An anatomy of a conversational AI agent – Business analyst edition

An anatomy of a conversational AI agent – Business analyst edition

I am a strong believer that if you want either to build something or effectively facilitate that something being built, you need to understand HOW something is put together, and how the process for doing so works and why.

That is the approach I’ve been faithfully following for years and years now, and I’m yet to find any flaws in it.

With that in mind, AI agents represent a new class of applications that didn’t even really exist two years ago and are now becoming rapidly ubiquitous and one of the most relevant applications of AI models, LLMs in particular (not counting creating slop to satisfy the social network algorithms – but that is besides the point of this talk).

The point here is that AI agents are something that we, Business analysts and Product owners, should become familiar with if we’re to remain relevant and helpful when we inevitably land on some AI-focused project in the future. And when I say familiar with, I mean understand how they work under the hood and what makes them tick.

So, in this talk, we’ll dissect one such agent, a conversational, multilingual AI agent for tourist recommendations and general information. The agent which I’ve built, partly to gain understanding about the inner workings of such systems, and partly because I really like development and keeping myself technically “fit” and up-to-date.

However, if you are not so technically inclined, then this is the talk for you – I did it, so you don’t have to.

In this talk, I’ll discuss challenges presented by multilingual AI agents in general, what are LLMs good for – and what they are not, how and why I went from classic “tool-using-LLM” to more modern intent-driven-state-machine architecture, how I addressed users drifting from topics that my agent could answer reliably, how the streaming of structured data work and a few others. All in all, after this talk, you should have a decent overview of what AI agents actually are (they are NOT wrappers around GPT – and haven’t been for quite some time now) and where they differ from classic applications.

That should come very handy on your next AI-focused project, or at least it will make all those chatbots and agentic AIs you’re using regularly much less opaque and residing within “there be dragons” territory and much more understandable and down to earth.