In my (admittedly limited) experience, the difficulty of working with Selenium doesn't stem so much from finding appropriate selectors but from britleness: network delays/timeouts, objects not yet being yet lodaded into the DOM, dynamically generated classnames (think scraping obfuscation).

I do see a niche for an system (AI or not) that deals with that last case (i.e. it automatically grabs some correlated selectors to fallback to), or as a quick tool to scrape static sites.

However (and this may be out of your proposed scope, and that's fine) I see more value on something that can construct a sort of DOM timeline - in order to accurately know what is available and when. If you start "recording" network and user events since page load, you may be able to reconstruct a) which nodes the user interacted with b) which preconditions there are for these nodes to be consistently available, no matter which stochastic delays/errors are present.

This is tricky and time consuming even when done manually, so I'm not sure it can be AI-mplemented. But that's maybe an idea to explore down the line :)