manually rebuilt quotes
A quote rebuilt by hand does not only cost ten minutes of typing.
It costs time. The customer waits until the next day, sometimes longer. The sales rep has to remember what was promised. The sales assistant has to interpret a note that was not written for her. The technician has to answer a question while already working on another job.
It also costs quality. A missing dimension, an unclear installation type, an unconfirmed discount, an incomplete address, a customer status entered incorrectly. None of these issues looks dramatic on its own. But they create back-and-forth, corrections, frustration and commercial risk.
The interesting part is that the company already had useful building blocks. It had quote software. It had a CRM. It had products and prices. It had a sales assistant who knew the operating rules very well. It had field teams that could describe a situation clearly when given the right channel.
It was not missing "AI". It was missing a system connecting those building blocks.
That is often where the real AI automation work sits. Not in a magic prompt. In turning a free-form, human, imperfect request into a reliable action in production tools.
Before the agent: clean up just enough
The first temptation on this kind of project is to rebuild everything. New CRM, new catalog, new quoting tool, new end-to-end process. That is rarely the right move for a small business that still needs to sell and install while the system improves.
So we did the opposite: clean up just enough for the agent to work properly.
With the company, we reviewed the products, options, prices and discount rules available in their existing environment. The goal was not to create the perfect catalog. The goal was to make the data required for quotes reliable enough to be used by an automated system.
This part is less spectacular than the voice agent, but it is decisive. An AI agent connected to messy data produces mistakes faster than a human. It does not magically infer the right business structure. It applies what it receives. If products are ambiguous, if prices live in three places, if commercial exceptions exist only in one person's head, you have to clarify before automating.
I also spent time with the sales assistant, because she held the operational truth. Not the theoretical process, but the real rules: what must always be asked, what can wait, what blocks a quote from being sent, what must be checked manually, what can be prepared automatically.
That is one lesson from this project: an AI agent is not built only with the business owner. It is built with the people who live with the exceptions.
The channel choice: not an app, a voice note
Once the business frame was clear, one practical question remained: how would sales reps and technicians send information to the agent?
We could have built a web interface. A clean form, with fields, dropdowns, authentication and history. Technically, that was possible.
But in the field, it was not the right gesture.
A sales rep leaving an appointment does not want to fill out an 18-field form in the car. An installer finishing a job does not want to find the right screen while the customer is still talking. But everyone knows how to send a voice note.
So we chose a Telegram interface, aligned with the company's internal habits and simple enough to be adopted. The sales rep or technician can send a text message or a voice note. They explain the situation as they would to the sales assistant:
For Mr. Smith, we need two custom windows, renovation installation, dimensions taken on site, 8% discount, preferred delivery in June. He is a residential customer, address to confirm, and he also wants a mosquito screen option if the price stays reasonable.
This message is not a form. It is not perfectly structured. It looks like a real field request.
The agent's job is precisely to turn that natural language into usable data.
What the agent does, step by step
When a message arrives in Telegram, the system first retrieves the content. If it is a voice note, it is transcribed. If it is text, it is analyzed directly. Then the agent extracts useful elements: customer, customer type, requested products, dimensions, options, discount, installation constraints, contact details, urgency, internal notes.
Then it compares this information with the rules defined with the company.
If a blocking piece of data is missing, the agent does not create a weak quote. It replies in Telegram a few seconds later, interactively:
I am missing the customer's full address and the installation type. Is this renovation installation or full removal?
Or:
You mention a discount, but not the rate. What rate should I apply?
Or:
I cannot finalize the quote without knowing whether the customer is an individual or a company. Can you confirm?
This changes a lot. Before, the sales assistant discovered gaps the next day. Now, the question goes back immediately to the person who still has the context in mind.
When all mandatory information is present, the agent acts. It creates the quote in the business tool with the right lines. It fills or updates the CRM. It adds useful context so the office understands where the request comes from. Then it notifies the sales assistant through the internal channel used by the company, Telegram or WhatsApp depending on the case.
The quote is not sent directly to the customer. That was an intentional choice.
The company wanted to keep a human double-check before sending. For work quotes, that is healthy. AI prepares, structures, checks missing information and fills the tools. A human validates the final commercial commitment. For a small business, this compromise is often more robust than full automatic sending.
Why this is not "just a ChatGPT prompt"
From the outside, you could summarize the project as: "an agent that turns a voice note into a quote." True, but incomplete.
The value does not come from one isolated prompt. It comes from the whole system:
- Products and prices cleaned up
- Business rules made explicit with the sales assistant
- An input channel adapted to field teams
- Active checks for missing information
- Real connection to the CRM and quote software
- Internal notification for human validation
- A test period with actual users
ChatGPT alone can help draft a quote. By default, it does not know which products are sold, which prices are valid, which discounts are allowed, which CRM field must be filled, or who must review before sending.
That is the difference between using AI and integrating AI.
Using AI means opening a chat and copy-pasting the result elsewhere. Integrating AI means connecting it to the tools, rules and responsibilities of the company.
