Legal AI Talk: The Rise of AI Agents

In this in-depth conversation between David A. Bloch, CEO of Legartis, and Gordian Berger, CTO of Legartis, the rapid evolution of Legal AI is explored—from foundational language models to autonomous agents.

They discuss practical applications, challenges such as data privacy and source verification, and the expanding role of AI in legal workflows.

The dialogue offers valuable insight into where the legal profession is heading—and what this transformation means for both innovation and legal integrity.

What do we already know about AI?

David: Could you give us a brief overview of what has happened in the past few years with AI?

Gordian: I think it’s important to briefly reflect on the past few years, because while AI has existed for many years—perhaps even decades—the most significant leap in its evolution occurred when OpenAI released ChatGPT.

That release marked the moment when the Transformer architecture, which underpins today’s AI, became widely known and usable for the general public—and it worked impressively well. That breakthrough, I believe, can be seen as the true beginning of the current AI evolution. Interestingly, that happened at the end of 2022. It feels like it’s been much longer, but in reality, it wasn’t that long ago.

Since 2023 and into 2024, we’ve witnessed many improvements. Following ChatGPT’s release, other providers quickly followed suit. We saw major gains in efficiency—not only in 2022, but also throughout 2023.

Now, looking at 2024, there was a surprise: the underlying technology didn’t improve that dramatically. While the models have certainly gotten better, I think many users—especially those who’ve used ChatGPT from the start—may not have noticed a huge leap in performance, at least not on the surface. And that’s important to highlight: 2024 has not brought a second jump of the same magnitude as the first.

However, a lot has changed behind the scenes—changes the average user might not immediately notice when simply using ChatGPT. The models have improved incrementally, yes, but at the same time, we can now achieve comparable results with much less hardware. AI has become significantly more cost-efficient.

That shift has led to a kind of "arms race" among model providers. Essentially, we’re now getting better results with much lower input requirements.

Another key development is that providers—like us at Legartis, but also others in the legal tech and AI space—have learned to use the existing models much more efficiently. And this is important because we’re still not using models to their full potential. It takes time to understand what you can get out of them. So even if the models themselves are improving only incrementally, the outputs are improving significantly—simply because we’re using them more effectively.

Additionally, we’re seeing encouraging trends around data privacy. While it doesn’t necessarily improve the quality of the models directly, it is becoming increasingly feasible to host your own AI models. That’s a huge win for organizations handling sensitive data.

That summarizes where we are in 2024. Looking ahead to 2025, I’m optimistic we’ll see another major leap forward—not because of radically new model architectures, but because of new, more efficient ways of using what we already have.

What the Rise of AI Means for the Development of Legartis?

David: Let’s take a closer look at what the evolution of AI over the past few years specifically means for a solution like Legartis. How has it shaped its development?

Gordian: Sure, let me walk you through the different phases we’ve seen. Starting with what I would call the early days—roughly between 2017 and 2020—the development of Legal AI was slow and expensive. Building these systems was incredibly resource-intensive. To give you an example from Legartis: developing an AI model for just one contract type used to take us more than a year.

So, the development time was extremely long. The result was a model that could analyze one specific type of contract and deliver an efficiency gain of around 30%—which was good and laid the foundation. But the challenges were significant. We had very limited access to high-quality legal training data. The models required extensive manual supervision and fine-tuning. Multilingual capabilities were extremely limited. Expanding from one language to another basically meant starting from scratch.

That brought us into the acceleration phase—from 2021 to almost 2024. During this time, we witnessed maturity in AI and big gains in efficiency. We moved from handling one contract type to multiple, and scaling to new types became much easier. The clear trend was toward hybrid workflows: AI-assisted, but still with legal professionals in the driver’s seat.

Now in 2024—and looking forward to 2025—we’re seeing scalability at an entirely new level. AI is no longer just reviewing contracts. It’s summarizing, negotiating, and suggesting revisions. We're expanding across multiple jurisdictions, supporting more languages, and onboarding new contract types with very little effort. That’s the speed of change we’re experiencing right now.

AI Agents and LLM Trends 2025: A New Chapter for Legal Innovation

David: And when we look ahead to 2025, could you introduce our audience to the era of AI agents? What can we expect, and what broader trends around LLMs should we be watching?

Gordian: Absolutely. Let me start by building on what I mentioned earlier: I don’t believe we’ll see massive breakthroughs in the underlying large language models themselves in 2025. Their development will likely continue incrementally.

But—and this is a big shift—the rise of AI agents will make a major impact. This trend is already emerging in 2025. What’s exciting is how much more we’ll be able to do with the technology we already have. AI agents will give us a kind of jump-start, even before the next big leap in model architecture arrives.

So, what is an AI agent? In contrast to a typical language model that answers user prompts based solely on its training data (which is often outdated), an AI agent has a defined goal or task—and access to tools that help it achieve that goal.

