The AI Chat UI

Here’s a fun exercise - what would it take to replace your software’s interface with a single chat box?

Here’s my thoughts working through it:

Define all the “Commands”

For this AI Chat Box to work well, it needs to have an appropriate understanding of all of the available commands / mutations / writes / whatever you want to call them.

More specifically, it needs to know the following:

1. What is the intent of the command?
2. What are the required parameters of the command?
3. What is the interface of the command?

Given that information, the AI Chat Box should be able to translate the intent of a user into a command, collect the required data, and then transmit the command to the correct destination.

Define all the “Read Models”

It’s not enough to issue commands to a system, you also need to query internal state. The AI would need to have an appropriate understanding of all of the available data, description of the fields, and an understanding of how to package that data in a way that is digestible for the end user.

Consider:

• Do we need to support in-chat graphs? Do those graphs need to further have additional selections?
• Would this become the preferred interface for all clients? Would other integrations need to “prompt” our system to return a JSON / XML response?
• How do we put in safe guards to prevent the unwary user from accidentally dumping an entire database?
• Is the LLM’s context window large enough to understand every field across every record in every database within our system?
• Would users want to type the exact same prompt every day just to generate the exact same report?
• How would you layer read policies on top of this in a way that is understandable to the AI while also ensuring security?

Will it ever just be a plain chat box?

At least right now… I don’t think so.

But, I do think that very quickly, we will have more intelligent “command” based agents.

The question I have is how we can create a uniform documentation process to allow ai agents to quickly digest the available commands within a system to enable the ai agent to know how to work within the system. And… of course, writing systems in a way that makes it easy for AI agents to interact with them.

But… that’s my 2 cents. What do you think?