1. Conversations about AI have shifted from what to how.

Last year, conversations at Legalweek centered around foundational questions like:

What is AI?How do large language models work?What does predicting tokens mean?

The focus was largely educational, with many legal professionals still wrapping their heads around the basics of AI.

This year, the conversation has shifted towards implementation — the nuts and bolts of integrating AI into legal practice. Law firms are now grappling with practical decisions like:

Should we build AI tools in-house? Or partner with vendors?How do we decide on which vendors to partner with?How might our security programs change as a result of implementing AI applications?

So, what are firms actually doing?

I noticed many are choosing to build basic capabilities themselves — like chatbot wrappers around LLMs — and turning to vendors for more complex applications.

2. Despite all the buzz, most law firms are still in the early stages of AI adoption.

Working in the legal AI space every day, it's easy to think we’re further along as an industry than we actually are. But being at Legalweek helped me realize we're still very much in the early stages of AI adoption.

In fact, one person I spoke with told me his firm was in the top 10% of law firms on innovation simply because they deployed a ChatGPT wrapper to every user in their firm.

I found that very surprising.

Especially because a recent report suggested that 79% of lawyers are using AI in their practice. This number is commonly referenced across the industry but is not entirely accurate. What’s often left out of that metric is that it includes those who merely intend to use AI in their practice.

The reality is that:

The actual adoption rate is likely closer to 30-40%, with most lawyers using only basic tools like ChatGPT or Copilot.Most firms are just figuring out how to roll out AI across their organizations.

3. Training and adoption are the biggest barriers to AI success in law firms.

As a result, firms are hiring armies of people to help encourage AI adoption by:

Training employees on how to use ChatGPTLearning what lawyers want and encouraging them to try existing toolsCreating internal AI tools to improve existing workflows

As a result, firms are converting a lot of traditional lawyers into what they call "innovation lawyers" — practicing attorneys who have transitioned into positions focused on implementing AI tools across the firm. But lawyers are creatures of habit. Despite these investments in human capital, progress remains slower than expected.

For example, when I ask lawyers how they use AI in their practice, they typically highlight:

Summarizing emailsCreating timelinesExplaining contract clausesRewriting documents

And although these use cases are valuable, they're still just tinkering around the margins of the practice of law. I would guess that today, AI is increasing attorney efficiency by no more than 5-10%. And lawyers are not yet applying AI to tasks that eat up big blocks of time, where they may be able to reap 2-3x efficiency when applying AI the right way.

4. When it comes to AI agents, there’s a true spectrum of “agenticness.”

“AI agents” were the buzzword of the conference, but there's definitely a fair amount of confusion about what exactly constitutes an agent and how it differs from existing tools.

Last summer, I was interviewed for an article about AI agents. They asked me what I thought about agents, and I told them they don’t exist. Why? At the time,  “agents” were just LLMs prompting other LLMs to do things. In fact, that’s how most “agents” continue to work today.

When clients ask if we're using agents in our products at PointOne, we can technically say yes because our systems involve large language models directing each other. But do these truly qualify as agents? I’m not sure.

There’s recently been a new wave of advanced systems — like OpenAI's Operator — that can execute tasks with greater autonomy. These are probably more along the lines of what I would consider an agent.

But rather than thinking about what is and what isn’t an agent, here’s a better way to think about agents.

You should view agents on a spectrum of agenticness — from input/output models to complex systems that might eventually function independently:

Level 1 (GPT3): Simple input-output modelLevel 2 (Reasoning Models): LLM talking to itselfLevel 3 (Collection of LLMs): LLMs working together in a systemLevel 4 (Collection of LLMs with Operators): LLMs working together with an operator that enables it to work independently in the wildLevel 5 (AI Employees): AI associates and paralegals

5. Clear winners are emerging in an increasingly crowded legal tech marketplace.

The legal tech ecosystem is maturing rapidly, with over 400 AI startups now competing for attention.

What's becoming increasingly clear is the stratification of the market. Some startups are starting to gain significant traction — securing notable clients, publishing case studies, and growing quickly.

A year ago, it felt like a wide-open market without any mature players. This year, that's starting to change.

This is great news for everyone in the industry because it’s starting to feel less like a scrappy band of startups and more like we’re actually figuring out how to fit into the market.

So there you have it, my five takeaways from Legalweek NYC:

Conversations about AI have shifted from what to how.Despite all the buzz, most law firms are still in the early stages of AI adoption.Training and adoption are the biggest barriers to AI success in law firms.When it comes to AI agents, there’s a true spectrum of “agenticness.”Clear winners are emerging in an increasingly crowded legal tech marketplace.

What are you seeing in your practice? Has AI made a real difference for you yet?