Sure! In classic accounting firms you'd typically have one accountant doing the books and then the other reviewing the first accountant's work. Given the model's accuracy (and models will only get better), only by automating the "doing" and keeping a human to do the review part we will already shave huge amount of time and costs without risking the compliance part.
Does the reviewer learn to be a good reviewer by first doing though?
What if the model misses something entirely, so it's not in the review? ;)
It is given a list of transactions as an input. If it misses one or more of them, that will be noticeable immediately. In the VAT return, one can see if something is missing very easily if the numbers do not align.
At some point I feel like you're better off making the model build a tool to scrape all the values...