This seems very complex given that I know little about accounting. Obviously, you'd need to invent some kind of standardized test for real data.

For example, if you administered a leetcode style examination to me, even at the easy level, I'd probably fail miserably compared to an AI.

However, I am a far more competent professional programmer than an AI.

Is accounting all just rote knowledge? Is there some degree of correctness that can be measured?

Will be interested to see what you come up with if you plan to put data behind your business model.

My wife is a CPA accountant and I'll go out on a limb and say that 80% of her time is spent doing the equivalent of data cleaning and validation. The balance of her time is making judgement calls so she can appropriately reconcile notional rules against the messiness of real world business transactions.

This seems similar to basically all knowledge-work.

99% hampered with low-cognitive throughput tasks, where most of your business value is generated in the 1% of what you do.

It's possible that AI isn't close to replacing human efforts.