Simply because current tools are somewhat adaptable doesn't imply this tool must be more so. There is actually zero guarantee of that. You're expecting of level of generalization that LLMs simply don't exhibit. If the website is within its training set or very similar to a website within its training set, it should work. But, website design changes fairly regularly. Amd many changes designed to promote usability actually confuse human users and result in lessened usability until the person becomes accustomed to it. That is simple for a human to do, but much more complicated for an LLM. Any a change that confuses a person is likely to also confuse an LLM like this. I also question the idea of using this product for web interaction generally. The applications where it might make sense are better served by dedicated APIs because those are more direct interactions and programs interfacing with them are less susceptible to errors.
As to this being a solution to data entry, I think that is questionable as well. Many such positions have been replaced by automation already where it makes sense. And that automation is much more reliable than anything demonstrated by OpenAI. Your belief that lower accuracy is rates would still be beneficial is also questionable. Garbage in/ garbage out is common saying in the industry for a reason. High volume of a garbage in isn't inherently useful just because there is a lot of it. Usually it just means it will cost more to fix.