Thousands of WhatsApp Chats, a Two-Person Team: Letting the Owner Monitor Without Reading Everything

This retail business's marketing team is just two people, but they handle every WhatsApp conversation with customers: product questions, complaints, and closed deals. The owner is busy running other operations, but still wants to know how the team is doing. The problem is, reading every raw chat every night would be far too much information to take in.

I built a system that stores every one of those WhatsApp chats on infrastructure built to handle high volume, then every night at 9pm, AI puts together an automatic recap for the owner: who complained, who closed a deal, and who still needs a follow-up.

The problem: the old system wasn't future-proof

The system originally ran on Google Apps Script, storing chat history in Sheets, the fastest way to get started. Once incoming chats grew into the hundreds and then thousands, this stopped being a question of "is it enough," it became a question of not being future-proof. Sheets itself has a cell limit and a script execution limit, and marketing chat is exactly the kind of data that keeps piling up every single day.

The second problem was about the owner: he wanted to monitor the team's work, but just two people handling WhatsApp could generate hundreds of chats a day. There was no way he could read all of it every night on top of everything else he had to run.

The approach: store every chat somewhere built to scale, and let AI summarize it

1. 1Chats in and outThrough Wablas, the WhatsApp gateway the team uses
2. 2Hooked into a Cloudflare WorkerEvery chat gets processed the moment it comes in
3. 3Text to D1, media to R2Stored cleanly, built to handle high volume
4. 4Automatic recap at 9pmAI reads that day's chats and summarizes them

Chats go in and out through Wablas, hooked into a Cloudflare Worker that acts as the brain. Conversation text is stored in Cloudflare D1, and media (product photos, transfer receipts, and the like) is stored separately in Cloudflare R2.

What I built

An automatic recap, not raw chat logs

Every night at 9pm, AI reads through that entire day's chats from D1, then puts together a recap: who complained, who closed a deal, who's still just asking questions, and who needs a follow-up. The result is delivered as an artifact the owner can open anytime, and if anything needs more context, the owner can just ask the AI directly.

This is different from Kirana, another AI assistant I built earlier. Kirana replies to chats directly, while here the AI never touches the customer conversation at all. The two marketing team members still handle every reply to customers themselves, the AI's only job is putting together the summary for the owner.

A gap that showed up later: phone calls

After running for a while, a pattern showed up: not every important conversation happens on WhatsApp. Sometimes a customer or the team just picks up the phone instead, and a phone conversation has no way of being captured by the system, since there's no text to log.

The fix wasn't more technology, it was a simple SOP: after every phone call with a customer, the team has to send a WhatsApp message confirming a summary of what was just discussed. That way, a phone conversation still leaves a text trail the system can capture, and it still makes it into that night's recap.

Results

Time to build, from the old system to the new one

1.5 months

When the automatic recap goes out to the owner every night

9 PM

The marketing team handling every customer chat

2 people

Wrapping up

What's interesting about this project is that the most important gap wasn't a technology problem at all. The infrastructure was already built to handle high volume, the AI could already summarize automatically, but there was still a gap in the one channel that never went through text. The fix wasn't new technology either, just a simple SOP bridging the people to the system that was already built.