Newsletter

As trailed, my first newsletter. Thanks for subscribing, I appreciate it.

The first question I got after launching my blog was "What is an AI agent?" It's a fair question. Like so much related to AI, the term seems to have arrived out of nowhere.

Speaking with IT guru and fellow Society for Computers & Law member Bill Blackburn at a recent event, I was reminded that the concept of agents is not new: systems programmed to take decisions are all around us, from thermostats to automatic doors.

Of course, recent AI developments like LLMs have enabled agentic systems to become more complex, using many more types of tools (computer programs such as web browsers, email clients, APIs, etc.) to pursue much more ambitious outcomes. Goal-based agents can now do our bidding; utility-based agents can maximise results, learning agents can improve with experience, somewhat like humans do.

The most famous AI-based agentic system – the one that invented its own religion – is OpenClaw. Professor Hannah Fry's YouTube video about her experiences using OpenClaw is ideal if you are new to this space, and entertaining even if you are not.

The promise of using AI-based intelligent agents to take on legal work is what we're excited about here. Can we build agents where accuracy, not evangelism, is key?

My own agent

The Plan

My plan is to build agents that can help a busy GC or in-house legal team. I previously flagged why the monitoring of legal and regulatory developments is a good use-case:

... having a defined process to ensure that the organisation is across legal developments is often a mandatory control, one which typically falls to Legal and/or Compliance. The investment case for this work is clear. Impacts spotted and mitigated late can be very challenging (…been there, done that…). The deployment of resource is therefore necessary, but time-consuming and repetitive. It is often manual.

Progress

My agent got off to a stellar start. I gave it the persona of a paralegal skilled in legal research. I was fretting a bit about the data model, but I needn't have worried.

Within an hour it had a published its first results to the web. Take a look at the knowledge-base it created. My write-up of the experience is here.

My virtual paralegal doesn't have the kind of freedom that Prof. Fry gave her OpenClaw agent. For a start, it lives on my laptop and I have to trigger it manually. That's not a bad thing because it makes it easier to build and maintain. It is eminently doable by any busy lawyer with a copy of Claude Desktop. If you haven't yet dipped a toe, I'd encourage you to give it a go.

Manual oversight can also be a virtue. I've only run it twice and it has already misbehaved, inferring tenuous links and ignoring my instructions to keep topics separate. It was easily corrected, although it took a few tries. Not yet ready for prime-time.

Next steps

The plan is to:

1. Refine its persona and prompts, to reduce risk of mistakes.
2. Give it 'life' (?) ... put it on its own computer and make it run itself, OpenClaw style.

Now this is where it gets interesting. Is this wise? How easy is it to do? How much does it cost? What guardrails do I need to put in place?

I am starting small because I want the 'from-scratch', hands-on experience: How do I make sure that my agent’s work is accurate and good quality? How will I spot if it makes false inferences or hallucinations, especially in areas of law I'm less familiar with? How much does this really matter as an input for a busy senior lawyer, c.f. if we were to expose this agent directly to the business?

Please send me feedback. Your own attempts? Connections to people doing this at scale? Thoughts? E.g. Am I pitched at the right level? Anything you'd like me to cover?

FYI, I think there will be a narrative arc:

• create basic agent ✅
• refine its persona 🧑
• give it its own computer 💻
• give it better tools 🔧
• harness it (guardrails) ⚠️
• refine its workflows to reduce cost 💰
• put it into production 🎉

... but we'll see. I'll keep you posted.

From around the web

Here’s a round-up of what I’ve come across on the web:

• Claude for Legal has caused quite a stir. BigTech (frontier model vendors) getting into LegalTech. We knew it was coming – the market had already reacted. But it's still significant to see it, try it. Is it revolutionary? Claude is, yes - as are other frontier LLMs. Claude for Legal is a build on top; a great template to help the industry implement. See for example, Helen's Legal AI Lab.

• Harvey and Legora greeted the news positively, as validation. But will these platforms, sitting atop models built by the tech titans like Anthropic, Google, OpenAI and Mistral, be able to compete with well-tuned retail offerings from those very same entities? For now, very likely yes; there's lots of room for everyone. See analysis by Artificial Lawyer; JD Supra; Non-Billable.

• Startups are open-sourcing their platforms e.g. MikeOSS (YouTube Walkthrough), Lavern.AI. Perhaps hoping to compete with the big players by building the next RedHat Linux? I.e. the core product is open; support and professional services is monetised. Will it work?

• In the SaaS era, many would have considered it unlikely, but this could be the epoch for bespoke, customised computing. Why adopt what your competitor uses when you can instead use AI to build something better? Is the legal industry ready for this mindset? K&E is.

• Others just want to vibecode. Check out Vibecode.law. They are organising "the world's largest" legal 'vibeathon' if you fancy a go. No coding experience necessary: the AI does it for you.

• Now is also the time for open legal datasets to shine. Models are most accurate when they have access to high-quality, curated inputs. Many legal datasets are proprietary but there is good work to catalogue and (ideally) make available open legal data to LLMs: Legal Data Hunter, Awesome Legal Data.

• Interesting to see competitive alliances forming in the closed dataset space? LexisNexis asked to be removed from Claude for Legal's code base. "Left untouched: Thomson Reuters / Westlaw / CoCounsel".

• I've been speaking with people building agentic AI systems for production use. A shout-out to Nejc Novak from Arna AI who gave me a behind-the-scenes demo of how they've built out a harness to allow legal teams to expose AI agents directly to their internal clients. Thanks Nejc!

• This made me smile: Artur Kulembetov over on LinkedIn: "...the current legal stack is basically twelve tabs, five folders, three portals, and someone trying to remember where the signed version went.". We can do better than that!

I don't have a fixed schedule for these emails yet – it will depend upon how much progress I make, but drop me a line if you have observations, feedback or questions. Or just to say hi. I'd love to hear from you.

Best

Chris