Automated Agents: A Book About Building Chatbots

When I joined a tiny startup with big ambitions, I had no idea we’d be building an eCommerce chatbot from scratch. It was my first real-world dive into AI, and I was starting with zero experience beyond a few online courses. As the first engineer on the team and the third employee overall, I had a simple goal: figure it out and make it work.

Back then, practical information on building chatbots was almost nonexistent. Everything I found was either too basic or so focused on research that it wasn’t useful in the real world. Fast forward to today, and nearly every chatbot tutorial revolves around plugging into ChatGPT or another Large Language Model (LLM). While those tools are powerful, they’re often overkill for solving the specific, focused problems most businesses face.

Customer service is a niche. During the chatbot hype, a lot of companies were racing to create agents that could hold a general conversation. Demo after demo featured people asking chatbots random questions like, “How tall is the Statue of Liberty?” or “Who’s the president of Russia?” Some threw in jokes or tried to trip the bot up with tricky questions. It was impressive, but it missed the point entirely.

Most customers aren’t looking to chit-chat. They come to customer service with a specific issue related to their current or future purchase. Maybe they want to know about a product, their order status, or your return policy.

In my experience, a chatbot often ends up being a Band-Aid for bad user experience (UX). In an ideal world, customers wouldn’t need to reach out to support at all. They’d find the solution on the website by themselves. But the reality is different. For example, if a customer says they can’t find their order, a chatbot that provides their order information is helpful. But wouldn’t it be better if the order was easy to find on the site in the first place? Maybe the UI is confusing or has too many steps. If the issue is common, the best solution is to fix the UX.

From an AI perspective, you don’t need an LLM to solve these problems. A good classifier and entity recognition system can take you far, sometimes all the way. These solutions are easier to build, require less computing power, and are cheaper to run.

I started compiling my thoughts on all this in a series of articles. But the more I wrote, the more it started to feel like a book. After posting about it online, I got a lot of interest from people wanting to follow along. So, I decided to write the book publicly.

Right now, the book is split into three parts. Part one is an overview that explains key concepts and ideas in an accessible way. Great for non-technical readers like product managers who want to understand how a chatbot works and how to implement one. Part two dives deeper into the technical architecture, aimed at developers who want to see how the pieces fit together. Finally, part three is a bonus section exploring how the same chatbot principles can be applied beyond eCommerce, like in virtual assistants.

You can follow the progress on GitHub at Automated Agents, and you can already read it online at automatedagentsbook.com. If you spot any issues, feel free to raise them there.

The startup I joined ended up thriving and was acquired by a larger platform. That success validated the ideas that form the backbone of this book. I hope the lessons I’ve learned along the way will help anyone else interested in building chatbots.

---