I wanted an AI to answer my customers on WeChat. One evening of testing told me why that still isn't possible.

I run a small medical supplies business. Customer messages come in on WeChat all day, a lot of them are the same questions, and I keep the records in Excel. I wanted an assistant that reads the incoming message, pulls the right record, and replies while I keep working.
So I spent an evening testing whether the building blocks exist. They partly do, and the part that is missing is the part I needed most.
What worked: an open source driver read my live WeChat window as 138 structured elements, including full message text. It located the input box by a stable element id rather than a screen coordinate, which matters because my older scripts clicked hardcoded pixels and broke the moment a window moved. And it did all of this while I kept typing in my browser. My foreground app never changed once.
What did not work: it could not type a single character while staying in the background. Three different apps refused, and the driver told me plainly that the surface drops posted input. I confirmed the empty input box by counting pixels rather than trusting the response code.
The only path for input is a brief foreground swap, which means if I happen to be typing at that moment, my keystrokes can land in the wrong window. For a customer conversation that is not a risk I am willing to take.
So I am buying a second machine. A cheap one. The agent runs there, takes the foreground when it needs to type, and never competes with me. That turns out to be the honest answer to "can one computer do both jobs", and it is no.
One thing I did not expect from a tool I was evaluating: it made me fix my own code. I had measured my desktop automation at 44 seconds per call. A different PowerShell flag brought it to 0.99 seconds. The bottleneck I had been blaming was a startup cost, and I had shipped around it for months without checking.
Full write-up with the numbers: https://saas.pet/reviews/cua-driver/