For instance, an agent might be a customer support assistant, a travel planner, or a legal policy drafter. The key difference is that an agent doesn’t just answer questions—an AI agent actively works toward outcomes. It can call APIs, access emails, use search functions, or even communicate with other language models or agents.

This access to tools dramatically enhances its capabilities. For example, if a model isn’t great at calculations, you can give it access to a calculator, and suddenly it becomes excellent at math. So even if the base technology hasn’t improved drastically, the user experience does—simply because the agent can interact with the right tools.

Another key development is long-term memory. Unlike regular models that only rely on static training data, an agent can maintain its own database. It can store user preferences, past interactions, or other relevant information that helps it perform better in future tasks.

And then, there’s the concept of feedback loops—which I find especially exciting. Agents can talk to each other. One agent might be tasked with creating a travel plan, another might handle bookings, and a third might coordinate both. You can even have an agent that supervises or trains others. This creates a more self-sufficient and intelligent ecosystem.

These feedback loops—where agents collaborate and learn from one another—are still in their infancy. They’re hard to control and manage, but they’re starting to emerge. And that’s where I believe we’ll see a lot of innovation in 2025.

How AI Agents Manage Other Models

David: The idea of agents becoming increasingly intelligent is really exciting. But could you explain in more detail—what does it actually mean when agents manage or control other models? 

Gordian: Sure. You can think of it as a larger workflow system. In such a setup, multiple agents can collaborate on a broader task. For example, you might have a coordinator agent whose role is to oversee the entire process and assign responsibilities to other agents. While this idea often comes up in travel planning, it’s just as relevant in legal scenarios.

The coordinator agent defines the overarching goal and delegates subtasks to others. This is where things become both powerful and complex. One of the main issues in AI today is hallucination—where models generate information that isn’t accurate. To avoid this, you need an agent to monitor the input and output between other agents. This kind of oversight enables the creation of a bigger, more reliable workflow system in which everything fits together coherently.

We plan to showcase some legal use cases for this setup soon. There’s a lot of potential to go into greater depth.

For users, this shift means that while AI agents will continue to act as assistants rather than full replacements in 2025, their impact will still be significant. I don’t think we should suggest that agents will be fully autonomous anytime soon—that would be unrealistic. We’re still in a phase where humans remain in the loop, and that’s important.

However, one big change users will notice is the reduction in prompting effort. Right now, if you want good results, you often have to prompt an AI system multiple times, refining your input along the way. Agents can handle much of that in the background. The result is a smoother user experience. You’ll still interact with the system, but you won’t need to be an expert in prompt engineering. That’s a big step toward making AI accessible to a wider audience.

But it’s not just about agents. There are other trends we should talk about too—especially the availability of different models. Currently, there’s still a strong reliance on U.S.-based AI models. In the future, it will become more important for users and providers to be able to select and switch between various models, especially when agents are involved. Flexibility here is key, particularly in terms of data protection.

Another ongoing issue is source verification. Hallucinations are still a major concern. And when multiple models are communicating, the risk increases. You definitely don’t want models feeding one another false information endlessly—that’s not only unhelpful, it could be dangerous depending on the context. So, verifying sources and validating what’s being exchanged between models is becoming increasingly essential.

Lastly, model switching and deployment will play a growing role. As models become smaller and more efficient, we’re seeing more demand for on-premise and private cloud solutions. This is especially true in Europe, where many companies are hesitant to rely on major U.S.-based cloud providers. In highly sensitive fields like legal work, organizations are understandably reluctant to expose their IP or confidential data to external systems. The need for local, controlled infrastructure is only going to grow.

Practical Use Cases for AI Agents in Legal

David: That was a bit more on the generic side. But when it comes to actual use cases for AI agents—Gordian, where do you see real, practical applications of this technology in the legal field?

Gordian: There are certainly many possible use cases, but let me just highlight a few from a high-level perspective. First and foremost—especially for risk reasons—AI agents should still be seen as assistants. That means the human must remain in the loop. The ideal scenario is a cooperative process where AI and humans complement each other.

A great example is legal research. Imagine a lawyer preparing for a case who needs to gather and review large amounts of information. An AI agent could help by collecting relevant materials, organizing them, summarizing key points, and even referencing sources. The agent could then present that summary to the lawyer for feedback, making it a smooth back-and-forth process. The lawyer stays in control, but benefits from increased speed and accuracy—and the agent might even uncover insights the human wouldn’t have found alone. This is a strong use case that’s already being explored by several startups.

Another promising area is contract negotiation and redlining. Companies often have specific policies and playbooks that define what is acceptable or negotiable in contracts. An AI agent can be trained on these internal guidelines and then assist in reviewing or negotiating contracts with counterparties. It could make suggestions or even attempt automatic negotiation based on legal best practices and user inputs. Again, the assistant model is key—the user remains in charge and can decide whether to accept, reject, or modify the agent’s output.

