Jacob Lauritzen on AI Agents: Beyond Chat

In a recent presentation at AI Engineer Europe, Jacob Lauritzen, CTO of Legora, argued that the current reliance on chat interfaces for AI agents is insufficient for tackling complex, real-world tasks. He advocated for a shift towards more sophisticated, high-bandwidth collaboration methods between humans and AI.

Jacob Lauritzen's Perspective

Jacob Lauritzen, as CTO of Legora, a company focused on building collaborative AI workspaces for legal professionals, brings a practical, hands-on perspective to the capabilities and limitations of current AI agents. Legora itself boasts over 1,000 customers and a valuation exceeding $5.5 billion, indicating significant traction in the legal tech space. Lauritzen's talk emphasized that while AI has made strides, particularly in conversational AI, its true potential lies in its ability to handle complex, multi-step workflows that require more than just a simple question-and-answer format.

The Limitations of Chat Interfaces

Lauritzen illustrated the point with a hypothetical scenario: drafting a contract. He showed a detailed breakdown of the steps an AI agent would need to take, including research, reviewing various legal documents, identifying relevant clauses, and finally drafting the contract. When the AI agent was presented with a flawed clause, its response was to simply suggest using a precedent from another contract, a process that proved inefficient and error-prone. Lauritzen critiqued this as a demonstration of how chat-based interactions can become a bottleneck. He stated, "Chat is a one-dimensional, low-bandwidth interface." This limitation, he explained, prevents agents from effectively handling the nuances of complex tasks where multiple steps and data points are involved.

The Verifier's Rule and Task Solvability

Introducing the concept of the "verifier's rule," attributed to Jason Wei, Lauritzen explained that AI is most effective at tasks that are easily verifiable. He presented a spectrum ranging from math and coding (easily verifiable) to finance and legal work (less easily verifiable). Legal tasks, such as drafting contracts or developing litigation strategies, often fall into the "unverifiable" or "hard to verify" categories, making them challenging for AI agents relying solely on simple chat interfaces. He demonstrated how breaking down a complex task like contract drafting into smaller, more manageable, and verifiable sub-tasks, such as checking definitions or applying formatting, allows for more efficient and reliable AI execution.

Increasing Trust and Control with High-Bandwidth Artifacts

Lauritzen proposed that to overcome these limitations, humans and AI should collaborate using what he termed "high-bandwidth artifacts." These are not just simple text conversations but rather structured, interactive documents or interfaces that allow for more detailed and nuanced collaboration. He showed an example of a tabular review interface where an AI agent could flag specific clauses in multiple contracts, allowing human reviewers to quickly identify issues and make informed decisions. "Humans and agents should collaborate in high-bandwidth artifacts," he asserted. This approach allows for better control, as humans can steer the AI's work more effectively, and builds trust by making the AI's process more transparent and its outputs more easily verifiable.

The Future of AI Collaboration

The presentation concluded with a look at Legora's own platform, showcasing how it provides these richer, more interactive artifacts for legal professionals. Lauritzen emphasized that the future of AI agents lies not in simply making them more conversational, but in developing interfaces that allow for deeper, more structured collaboration, enabling agents to handle complex tasks with greater accuracy and efficiency. He believes that this convergence of UI will be key to unlocking the full potential of AI across various industries.