Legal Tech Unicorns: Reliable AI Agents for Legal Practice

Denisa Kopandi returns to Version Up to discuss what happens when legal AI agents meet the uneven, iterative work of an actual transaction. Drawing on her experience testing legal AI tools and building workflows, she explains why a sequence that looks convincing in a demo can be difficult to trust in practice.

We distinguish between agents working behind a chatbot’s answer and workflows intended to carry a legal matter through multiple stages. Denisa describes the gap between a neat progression from intake to analysis to drafting and the reality of documents arriving in batches, incomplete information, and clients changing the structure or timing of a deal.

A central theme is the legal analysis that happens before anything gets written down. A lawyer may open an apparently routine document and notice something that changes the entire assessment. That makes it difficult to specify every task or decide in advance where an agent should stop for review. Checking the final legal analysis may not reveal that an earlier agent overlooked a document or extracted the wrong information.

We explore a more manageable approach: specialized agents for narrower tasks, with lawyers deciding when to use them and how the results fit together. Examples include assessing a merger notification threshold or preparing a request for information using what is already known about the client. The question is how much useful work can be delegated without creating a verification burden that cancels out the time saved.

Denisa also shares an experiment in which an agent began deleting earlier versions of a due diligence report, leaving her to transfer the analysis manually. We discuss what that experience says about reliability, alongside the difficulty of trusting a batch of contract analyses after finding an error in one.

There is a training dimension too. Denisa asked interns to use AI for research and submit their full conversations, so she could review how they questioned the model. We consider whether narrower tasks and visible review can help junior lawyers develop judgment, and how quickly improving tools affect the decision to invest in building a workflow.

Denisa Kopandi is a senior associate at bpv GRIGORESCU ȘTEFĂNICĂ in Romania, working across technology, competition, and commercial matters. Alongside her client work, she evaluates and implements legal AI tools and explores how they can support specific legal tasks and the development of junior lawyers. You can follow her on LinkedIn.

* * *

Version Up Members is a free, invitation-only community for lawyers and legal professionals putting AI to work at law firms and legal departments. Request an invitation: versionup.ai/members