By Paul Walker, Global Solutions Director, iManage. Law firms have always delegated work to those with less experience, lower billing rates, and narrower scope. Agentic AI is the next evolution of this delegation, but firms must first build a robust information infrastructure to delegate safely to non-human agents. This article outlines a staged roadmap for firms to achieve trustworthy agentic AI capabilities.
Moving up the delegation gradient AI's progression within the legal industry mirrors a natural 'delegation gradient' observed in law firms. Initially, AI tools were confined to browser tabs, isolated from a firm's internal knowledge and operational workflows. This evolved into AI acting as an integrated assistant within common legal tools, aiding with specific tasks. The current phase sees AI beginning to orchestrate more complex, multi-step tasks across various firm systems, demonstrating a higher level of integration and autonomy. The next significant leap is the emergence of delegated agents capable of completing entire jobs, seeking human approval only at crucial junctures. Ultimately, this will lead to fully integrated agents that can proactively act on triggers without direct human instruction. The majority of law firms today find themselves somewhere along this gradient, typically navigating the stages between basic AI assistance and early orchestration, highlighting the significant journey ahead for many in harnessing the full potential of agentic AI. Stage one – get the information architecture in order The foundational and most critical step for any law firm aiming for successful implementation of agentic AI is to meticulously organize its information architecture. Without a robust and centralized system, valuable work product remains scattered across disparate locations such as individual inboxes, personal network drives, and various ad hoc collaboration platforms. This disorganization makes it highly probable that an AI agent will either overlook crucial material or inadvertently access outdated or incorrect versions of documents. Therefore, the initial and paramount action is to establish a single, comprehensive matter-centric record where every document, its versions, redlines, and accepted changes are systematically stored. This centralized repository inherently captures and preserves institutional judgment. Following this, files must be thoroughly cleaned and accurately tagged. AI-driven enrichment tools can significantly expedite this process, employing sophisticated classification and extraction models to assign attributes like governing law or deal type to legacy documents that were never formally structured. This transformation effectively converts a chaotic archive into a precisely navigable resource for AI agents, dramatically enhancing both their speed and accuracy in information retrieval and processing. Stage two – extend governance to non-human users Just as a newly hired trainee lawyer is not immediately granted unrestricted access to all client files upon their first day, AI agents must be subjected to an equally rigorous system of governance. It is imperative that when a lawyer deploys an AI agent, that agent automatically inherits the precise permissions and access rights of the initiating lawyer. This ensures that crucial ethical walls and information barriers, which are fundamental to legal practice, are strictly enforced at the point of data retrieval. In practical terms, this necessitates treating each AI agent as a distinct identity within the firm’s security framework, endowed with defined, least-privilege access that is inextricably linked to the human user on whose behalf it is operating. Furthermore, the aspect of auditability is non-negotiable and equally important. If an AI agent contributes to any legal advice or work product, the firm must possess the capability to meticulously reconstruct exactly what information the agent was exposed to and precisely what actions it undertook. Without such a detailed and accountable record, a firm would be unable to provide a verifiable basis for its own legal advice, potentially leading to significant compliance and ethical risks. Stage three – deliver context, not just content Simply connecting AI tools to an ever-growing number of repositories, even through modern open standards like the Model Context Protocol (MCP), is not sufficient to enhance an agent's judgment or efficacy. While widening an agent's access to information, it does not inherently improve its ability to reason or make sound decisions. Large language models have finite "context windows," and a known challenge is that they tend to under-weight or even overlook material that is buried in the middle of excessively long inputs. Consequently, indiscriminately overwhelming an AI agent with a deluge of loosely relevant documents not only degrades its reasoning capabilities but also unnecessarily inflates token costs, making the operation inefficient and expensive. The ultimate objective is to provide governed context. This means that information retrieval should be intelligently ranked by its direct relevance to the specific matter at hand. Rather than feeding entire files, the system should return structured excerpts and pertinent metadata. Crucially, these retrieved documents should be seamlessly linked to the associated parties, communications, and relationships that provide them with their full meaning. This deepens