My recommendation: Do manual and automated solutions. I worked on a similar issue back in ~redactedwhere it was of utmost importance that all the numbers were exactly right and I handled the coding of automated extracting tools while the rest of the team manually did the same on their end. Both methods were flawed, but crucially they were flawed in different ways so we were able to dramatically up the accuracy by combing the two approaches. (Automated failures tend to be on hard to read spots that manual readers naturally pay more attention to. Manual failures are when things get boring, which automated solutions excel at). If you do want to avoid the manual time & cost, I'd use local LLMs to sort your PDFs into level of difficulty and importance so you focus your spot checking on the ones that are likely to be wrong in very significant ways.

I wanted to add my thoughts here, albeit a bit late. I've seen Visual Language Model paired with deterministic reconciliation logic works extremely well. We've used gpt-4o-mini (as a VLM) previously which handles even handwritten stuff (as long as it's not scribbles). We've used it programmatically via OpenAI's API. The newer models "reason and think" too much but don't create better quality for this use case in my experience. Besides using pure VLMs, the one thing which will tremendously help is to make sure your algorithm (whether in Python code or n8n flow, etc) does a mathematical reconciliation if possible. For example, ensure that all line item amounts adds up to the subtotal, and/or if you can see the tax percentage, ensure that the tax amount is tax_perc * subtotal, etc. The combination of non-deterministic VLM and deterministic math tremendously increases accuracy. And because GPT-4o-mini is now considered an older model, the cost is ridiculously low. Here's the math, I just did one invoice: it took 3 seconds 25,600 input tokens (scanned image) and 122 output tokens. Cost for 1 invoice: 25600*0.15/redacted+ 122*0.60/redacted = $redacted for 20,000 invoices: $80 and 16.7 hours (the cost of developing the automation system will be much higher unless you have the expertise in-house. One caveat, as is the case with any non-deterministic AI system, you want to implement as many checks and balances outside of the AI, I mentioned the per invoice math formulas, but you can and should also identify ways in which the total amount of invoices processed for a month should match a certain amount, or per customer, etc.