From AI Tools to Agentic AI – A Roadmap

By Paul Walker, Global Solutions Director, iManage.

Law firms have always delegated work to those with less experience, lower billing rates, and narrower scope: partners pass work to associates, associates pass it to trainees and paralegals. Agentic AI is the next form of that delegation, but it’s an open question whether firms have built the infrastructure to delegate safely to a non-human agent.

Most have not, and a law firm’s success with agentic AI will depend not so much on the sophistication of the technology it has purchased as the quality of the information infrastructure supporting it. To achieve trustworthy agentic AI capability, firms need a clearly staged roadmap, one that can safely guide them forward while avoiding any wrong turns or dead ends.

Moving up the delegation gradient

AI’s journey in the legal space has followed something of a ‘delegation gradient’. AI began in a browser tab, cut off from the firm’s knowledge and workflows. It then moved into everyday tools as an assistant, and is now starting to orchestrate multi-step tasks across firm systems. The next phase of the gradient involves delegated agents that complete whole jobs and seek approval at key moments, and eventually there will be integrated agents that act on triggers without being asked.

Most firms sit somewhere between assistance and orchestration. How do they ensure they have built the infrastructure to advance along the gradient and delegate safely to a non-human agent? The path forward is best approached in several distinct stages.

Stage one – get the information architecture in order

If work product is scattered across inboxes, personal drives, and ad hoc collaboration sites, an agent will miss critical material or draw on the wrong version of it. The first step is a single matter-centric record. Every version, redline, and accepted change in that record captures a decision, making it a store of institutional judgement.

The next step is cleaning up and tagging those files – and AI-driven enrichment can accelerate the work. Classification and extraction models can tag legacy documents with attributes that were never captured in a structured way, such as governing law or deal type, turning an archive into something an agent can navigate with precision, improving speed and accuracy.

Stage two – extend governance to non-human users

A trainee does not receive the keys to every client file on day one, and agents should not either. When a lawyer initiates an agent, it should inherit exactly that lawyer’s permissions, with ethical walls and information barriers enforced at the point of retrieval. In practice, that means treating each agent as an identity with defined, least-privilege access, tied to the user on whose behalf it acts.

Auditability matters just as much. When an agent contributes to advice, the firm must be able to reconstruct what it was shown and what it did. Without that record, a firm cannot account for the basis of its own advice.

Stage three – deliver context, not just content

Connecting AI tools to more repositories – increasingly through open standards such as the Model Context Protocol (MCP) – widens an agent’s reach but does not improve its judgement. Context windows are finite, and models tend to under-weight material buried in the middle of long inputs.

Flooding an agent with loosely relevant documents degrades its reasoning and inflates token costs. A well-built MCP connection exposes the firm’s information architecture so the agent knows where to look.

The goal is governed context: retrieval ranked by relevance to the matter at hand, returning structured excerpts and metadata rather than whole files, and linking documents to the parties, communications, and relationships that give them meaning. These all deepen the context further, aiding the agent in carrying out its task.

Stage four – delegate action within guardrails

The first truly delegated agents are likely to monitor rather than modify. An agent tracking regulatory change, for instance, could identify every affected client agreement and recommend repapering before the client asks. Obligations tracking and conflicts checking are similar candidates, turning reactive work into proactive service.

What’s important is that supervision should always be proportionate to risk, with anything client-facing reviewed as a partner would review a junior lawyer’s draft – and there should be clear escalation points that hand control back to a lawyer when a task exceeds the agent’s authority.

A familiar discipline for a new technology

The discipline agentic AI demands is not new. Firms have always understood that delegation is only as safe as the structure around it: clear scope, appropriate access, proper supervision, and an accountable record. What has changed is who – or what – is now doing the work.

Firms that invest now in the information infrastructure to support this shift will be able to hand meaningful work to agents with confidence. Those that do not will find their most sophisticated AI investments constrained by the weakest part of their foundation, with a bumpy road ahead.

To learn more about iManage, see here.

—

[ This is a sponsored thought leadership article for Artificial Lawyer by iManage. ]