the contextual understanding for the AI agent, significantly improving its capacity to accurately and effectively carry out its assigned tasks. Stage four – delegate action within guardrails When firms begin to truly delegate tasks to AI agents, these initial deployments are most prudently focused on monitoring functions rather than direct modification or creation, particularly in high-stakes areas. For instance, an agent could be tasked with continuously tracking evolving regulatory changes, automatically identifying every client agreement potentially affected by these changes, and proactively recommending necessary repapering before a client even realizes the need. Similarly, obligations tracking and conflicts checking are prime candidates for such delegated AI, transforming traditionally reactive and time-consuming work into highly efficient, proactive client service. The paramount principle guiding this delegation is that human supervision must always be proportionate to the inherent risk of the task. Any output or contribution from an AI agent that is client-facing must undergo the same rigorous review process as a draft prepared by a junior lawyer, typically requiring review by a partner. Furthermore, it is essential to establish clear and robust escalation points. These mechanisms ensure that control is immediately handed back to a human lawyer whenever a task exceeds the predefined authority or capabilities of the AI agent, preventing potential errors or overreach and maintaining human oversight where it is most critical. A familiar discipline for a new technology The rigorous discipline and structured approach demanded by the effective and safe implementation of agentic AI are, in essence, not fundamentally new to the legal profession. Law firms have a long-standing understanding that the practice of delegation—whether to junior lawyers, paralegals, or administrative staff—is only as secure and reliable as the overarching structure built around it. This traditional structure has always necessitated clear definitions of scope for delegated tasks, ensuring appropriate access levels to sensitive information, establishing proper supervision protocols to guide and review work, and maintaining an accountable record of actions taken. What has profoundly changed in the era of agentic AI is the nature of the entity performing the delegated work; it is now a non-human agent rather than a person. Firms that proactively invest now in building and refining the robust information infrastructure required to support this technological shift will be uniquely positioned to confidently assign meaningful and complex work to AI agents. Conversely, those firms that neglect to fortify their foundational infrastructure will likely discover that their most advanced and costly AI investments are severely constrained and undermined by the inherent weaknesses in their underlying data and governance systems, leading to a challenging and inefficient adoption pathway.
Moving up the delegation gradient
AI's progression within the legal industry mirrors a natural 'delegation gradient' observed in law firms. Initially, AI tools were confined to browser tabs, isolated from a firm's internal knowledge and operational workflows. This evolved into AI acting as an integrated assistant within common legal tools, aiding with specific tasks. The current phase sees AI beginning to orchestrate more complex, multi-step tasks across various firm systems, demonstrating a higher level of integration and autonomy. The next significant leap is the emergence of delegated agents capable of completing entire jobs, seeking human approval only at crucial junctures. Ultimately, this will lead to fully integrated agents that can proactively act on triggers without direct human instruction. The majority of law firms today find themselves somewhere along this gradient, typically navigating the stages between basic AI assistance and early orchestration, highlighting the significant journey ahead for many in harnessing the full potential of agentic AI.
Stage one – get the information architecture in order
The foundational and most critical step for any law firm aiming for successful implementation of agentic AI is to meticulously organize its information architecture. Without a robust and centralized system, valuable work product remains scattered across disparate locations such as individual inboxes, personal network drives, and various ad hoc collaboration platforms. This disorganization makes it highly probable that an AI agent will either overlook crucial material or inadvertently access outdated or incorrect versions of documents. Therefore, the initial and paramount action is to establish a single, comprehensive matter-centric record where every document, its versions, redlines, and accepted changes are systematically stored. This centralized repository inherently captures and preserves institutional judgment. Following this, files must be thoroughly cleaned and accurately tagged. AI-driven enrichment tools can significantly expedite this process, employing sophisticated classification and extraction models to assign attributes like governing law or deal type to legacy documents that were never formally structured. This transformation effectively converts a chaotic archive into a precisely navigable resource for AI agents, dramatically enhancing both their speed and accuracy in information retrieval and processing.
