Because it takes a dedicated team reviewing commands all day, every day, and dev team implementing event handlers, coordinating with other product teams. It takes lots of dedication.

I used to run a chatbot that screened people. We thought we were asking simple enough questions, and yet two million chats in, we still discovered at least a Dozen unexpected/misdetected intents that required care. A day. And here is the problem: management will ask: "so if we add handling for 'rotate the map 30 deg', how many people will appreciate it?". The answer is usually "one person every 400k sessions". For an MbA trained manager that means there is no reason to do it. And its the same with every other similar discovery.

Running an NLP based tool requires staffing a 10-20 people full time, telling them to keep the quality high and leaving them the fck alone. Thats anathema for most companies that proclaim to be "metric based". Google is a prime example of this behavior.