I’m curious how this increased throughput happens.
You’ve accurately stated that AI isn’t as rigorous as a trained attorney. Doesn’t that mean that every single datapoint must be confirmed by a human?
How is that quicker than just using a human to read the content and make the call? Data entry savings?
Yes, remember that these are effectively random PDFs in various different designs and formats, some of them not editable or even OCR'd.
It took a human attorney 20-30 minutes on average to manually copy-paste data from these PDFs into a spreadsheet (while also fixing any errors they found in the document and re-checking for quality).
Now, the AI copies everything into the spreadsheet in a small amount of time, and then the human reviews it. It takes maybe ~5-7 minutes to scroll to the appropriate pages in the document, read the lines vs the spreadsheet, and make corrections. So you've gone from 2-3 items an hour to ~8-10 items an hour.
Maybe you could pay someone to develop an OCR/ML application that could do this. But that project would never be profitable, even with the time savings. At the cost of a couple Claude subscriptions, it makes sense.
Does the human find enough bugs that they stay on guard, or just rubber stamp everything without really looking at it? It’s hard to stay vigilant when stuff looks plausible.
Complete speculation: you could instruct Claude to hide one random mistake in every document.
And Claude should write down the mistake in a sealed envelope, so it doesn't make into the database.
A review that doesn't find the mistake counts as invalid.
This is what bag scanners at airports do - the hit rate is so low and the job so boring the software projects fake contraband onto the imagery. Fail to spot the knuckledusters and expect a chat with the manager.
Don't forget that humans have a not insignificant error rate when copy/pasting or copy/typing data.
And it's possible to run each document through the LLM pipeline multiple times, using different models and/or prompts each time, to check for errors and inconsistencies. That will take more time and cost more, but it can reduce the error and hallucination rate significantly.
I'm doing some public court records processing for bankruptcy cases (interested mostly to seek out corruption in big national cases), and yes, the "variousness" of random PDFs is exactly the issue. Trying to get the cost for a whole case down to a minimum.
Sample is around 300 court dates, shy under 1k files.
At best I'm building a claude skills file.
> Maybe you could pay someone to develop an OCR/ML application that could do this. But that project would never be profitable, even with the time savings. At the cost of a couple Claude subscriptions, it makes sense.
A better use of these Claude subscription would be to develop the app (which it can pretty much do at that point) and you could iterate to make the workflow even more efficient than your current one.
Nobody working there has the requisite experience to do this in a reasonable amount of time. These are not particularly tech-savvy folks, Claude use aside.
Yes. And it might not even be worth it, as the AI agents gets cheaper and cheaper.
Keep in mind that the task is fixed, so as the frontier of AI advances, you can switch to a cheaper trailing edge system and still get the same or even better performance for this task.
Thanks for sharing the details. Does the attorney check that the AI copied the data accurately? Or is it just assumed to be correct?
Your experience mirrors my own. AI is great for parsing data that can take up a huge amount of time. My only concern is whether or not it’s done accurately. I wouldn’t use it for anything where mistakes cause serious consequences.
You don’t need a trained attorney to schematize data. The LLMs are used to make the data easier to understand and manipulate.
They'll also hallucinate and change meaning in the process of extraction and "schematization"
Not necessarily. Depends how you use it.
"Write a python script that breaks down this PDF by X feature" would not hallucinate anything in the PDF. Certainly you could trivially double check that all text in the extracted JSON was in the text layer of the PDF.
How much experience do you have with LLMs exactly? It would be consistent with my experience if Claude stuck in a line of python that just emits a JSON literal with no justification, potentially buried in a large program where an untrained person might not notice it. I don't even trust them if the output consists of structured data paired with source images from the PDF, because I've experienced LLMs fabricating the source rectangles to match the output. I only use tools like this by asking for programs, because as you note LLMs are good at that, and the verification process consists of tool calls to legitimate PDF manipulation tools so I have some confidence everything is above board. Even then I only do this for hobbies, not anything that matters.
Lawyer here. I used to trust Claude as hallucinations are near non-existent now. However for large volume tasks such as due diligence exercises, they still happen. We also tried Legora's tabular review, there were also numerous halucinated provisions in our due diligence exercise.
Is it possible to catch those hallucinations using another LLM with a strong fact checker prompt with sources provided in output for human validation?
Junior associates hallucinate too...
And when they do, you can train them or fire them, and they learn not to do it.
LLMs change not a whit, and there's no one to take responsibility for the failure (and thus no way to fix it).
As the new variation on the old theme has it, "A computer can never be held accountable, and so very many people are trying to get them make management decisions."
You can’t train people to never make a mistake, particularly when doing highly repetitive work like this. You must build your systems to account for that regardless.
You can scold juniors and they will learn. You can't scold Claude.
Surely the rate of improvement in new LLM models is the equivalent mechanism?
You can scold Claude. Just doesn't make a difference.
... until you hit Claude's risable "model welfare" protection.
That's a recipe for disaster in my experience. I tried it (with Claude) on a simple tabular bank statement PDF, and it transposed two amounts, placinh each against the other's description. And the bot assured me the result was cotrect. The chance of a human checker catching such corruption is low.
Interesting. Did that PDF have a text layer or did you ask Claude to OCR it? If the latter I'm not surprised at all.
Doing similar-ish things with Claude, it's helpful to have something to ground it.
For instance, if you can say:
"Refer to the database schema in x.sql as your source of truth for the database structure we want to import into*. Do not invent data, tables or columns that do not exist. Carefully match all output against this database schema and do not create output that doesn't exist if it does not match the schema, simply skip it."
You will end up with a far better result in my experience.
Gotta treat it like a child.
> Gotta treat it like a child.
"Sorry for that, Your Honour, but we gave that case prep work to a child."
They do, sometimes. That's why the review still has to happen.
But now it's comparing already filled columns on a spreadsheet, not copy-pasting every single thing from an (often uncopyable) PDF.
> But now it's comparing already filled columns on a spreadsheet
... with a PDF, right?
> not copy-pasting every single thing from an (often uncopyable) PDF
Obviously the PDF is copyable, else your bot would not be copying it.
Reviewing something takes less time than producing it.
I...don't think that is universally true.
Few things are. It also doesn’t require standards. “Stamp this diff” culture is everywhere even before AI. A stamp is literally easier than anything else.
Whether that is useful measurement I suppose depends on the circumstances.
Yes, but empirically in this case, it seems to be true.