A bot that answers a store’s WhatsApp and Instagram chats using the store’s real stock and prices, and hands the conversation to a person when it should. Today it handles around 6,000 conversations a month.
Sales bots wired into your systems
- Channels
- WhatsApp and Instagram
- Connected to
- the store’s catalog and stock, the daily exchange rate, and the trade-in price sheet
- Volume
- about 6,000 conversations and 23,000 incoming messages a month
- In production since
- July 2026
- Model cost
- under two US cents per reply
- Runs on
- Chatwoot plus Chatty, the platform I built to host these bots
Numbers from August 2026, for a tech retailer with more than 90,000 Instagram followers.
The problem
The store gets hundreds of messages a day. People ask for a price, whether a model is in stock, how much a used phone is worth as a trade-in, and what the payment plans are. Most of those questions have an exact answer somewhere in the store’s systems.
With a small team, the messages that arrive at night or on weekends wait until the next day, and some never get an answer. Every one of those is a sale that may go somewhere else.
A regular chatbot does not fix this. It does not know what is in stock, and when it does not know, it makes something up. A wrong price in a chat is worse than no answer at all.
How it works
Every message goes through the same four steps.
-
The message arrives
If the customer sends several messages in a row, the bot waits a moment and answers them as one. Photos and voice notes are turned into text first.
-
The bot looks things up
Stock, prices, trade-in values, and payment plans come from the store’s own data, updated every half hour. The bot never answers from memory.
-
The reply is checked before it goes out
Twenty-eight rules review the draft. If it quotes a price that did not come from the data, or claims stock that is not there, the draft is thrown away and written again.
-
A person steps in when needed
Repairs, payments, and reservations go to a human agent, with a summary of the chat already written. The bot stops talking the moment it hands off.
- Hi! How much is the 15 Pro, 256 GB?
- It’s $1,200,000.
- iPhone 15 Pro 256 GB, sealed, in stock. USD 980, or 1,533,700 pesos at today’s rate. Want the payment plans?
What it does not do
- It does not make up prices or stock. If it did not look it up, it does not say it.
- It does not do math in its head. Installments, trade-ins, and currency conversions come from a calculator, not from the model.
- It does not give opinions on repairs or warranties. A person handles those.
- It never confirms a payment as received.
- It does not keep talking after handing the chat to a person.
- It introduces itself as an automated assistant, and you can always ask for a person.
What the store gets
Answering chats is half of it. The other half is what the owner learns from them.
- A daily report, by chat: what people asked for and the store did not have, who is on the waiting list, reservations, and which ad each customer came from.
- A plain-language note for every improvement to the bot, written for the store, not for engineers.
- A place to report when the bot gets something wrong, and to see when it has been fixed.
Questions people ask
- How long does it take?
- A few weeks for a first version connected to your data. After that the bot improves with use: every mistake you report becomes a test it can never fail again.
- What do I need on my side?
- A place where your prices and stock already live (a spreadsheet, a system, a website) and a WhatsApp Business number or an Instagram account. I take care of the rest.
- What happens when the bot does not know?
- It says so and passes the chat to a person. I would rather ship a bot that hands off than one that guesses.
- How much does it cost?
- It depends on the channels and the systems to connect. Write to me with what you have and I will give you a number.
- What about my customers’ data?
- It stays in your CRM. The bot reads what it needs to answer, and nothing is shared with anyone else.
How it is built
For the engineers in the room, I am writing up the inside of this system one piece at a time.
- Inside Chatty
The map of the whole system, kept up to date.
- You do not stop hallucinations with the prompt
Data first, then tools, then hard rules. The prompt comes last.
- A rule that throws the reply away
How a draft gets rejected and rewritten inside the same turn.
- The tests I actually run
Scripted conversations as a gate, simulated customers as observation.
- When the bot hands off to a person
The model asks, a rule forces it, or there was no reply. Then the bot goes quiet.
- The maintenance loop is a set of skills
Evidence to open an issue, a failing test to close it, and a note in the store’s language.
- The bot captures, the MCP consumes
Questions from the store’s own Claude, a daily report, and why reporting is not the bot’s job.
Want one for your business?
Tell me what your customers ask and where your prices and stock live. No forms, no commitment: write me and we will take a look.
or email me at pedro@pedrosimon.dev
Pedro Simon, Rosario, Argentina.