That is what I build in AI agent projects for businesses. Not pleasant demos. Systems that enter daily work.
Adoption: the real test
The biggest risk was not technical. The biggest risk was that no one would use it.
In many automation projects, the tool is designed from the office, then pushed onto the field. It makes sense on a diagram, but it is too heavy in real life. The result: the team keeps using WhatsApp, paper, voice notes, and the new system becomes one more tool to maintain.
Here, I had to adapt to each stakeholder. The owner wanted to understand the gain. The sales assistant wanted to keep control over quality. Sales reps wanted it not to slow down appointments. Technicians wanted to speak normally, without learning new software.
So we presented the agent as an assistant for them, not as a monitoring tool. It is not there to check whether the sales rep did the job correctly. It is there so the rep does not have to explain the same thing three times. It is not there to replace the sales assistant. It is there to give her a quote that is already prepared, with information placed in the right system.
That nuance matters. Automation that feels like it takes work away from teams creates resistance. Automation that removes a painful task creates adoption.
After training, we kept a test period. The team played with the agent, sent real cases, spotted ambiguous wording and asked for adjustments. Since the beginning of this week, the system has been running in production.
The test period mattered because it revealed things no scoping workshop can see. Real messages are shorter than expected. Voice notes contain hesitations. Technicians do not always name products like the catalog. Sales reps sometimes use shortcuts only they understand.
A good agent must absorb that reality, not ask humans to become walking forms.
What the company actually gains
I will not invent a productivity percentage. The system has just entered production, and usage needs to be observed over time.
But the expected gains are already clear.
The sales assistant receives fewer incomplete files. Questions go back to the right person faster. Quotes are prepared in the right tools without full retyping. The CRM is better maintained because updating it is part of the flow. Sales reps and technicians keep a simple channel, suited to their day-to-day work. The owner gets a more consistent process without brutally changing the whole organization.
The most important gain may not even be time. It is reliability.
As a company grows, small process variations become expensive. Everyone has their method, then exceptions pile up, then the office compensates. The agent imposes structure without imposing friction. It lets teams speak naturally, but it forces the system to ask for missing information before creating the quote.
That is where AI becomes profitable: when it reduces errors, delays and invisible rework.
What this project taught me
This project reminded me of a simple thing: the best automations do not start with a tool. They start with a tour of real usage.
You have to listen to the business owner, but also to the person correcting quotes. You have to understand the CRM, but also the paper notes. You have to look at the catalog, but also at the words used in the field. You have to get the team talking without making them feel judged.
It also confirmed that training is part of the product. Delivering an agent without explaining how to use it, when to trust it, when to verify, and how to phrase a request is delivering half the system.
Finally, it reminded me that small businesses do not need an abstract "AI transformation". They need one use case that works. A quote. A follow-up. An invoice. Lead qualification. Call analysis. A handoff between two tools.
I had already seen this logic on a pipeline for AI sales call analysis with n8n, Whisper and GPT-4o, then on Pennylane invoicing automation with n8n. This new case confirms it: useful AI is rarely isolated. It lives inside a workflow.
Takeaways if you run a small business
If you feel like "AI is everywhere" but it has not changed your daily operations yet, it does not necessarily mean you are late. It may mean you tested AI at the wrong level.
ChatGPT in the browser is a good individual tool. But to run a business process, you need to answer different questions:
- What data can the agent read?
- What actions is it allowed to take?
- What information must it check before acting?
- Who validates sensitive decisions?
- In which tool should the final action appear?
- Which channel will the team actually use?
When these questions are handled, AI becomes much less vague.
For this joinery company, the answer was not a huge platform, a full rebuild, or a three-day theoretical training session. It was an agent available from Telegram, connected to the existing tools, able to ask the right questions and prepare a usable quote.
It is not as spectacular as a LinkedIn demo. It is more useful.
If you want to identify the same kind of use case in your company, I can help you scope, build and deploy this kind of system. The closest entry point is my Automation & Workflows page, and if your topic involves assistants that act inside your tools, the AI Agents page explains the approach.
First 30-minute conversation to understand your context: book a slot.
FAQ
How do you automate quote creation with an AI agent?
You first structure the products, business rules and mandatory information. The agent can then collect requests from a simple channel like Telegram, check missing data, create the quote in the business tool, update the CRM and notify the team for human validation.
Why use Telegram or WhatsApp instead of a dedicated app?
For field teams, the best channel is often the one they already use. A Telegram or WhatsApp voice note takes less effort than a full form, especially between appointments or on site. The agent then turns the free-form message into structured data.
Can an AI agent automatically send a quote to the customer?
Technically yes, but it is not always the right choice. In this project, the company kept a manual double-check by the sales assistant before sending. That is often the right compromise: AI prepares, checks and fills; the human validates the commercial commitment.
What are the limits of an AI agent for commercial quotes?
The agent depends on the quality of available data, defined business rules and tool access. If prices, products or commercial exceptions are unclear, they must be clarified first. AI does not replace that framing work; it executes it quickly once it is clean.
Questions? Feel free to reach out on LinkedIn or use the contact form.