I've tried using the Team version and tried building Actions to connect it with a SQL database. The database is old and doesn't have a complete information schema, to be fair, so I've created one in Excel and uploaded it.

I've found that ChatGPT, even with a lot of instruction, can't figure out how to write good SQL queries.

2 out of 5 runs will result in calls to objects that don't exist.

Aside from that, building the Action to use the queries directly against the database has proven difficult. You don't get any errors back so it's nearly impossible to troubleshoot why it fails an API call.

Finally, it doesn't remember new information, so it has to learn all the tables and columns again with each new chat, which means it makes the same errors repeatedly, which means the first question on any conversation takes around 15 to 20 seconds to get anything back at all.

Overall, it pretty much sucks.