AI assistants: a case study
The rapid rise of artificial intelligence brings many new possibilities. AI assistants are one of them. This case study is about our first experiences with this technology.
What is an AI assistant?
An AI assistant can answer visitors' questions without the need for a website or an employee. They usually do this using commonly used Large Language Models (LLMs) like ChatGPT and Gemini.
The assistant receives specific instructions and is provided with additional information on a specific topic. Based on these instructions, the assistant can answer a wide range of questions.
Ronnie.AI
Ronnie.AI is our own experiment. Its goal is to answer questions our visitors have about our experience, working methods, team, and more. Since Ronnie is a prototype, it certainly doesn't always go smoothly, as you may have noticed.
In any case, Ronnie helps us better understand how AI assistants work and what they can mean for our customers.
This assistant is built using Google's Vertex AI agent builder.
Instructions
An AI assistant receives instructions for answering questions. This can be anything, such as:
- Tone of voice: a way of writing
- Language: a language in which the assistant should respond
- Profanity: instructions to avoid using swear words
- A reference to relevant documents that can serve as input.
Example Conversations
To help the AI better understand how it should respond, you can provide it with examples. This is a series of common questions and answers from which the AI can deduce the intention.
You can later supplement these examples based on conversations the assistant has had.
Documents
In the instructions, you can refer to documents containing relevant information. This can be anything, and the AI will try to incorporate this into answering questions.
In Practice
This all sounds very simple, but in practice, it is difficult to really gain control over the assistant. Errors are hard to predict, and even with very specific instructions, the assistant doesn't always listen to what you want. The LLMs are trained on a gigantic amount of information, which often has nothing to do with your goal.
That is why our Ronnie remains an experiment for the time being, but we continue to explore how AI can further benefit us and our customers.
