Ed. note: This article first appeared in ILTA’s Peer-to-Peer Magazine.
Despite widespread predictions that 2025 would usher in the era of agentic AI, practical deployment remains in the very early stages. One critical barrier was that AI agents could not communicate effectively with each other or with external systems. To a certain extent, Model Context Protocol (MCP) — the open-source standard enabling seamless AI model integration with tools and data sources — has addressed this barrier, but progress remains slow. As the next year unfolds, what might the evolution of AI agents within the legal sphere look like?
AI agents already exist within legal settings; however, they are a bit limited, at the moment, in what they can actually do. The primary focus of AI agents in legal settings is creating vaults, or knowledge repositories of contracts, documents, or other content that an AI agent can dip into, review, and then return answers to a legal professional. In essence, the AI agent is performing basic search and retrieval.
While this functionality is basic, that does not mean it lacks utility; quite the contrary. The ability to efficiently perform and complete this type of narrow task can reduce the time required for some particularly onerous legal work from hours to minutes.
As mentioned earlier, one of the things preventing agents from moving beyond basic search and retrieval to more agentic workflows was the inability to communicate with different systems and hand off the next task in that workflow to them. MCP has gone a long way towards making these handoffs feasible, removing one of the significant barriers preventing the realization of truly agentic AI.
Better yet, studio tools have emerged that allow end users to easily create AI agents without coding them from scratch. As recently as three months ago, that was not the case. You needed someone with coding expertise to help develop an AI agent. The fact that this capability is now in the hands of non-technical legal professionals and end users removes another barrier, further clearing the runway.
Shifting to Multi-step Processes
With different systems easily able to communicate with one another thanks to MCP and the possibility of developing agentic workflows without ever having to see a single line of code, we are rapidly entering an era where AI agents can complete more and more steps of a legal workflow.
Here is what this shift might look like. Many organizations today have channels that enable employees to interact with the legal department and ask questions, such as Slack channels, online forms, or dedicated email aliases. Regardless of the channel, the process is the same: a legal question is submitted, and a human being responds.
Taking that process one step further involves layering in an AI chatbot that can answer questions. The AI engine has been seeded with all the appropriate policies, contracts, and other high-quality content, enabling it to provide intelligent responses to people when they ask legal questions. Moreover, a human stays in the loop. The AI routes the answer to a human in the legal department for validation and approval before sending it. That is an agentic workflow because of its multi-step nature. The AI consults a knowledge source, receives an answer, and routes it to a human in the loop for validation. This relatively simple use case makes it easy to envision the far more sophisticated agentic workflows that could emerge.
Agentic AI Creates Stellar Contracts
One of the most near-term applications of an AI agent — primarily when used within the framework of a knowledge vault or repository — is monitoring contracts and tracking the obligations they contain. This process starts in a familiar enough manner: a legal professional interacts with the AI agent/chatbot and asks it to monitor a pile of contracts and look for specific items, such as certain kinds of problems, specific dates, or eventualities. Maybe they want to look for currency exchange fluctuations that might affect the organization’s supply chain, or termination terms and conditions that are far outside of the normal range. Or maybe they want the AI agent to monitor things like expiry and take action as those expiries approach.
Before, AI might have been able to identify that an expiry was approaching and notify the human in the loop. Agentic AI can go several steps further down the workflow path. It knows that an expiry is approaching, then drafts an email that serves as a starting point for negotiating a renewal, routes it to a human for approval, and sends it to the counterparty. AI agents can draft a new contract and route it to a lawyer for a final review. Once approved by the lawyer, the contract is sent to the other party. Once the countersigned contract is returned, the AI agent updates the contract base, and then — bringing things full circle — starts monitoring that new contract. In this way, the AI agent can start chipping away at more and more of the pieces of a legal workflow while still keeping a human in the loop. That future is now nearer to reality than ever before.
From Promise to Practice
A useful parallel comes from the automotive world, of all places. Autonomous vehicles are often described in terms of levels of autonomy, ranging from basic driver assistance to full self-driving capability. At the lowest levels, the car supports the driver with alerts or minor corrections. As autonomy increases, the vehicle begins to handle more complex tasks — such as lane changes, adaptive cruising, and even navigating entire routes — while still requiring human oversight. At the highest levels, the system assumes nearly all responsibility, with humans stepping in only in rare circumstances.
Agentic AI workflows in the legal space are evolving in a strikingly similar way, progressing over time toward ever-more advanced autonomous actions while still keeping humans in the loop. This human oversight will continue to play a critical role as a safeguard, ensuring quality and building confidence in these emerging systems. While the pace of progress may feel slower than early predictions suggested when the buzz around agentic AI first started reverberating in the legal sphere, the trajectory is unmistakable: legal AI agents are steadily moving from promise to practice, reshaping workflows in ways that will soon feel as natural and inevitable as other technological transformations the legal space has already experienced, embraced, and embedded in their operations.
Paul Walker is the global solutions director at iManage.