What are agents?

In the evolving landscape of legal AI, the term "agents" has become very relevant. But what exactly are agents, and why should legal professionals pay attention to them?

Understanding AI agents

AI agents are autonomous or semi-autonomous systems designed to perform specific tasks. Unlike traditional GenAI models that usually respond directly to a prompt, agents can carry out complex, multi-step tasks with minimal human intervention. An AI Agent is an advanced system that perceives, decides, and acts within an environment. It can use a GenAI model as part of its reasoning but also incorporates memory, decision-making, and interaction capabilities.

Think of an agent as an assistant that not only responds directly to requests, but is also capable of, browsing a database, reviewing its own output, do complex maths or even generating more context before answering your question.

Extra: Generating more context is actually what reasoning models do before answering your question, generating a chain-of prompts (also called chain-of-thought prompting), and you could also call the reasoning models an Agentic framework around models like GPT4o.

Practical applications for lawyers

1. Grid Review (multiple legal documents undergo different types of AI processing in parallel)

Traditional AI: Lawyers need to manually analyze dozens of documents.

Agentic AI:

• Autonomous Execution: Agents trigger actions on multiple documents. Different agents are used for different tasks such as summarizing, re-prompting, fact extraction or text reviewing.
• Multi-Agent Collaboration: Reviewing, extraction, and summarization agents interact dynamically instead of being separate queries.
• Real-Time Refinement: Agents adjust prompts, correct errors, and validate citations automatically.

💡 Outcome: With Agentic AI, the grid review process becomes autonomous, adaptive, and scalable, reducing manual effort and increasing legal review accuracy.

2. Contract Drafting (Proactive & Context-Aware)

Traditional AI: A lawyer asks an AI to draft a non-compete clause.

Agentic AI:

• Researches relevant case law from a legal or contract database automatically before drafting.
• Analyzes the contract’s full context (e.g., industry, jurisdiction, company policies).
• Drafts the clause while incorporating tailored language for enforceability.
• Simulates a negotiation, generating alternative versions based on risk tolerance.
• Suggests revisions based on legal trends and opposing party arguments.

💡 Outcome: The agent handles the entire drafting-research-negotiation workflow, allowing the lawyer to finalize rather than start from scratch.

3. Contract Review & Lifecycle Management (Autonomous Updates & Alerts)

Traditional AI: AI highlights risky clauses when manually prompted.

Agentic AI:

• Monitors all contracts within a firm's system.
• Flags changes in relevant laws and automatically gives suggestions for updating contract templates.

• Identifies risks in client agreements and drafts suggested amendments.

• Initiates review workflows, assigning tasks to lawyers or clients if necessary.

• Negotiates with counterparties using automated chat-based reasoning.

💡 Outcome: The agent proactively tracks contracts, legal risks, and compliance—not just answering queries but improving contract governance.

4. Legal Research (Persistent Monitoring & Summarization)

Traditional AI: Lawyers ask AI to find relevant case law.

Agentic AI:

• Continuously scans legal databases, government websites, and court rulings.

• Identifies legal shifts relevant to ongoing cases or specific client contracts.

• Sends real-time alerts to lawyers about new case law affecting arguments.

• Drafts memos and legal briefs summarizing implications and suggesting legal strategies.

• Cross-references laws across jurisdictions for multinational clients.

💡 Outcome: Lawyers receive pre-digested legal insights instead of manually searching case law.

Challenges and considerations

While agents offer significant benefits, the biggest challenge is hallucination. Even if you have a 95% accuracy, setting up 5 agents in a row, results in only a 77% (0.95^5). Hallucinations can trigger incorrect decisions: one hallucinated task can consequently influence the quality of the subsequent tasks an AI agent needs to execute. Furthermore, it is extremely challenging to give agents access to different applications and databases. And finally, lots of tools and software are not build for AI agents (how ironic, you now have to fill in Captcha’s for your AI to lie to the software you’re using that the AI agent is not a robot... which it is).

So, how can agents assist you in enhancing your legal practice?

AI agents present an opportunity to redefine how legal professionals work. The biggest opportunity is to automate certain tasks, but if you ask me, the biggest win will be in the quality improvement of AI systems. Even just asking an AI agent to check the AI’s output increases its performance!

Traditional AI: user asks question > AI generates output

Simple agent: user asks question > AI generates output, but before sending it to the user, it reviews its own output > then improves it based on its review > and then sends it to the user.

Lastly, modern agentic frameworks can spot gaps in an AI model’s knowledge and decide to browse a database or the internet to fill that knowledge gap. This is actually already what agents like Deep Search from OpenAI work can do, very cool!