I've made a Slack bot for applying LLMs (Google's Gemini) to legal questions for my clients. This system excels at contract analysis and discussion of commercial & crypto legal issues, but could be applied to any domain. This blog post explains how it works, what it is, and how I got here. At the bottom you can read about trying it out yourself. I've been very impressed with the results, and so have some of the people who've tested this out: ***** - this is fantastic, I had a junior conduct this review for me last week.

Why Slack?

Most of my clients use Slack and it's a convenient way to give access to many people in a company to the tool. It also lets me see the outputs and anyone else can easily comment and interact with the bot. Slack also has concise messages, so it's a good source of facts that can be fed into the context window of the LLM, which means that the bot can answer based on the knowledge of the specific company. So each private Slack channel begins with a few sentences about the name of the company, where it is, risk tolerance, etc. LLMs without context about the client do not work well, just like lawyers without context.

Why Private Channels: Preserving Privilege & Avoiding A Bradley Heppner Outcome

A widely publicized US case recently showed the danger of using LLMs without a lawyer overseeing it, because the person lost privilege over the advice. That is very bad if you are a litigant. And it's unnecessary, because lawyers can grant access to their clients to use these tools, and then have oversight of the tool, and it will be within the scope of privilege. I'm very confident that if the person had been using a setup like mine, where the Slack channel was established and overseen by the client's lawyer, that privilege would not even have been an issue.

Providing Detailed Analysis As Pre-Review

My service works by loading extensive prompts that are used, plus the context of the Slack channel, plus code that identifies what the tasks are (which is also powered by an LLM, of course). It then produces a structured output that follows a general pattern, or, it will answer back with a single reply (useful for Q&A). It intelligently selects the best output based on the task at hand. This extra prompting and scaffolding is what makes the difference, especially for users who aren't as sophisticated as lawyers, because the main target audience for this is the lay public. The idea is to do a pre-review of the issues, so the person is aware that there even are issues, so that they can raise them earlier in the process. Too many employees avoid asking things of legal because that's too big a step, whereas asking an LLM is free and easy.

Getting Here: My LLM Journey With Law Applications

Regular readers will know that my starting point in law was as a programmer. I built a number of tech businesses myself at one point, but for the last decade I've been focused on helping my clients build great businesses as a tech-savvy lawyer. Anybody that says they are a tech-savvy lawyer should have been spending a lot of time with LLMs over the last few years, and I have! I've written an anti-money laundering PEP screening service that scrapes websites, book chapters about crypto law, an AML advice service, and an automated job board for crypto jobs. Last summer I made a service that did contract reviews by email (which I never published, but it worked pretty well). In addition to these programming projects, I've tried to apply the various LLMs to my work (which can't be posted online). In short: I've been deeply experimenting for the last three years, not just using LLMs but programming with them and using the APIs for larger-scale projects.

Recently, LLMs have gotten good enough to start doing useful work.

Technical Underpinnings: Making More Bots

The bot code is not vibe-coded! That's not a point of pride, but rather to say that it's somewhat difficult to put together a system that does this still. I don't expect that to be the case for long with the advancements in coding and platforms for launching agents. But it is not trivial today to get this working well. And the prompts that go around the bots require some legal knowledge. The fun part is that there's a mini-language for writing bots, which supports variables, loading files, and writing text, so that I can quickly make new bots for different use cases. If this is successful, perhaps clients can write their own too (I had that in mind with this format, because it's relatively safe).

I've found great results with Gemini Flash 3.1, and it's so cheap that this service is essentially free. In the world of legal services, this is what free looks like. So lawyers can offer this to their clients without breaking the bank (some models can be very expensive). This also requires being able to convert Word documents to PDF, which is something that happens under the hood, because Gemini only reads PDFs at the moment and lawyers are addicted to Word documents.

Pre-Review vs. Drafting

There's a lot that can be done with LLMs today in legal, but most lawyers aren't going to be replaced by bots. I think the current sweet spot is doing pre-reviews. And often the pre-reviews are very similar to what I would say as a lawyer (which is kind of the point.) Many issues are routine for lawyers but new for people who don't regularly deal with contracts, and LLMs can help to close that knowledge gap in a quick and easy way. That's where my bot idea comes in. Other tasks like drafting contracts, or writing legislation, are just not ready yet.

Try It Out!

If you want to try this out, I'm willing to make you a private Slack channel so you can use it yourself. Feel free to email me: addison@cameronhuff.com.