Ljupka Schwantes LinkedIn post on this week’s “Claude for Legal” announcements rings true. I am excited about leveraging the RelativityOne and other legal AI integrations for clients, however risk and cost control drive my consulting and matter practice. I have active integration projects where these bidirectional connections may slash development budgets and timelines. First, we have to set clear guidelines and requirements to prevent exposure of client content and attorney work product. Compliance controls and functionality always seem to lag innovation, especially in light of the AI sprint we are living in. My friend Kelly Twigger’s articles on the evolving caselaw on the intersection of AI and privilege highlight how the bench is struggling to give us guidance. Run your own AI summary to explore the functionality and risks associated with the 20+ legal connectors launched last week.
Rather than opine on the new offerings, I wanted to share feedback from my own AI research and vibe coding experiments. The only way to understand these systems is to use them yourself. That is how I have always kept on the cutting edge of enterprise (Purview), eDiscovery (RelativityOne) and as many of the tools as possible. Trust but verify through your own usage.
I resurrected an ancient passion project and started using precious weekend time to teach myself AI driven development on a modern environment. I am not a developer. I can kluge scripts and taught myself Cold Fusion web coding for one of my first startups. I do not consider these real development skills. A recent episode of Hardfork (Using Gemini 3 Is Here) inspired me to literally dust off a set of 1980’s notebooks filled with my poorly handwritten fiction research to see if Gemini 3.1 could convert scans into accurate text for a knowledge repository. It was fun doing glasswork scanning while watching cooking shows on the couch. Especially when I ran into ideas scrawled on HPD letterhead from my days in the crime lab (yes, I have always had some kind of startup brewing in my off hours).
Impressions from an old school geek:
- You still need to be able to read commented code to debug and tweak. But you no longer have to guess at syntax.
- AI can now guide you through the development process, especially if you are already kind of techie inclined. Ask it to interview you on your project goals and design staged prompts to get started.
- AI is insanely smart AND dumb at the same time. Verify everything. Example, Google AI Studio (Gemini) ran me in circles for hours insisting that I had to use Gemini-1.5-Pro or Flash when that version was not supported for my project or API keys. I made it write me a python snippet to generate a list of supported models from my API key to finally figure this out.
- The AI methods, calls and names are changing so fast that the coding models cannot keep up. This is similar to issues I have observed using Copilot to answer questions about Purview. It references outdated documents (or ingested training materials) and comingles answers from the legacy and modern Purview UIs.
- I was gifted a couple of great pre-prompt instructions to minimize hallucinations and put the model into ‘strictly factual’ mode. The instructions consome some query cache, but help ensure that I understand the source and the intersection of fact vs. reasoning. These are usually verbose, so you can feed your own as a prompt and ask for a model specific condensed version (use the AI against the AI).
- Speaking of AI vs. AI, feed the same prompts to multiple models for fast comparisons or the results from one model to another for evaluation.
- Free tier AI is incredibly constrained for more than simple smart search results. Worth experimenting with a couple prepaid or minimal monthly subscriptions while you are evaluating models.
- Any AI service processing actual documents or uploaded data is going to take processing time and will have request limits. Build delays and ‘cool down buffers’ into your code. This varies per model and subscription tier.
- Data Protection Agreement terms vary wildly. If you are using a public model, you should assume that anything you input may be used for training or otherwise exposed. Google’s NotebookLLM has great knowledge base features for my project, but their terms are not acceptable for real personal information or copyright purposes from a non-attorney perspective.
- My program processed 165 MB of scanned documents for 3-4 hours in the Gemini-3.1-Pro model. Much of that time was 30-60 second buffers. It output ~1.5M tokens worth of RTF files for a total cost of $4.75. Worth every penny to see the clean, formatted text with tables and even understandable diagram content.
Yes, I was always a geek and proud of it.
I hope my experiments encourage other practitioners to try vibe coding or similar AI projects. We all need to invest time (even non-billable hours) to keep up our skills sharp. Next trick is to see if Copilot (maybe using Claude) is better at generating my monthly time allocation reports. My early testing was not encouraging, but who knows now.
Circling back to the Claude for Legal integrations, I encourage everyone to ask your favorite AI to outline the documented functionality of any of the partners whose platforms you use. Here is my prompt in Researcher mode:
research recent release of Claude for Legal connectors with a focus on RelativityOne integration to outline functionality, workflows and potential content and legal workproduct exposure risk.
Start looking at low risk workflow or usage scenarios that do not involve actual eDiscovery content, strategy or work product. Relativity seems to have taken that approach for overall workspace management, reporting and similar administrative support. In contrast, Everlaw and Consilio Claude integrations involve actual content searches, classification or other higher risk functionality as Schwantes outlines in her post. So play, learn and test in responsible sandboxes before risking client confidentiality, privilege or sensitive information. I would love to hear about your experiences!
Greg Buckles wants your feedback, questions or project inquiries at Greg@eDJGroupInc.com. Reach out for a free 15 minute ‘Good Karma’ call if he has availability. He solves problems and creates eDiscovery solutions for enterprise and law firm clients.
Greg’s blog perspectives are personal opinions and should not be interpreted as a professional judgment or advice. Greg is no longer an investigative journalist and all perspectives are based on best public information. Blog content is neither approved nor reviewed by any providers prior to being published. Do you want to share your own perspective? Greg is looking for practical, professional informative perspectives free of marketing fluff, hidden agendas or personal/product bias. Outside blogs will clearly indicate the author, company and any relevant affiliations.
Greg’s latest nature, art and diving photographs on Instagram.
