A few months ago you couldn't talk about AI without mentioning ChatGPT in the same sentence.
I think those days might be changing.
The latest projects I’ve been testing with the AI tools are case analysis projects.
Now, I will stress that I have removed all personally identifiable data (PII) from the documents, because I still don’t trust the security for the browser-based chat tools (e.g., ChatGPT).
For this use case I’ve taken a few pieces of data to try to understand our veterans disability cases.
- Service Accident
- Branch of Service
- Current Rating Decision
- Main Goal
- Estimated Case Value
I could enlist a virtual assistant (VA) to try to help me analyze all this data (over 300 cases), but that would take weeks to get the information that I need today.
The alternative is an AI tool.
We have a few main questions that we want to answer:
- What injuries from service can contribute to certain disabilities?
- Which disabilities usually get denied by the Veterans Administration?
- Which injuries + disabilities combinations could help us predict a stronger case for our attorneys?
That’s a lot of data to handle manually, but luckily we have a few off the shelf tools that can help!
The Incumbent Tool: ChatGPT + Code Interpreter
Code interpreter for ChatGPT was a recent (within the last few months) addition that got the AI world excited.
Finally, we can feed PDFs or CSVs to ChatGPT to help us analyze data!
Well… it didn’t work out as well as I had hoped.
I used this prompt in ChatGPT:“Identify correlations or patterns between known VA service accident injuries and disabilities that have been denied”
Awesome, it realized that I took a CSV and turned it into a plain text file to reduce the size of the document. It also figured out that it should use a comma as a delimiter to make the data make sense.
Off to a good start.
But once it starts reviewing the data, that’s where it starts to get confused.
Well that doesn’t help me. The actual output was a bullet point list of common words that it found.
I could build another tool to do a deeper analysis, but that would take time. I really want some insights now to share with the team.
Claude 2 from Anthropic Enters the Race
I gave Claude 2 the exact same data and asked the same question.
It took a minute to analyze, but the response was impressive.
Sorry, I can’t actually give you the answers that Claude gave me because that would be too easy for another law firm working VA disability cases to improve their content calendar.
After reviewing the uploaded data, I can confidently say the insights and answers to my questions were fantastic.
This was the equivalent of hiring a VA that could crunch hundreds of rows of data and immediately give me answers to any follow up questions I asked.
Using Claude we were able to figure out what injuries are more likely to cause certain disabilities, why they may have been denied for those disabilities, and the estimated case value for each injury so we can prioritize.
The way it was broken down is simple so I can hand it to someone on my team who may not fully understand the topic (e.g., which cancers are related to burn pit exposure) and they’ll have a good direction to start drafting a content plan.
I asked multiple follow up questions and the responses were as good or better than if I had hired someone to review the data and provide a report on the findings.
Are You Really Giving Up on ChatGPT?
I’m not totally giving up on ChatGPT, but Claude definitely got a higher placement in my browser bookmarks bar.
The downside of Anthropic Claude, is that I don’t have access to their API (yet). I can’t directly connect other programs to expedite work like I can with OpenAI’s GPT4.
I’ve tried multiple document analysis exercises and the Claude responses consistently come out more accurate and helpful.
For some writing tasks, I think ChatGPT is still a good option, but Anthropic’s tool has an edge with document analysis.
Limitations of Current Options & What We're Currently Building
As I mentioned earlier, I removed all PII from the document. That’s not always easy to do, so another solution is needed.
This is the reason why we are building CaseScribe.
Gather important information from your sensitive case documents securely and safely.
We’re excited about the future of using AI-powered tools to make us more effective and efficient.
Reading through hundreds of pages of medical records can be tiring. Mistakes can be made.
By letting the tool take the first pass, we can limit errors and better position the case for our clients.
If you’re interested in getting on the beta list, let us know!