I Trained Apollo AI Directly with tawk.to Customer Interactions

Community Discussion
1

I’ll just come out and say it: tawk.to has the best AI integration of any CRM platform, largely because of how open and flexible it is.

I’ve been working on a project for a while that allows me to continuously train an AI model using real customer interaction data, and tawk.to made this possible without fighting the platform.

The system is built around three main steps:

  • Querying data
  • Data analysis
  • Model instruction output

Querying Data

I start by exporting 90 days of customer chat logs. This includes conversations handled by both our existing support bot and human agents. The goal is to capture a realistic mix of interaction types—good, bad, and everything in between.

Each chat is exported as its own file:
ticket-{uuid}.json

The logs follow a simple schema:

{
  "id": "{uuid}",
  "messages": [
    {
      "role": "user",
      "content": "hey i was wondering does it cost extra for shipping"
    },
    {
      "role": "assistant",
      "content": "Sorry, our live chat team seems to be busy with other customers right now! You can click here to open a ticket and we will get back to you shortly! \nhttps://store.com/contact-us/",
      "sender": "AI Helper"
    }
  ]
}

In total, I ended up with ~1,700 files, which turned out to be a goldmine of real-world data.

Data Analysis

I then use another AI model to parse through these logs and generate a report that identifies:

  • What went well in each conversation
  • What went poorly
  • Common failure points
  • Repeated customer questions

From this analysis, the system outputs a CSV containing FAQs, grouped by the most common customer query types.

Model Instruction Output

I built an internal tool that automatically exports this CSV to Google Drive. Any time I want to update the model, I simply rerun the script and the CSV updates in place.

Screenshot 2026-01-11 at 14.10.55
Screenshot 2026-01-11 at 14.10.551880×236 20.4 KB

The document is on an auto-updating schedule, where the tool analyzes the most recent 30 days of logs and appends new questions as they appear.

I’ll occasionally skim the CSV to sanity-check the output, but so far it’s been consistently accurate.

2 Likes
2

This is so awesome! Thanks for sharing.

3

This is a game changer for me! Thanks for sharing.

4

That’s a smart way to approach AI training — using real interaction data instead of assumptions makes a huge difference. Platforms that stay open and flexible really unlock this kind of innovation.

We’re seeing the same thing in sports streaming apps, where analyzing live user interactions and feedback helps improve stream quality, engagement, and real-time features. When the tech doesn’t get in your way, you can iterate and improve much faster.