Stage two – extend governance to non-human users
Just as a newly hired trainee lawyer is not immediately granted unrestricted access to all client files upon their first day, AI agents must be subjected to an equally rigorous system of governance. It is imperative that when a lawyer deploys an AI agent, that agent automatically inherits the precise permissions and access rights of the initiating lawyer. This ensures that crucial ethical walls and information barriers, which are fundamental to legal practice, are strictly enforced at the point of data retrieval. In practical terms, this necessitates treating each AI agent as a distinct identity within the firm’s security framework, endowed with defined, least-privilege access that is inextricably linked to the human user on whose behalf it is operating. Furthermore, the aspect of auditability is non-negotiable and equally important. If an AI agent contributes to any legal advice or work product, the firm must possess the capability to meticulously reconstruct exactly what information the agent was exposed to and precisely what actions it undertook. Without such a detailed and accountable record, a firm would be unable to provide a verifiable basis for its own legal advice, potentially leading to significant compliance and ethical risks.
Stage three – deliver context, not just content
Simply connecting AI tools to an ever-growing number of repositories, even through modern open standards like the Model Context Protocol (MCP), is not sufficient to enhance an agent's judgment or efficacy. While widening an agent's access to information, it does not inherently improve its ability to reason or make sound decisions. Large language models have finite "context windows," and a known challenge is that they tend to under-weight or even overlook material that is buried in the middle of excessively long inputs. Consequently, indiscriminately overwhelming an AI agent with a deluge of loosely relevant documents not only degrades its reasoning capabilities but also unnecessarily inflates token costs, making the operation inefficient and expensive. The ultimate objective is to provide governed context. This means that information retrieval should be intelligently ranked by its direct relevance to the specific matter at hand. Rather than feeding entire files, the system should return structured excerpts and pertinent metadata. Crucially, these retrieved documents should be seamlessly linked to the associated parties, communications, and relationships that provide them with their full meaning. This deepens the contextual understanding for the AI agent, significantly improving its capacity to accurately and effectively carry out its assigned tasks.
Stage four – delegate action within guardrails
When firms begin to truly delegate tasks to AI agents, these initial deployments are most prudently focused on monitoring functions rather than direct modification or creation, particularly in high-stakes areas. For instance, an agent could be tasked with continuously tracking evolving regulatory changes, automatically identifying every client agreement potentially affected by these changes, and proactively recommending necessary repapering before a client even realizes the need. Similarly, obligations tracking and conflicts checking are prime candidates for such delegated AI, transforming traditionally reactive and time-consuming work into highly efficient, proactive client service. The paramount principle guiding this delegation is that human supervision must always be proportionate to the inherent risk of the task. Any output or contribution from an AI agent that is client-facing must undergo the same rigorous review process as a draft prepared by a junior lawyer, typically requiring review by a partner. Furthermore, it is essential to establish clear and robust escalation points. These mechanisms ensure that control is immediately handed back to a human lawyer whenever a task exceeds the predefined authority or capabilities of the AI agent, preventing potential errors or overreach and maintaining human oversight where it is most critical.
A familiar discipline for a new technology
The rigorous discipline and structured approach demanded by the effective and safe implementation of agentic AI are, in essence, not fundamentally new to the legal profession. Law firms have a long-standing understanding that the practice of delegation—whether to junior lawyers, paralegals, or administrative staff—is only as secure and reliable as the overarching structure built around it. This traditional structure has always necessitated clear definitions of scope for delegated tasks, ensuring appropriate access levels to sensitive information, establishing proper supervision protocols to guide and review work, and maintaining an accountable record of actions taken. What has profoundly changed in the era of agentic AI is the nature of the entity performing the delegated work; it is now a non-human agent rather than a person. Firms that proactively invest now in building and refining the robust information infrastructure required to support this technological shift will be uniquely positioned to confidently assign meaningful and complex work to AI agents. Conversely, those firms that neglect to fortify their foundational infrastructure will likely discover that their most advanced and costly AI investments are severely constrained and undermined by the inherent weaknesses in their underlying data and governance systems, leading to a challenging and inefficient adoption pathway.