Another interesting scenario is risk assessment and contract lifecycle management. You could have one agent responsible for contract negotiation, and another that continuously monitors for regulatory changes. When changes occur, the monitoring agent could suggest updates to your legal department’s internal policies. Based on those updates, another agent could then automatically scan all existing contracts and flag risks or inconsistencies.

In this setup, you have different agents working in tandem: one focusing on legal content, another on regulatory developments, and a third on implementation. This collaborative agent architecture could significantly reduce manual work—while still keeping a legal professional in the loop to validate and control the final output.

When Will AI Agents Handle Contract Negotiations Alone?

David: So, from what I’m hearing, contract negotiation is likely to become increasingly automated. What’s your view—how long do you think it will take until two AI agents can negotiate a contract entirely on their own?

Gordian: That’s a tough question to answer. The truth is, we still don’t know how big the leap in 2025 will be. I do expect it to be significant, but I believe AI will still operate mostly at the assistant level.

So no, I don’t see fully autonomous agent-to-agent negotiations happening this year. However, I could imagine it becoming reality within three to five years. But again, it’s difficult to predict. Sometimes AI development moves incredibly fast within a year—and then it flattens out again.

What I do believe we can promise, though, is that AI will increasingly support users during negotiations. That alone will bring substantial efficiency gains.

How to Choose a Safe and Smart AI Provider

David: That makes sense. And when we talk about AI, it’s obvious that data privacy is a crucial topic. How can organizations mitigate data risks when choosing an AI provider? Are there ways to define clear data guidelines? Or more generally—how do you make a smart, safe provider choice?

Gordian: Yes, that’s a key issue. Part of the answer is still quite traditional—you have to understand your data and the regulatory framework around it, especially things like GDPR. Start by categorizing your data: which parts are more sensitive, which less so?

You should also consider involving your AI provider early on in the conversation. Ask them questions: What happens to your data? Will it be stored, processed, or even used for model training? Can it be anonymized? These are all critical considerations.

Depending on your preferences and legal obligations, you may decide that certain data should only be processed temporarily—or not leave your environment at all. In such cases, anonymization is one viable approach. But again, you need the tools—and the authority—to make those decisions yourself.

Some organizations may even prefer to run models entirely on-premise. That wasn’t a realistic option a few years ago because the models were simply too large and expensive to host. But today, that’s changing. We’re seeing more and more interest in both on-premise and private cloud deployments—especially in legal environments, where confidentiality is essential.

Will AI Replace Junior Legal Professionals?

David: So, will junior legal professionals be replaced by AI—or AI agents—in the future?

Gordian: That’s definitely a provocative question. But we’re already seeing two parallel trends.

On the one hand, I think it’s fair to say that more work will be accomplished with the same amount of input. In that sense, yes—there will be cost reductions. But on the other hand, the expectations for junior lawyers will increase. A junior lawyer working with AI will be able to deliver more than one without AI. That’s just the new reality.

At the same time, we’re also seeing other forces at play—like increased regulation, which drives up legal workload in general. And then there’s the demographic shift, which is also having an impact.

So in the end, I believe these trends might balance each other out. What I expect over the next few years is that overall productivity will rise significantly—but so will the workload. That means companies that don’t adopt AI will have a real problem. The amount of work expected from legal teams is growing, and without AI, it simply won’t be manageable with the same staffing levels.

Outlook

David: Gordian, could you give us a quick outlook on what you’re currently working on?

Gordian: Absolutely. We touched on this earlier, but one of the most challenging aspects right now is prompt creation. When users want to check new contract types, they often have to define what exactly they want to check, write the prompts accordingly, and then create matching test sets. That process is time-consuming and highly manual. This is exactly where our next-generation agents come in. We’re currently developing agents that can assist users with that entire setup. So in the near future, users will be able to check any contract—regardless of contract type—much faster and with minimal manual effort.

Another important development is quality control. Currently, we rely on our own test sets to ensure output quality. Going forward, we’ll introduce agents that can check the AI’s results automatically. This means we’ll be able to offer higher quality—while also tailoring the process to different legal departments’ risk tolerance and to the specific contract types in use.

That opens the door for something truly exciting: autonomous contract reviews. For simpler contract types like NDAs, we believe we can soon reach a level of confidence and quality where a contract is not only reviewed but also corrected—automatically. In such cases, the document could even be sent back to the counterparty without requiring manual review at all.

And as we’ve discussed, playbook management is another area where agents will help immensely. Many companies rarely update their contract guidelines or policies. With our system, agents will assist in creating or updating these internal policies—making it easier to set and maintain high-quality legal standards.

The same applies to existing playbooks. Regulations change. Industry standards evolve. And some positions may no longer be enforceable. Our AI agents will be able to detect these shifts and proactively suggest updates—ensuring your internal guidelines are always up to date.

And finally, thanks to large language models combined with agents, it’s now possible to expand to new languages much faster than before. This brings real scalability to Legal AI.

Would you like a quick overview of all your active contracts?