The whole point is that this OpenAI is positioned as a universal tool. It can be connected to any program as an external library, and the connection takes only 7 lines of code.
So, in theory, nothing prevents integrating this thing into the Heroes games, training it (for this, the AI only needs to play against itself), and then you're good to go. Although, there is a challenge: a game of Heroes is objectively longer and more complex, and besides, even in Dota, the AI was trained with limitations (they took a limited number of game modes and a very limited number of heroes that the AI can play as. It seems like only one hero).
If you're interested.
Article on Habr
Website dedicated to OpenAI
On GitHub
If only 7 lines of code are needed, then why were 12+ people needed in the team that created this bot for Dota (https://www.youtube.com/watch?v=WKzuUAY2c4o at 13:30)? Moreover, if you listen to what their team lead says, they were doing trivial things. Playing in the mid lane in Dota is, by its nature, almost a purely technical task, and besides, it's a task with open information. If I heard correctly, they were working on it for about a year. Considering that this is only for the Support role and only for playing in the mid lane. So don't worry).