Agentic AI begins with Model Context Protocol

Out of all the exciting projects on my plate right now at iManage, one stands out above the rest: Model Context Protocol (MCP). 

For those just catching up on this development, Model Context Protocol is an open, universal standard created by Anthropic. It enables AI models (like Claude or GPT-4) to securely connect with enterprise applications and workflows via an orchestration layer such as Microsoft Copilot or ChatGPT Enterprise. 

With MCP, AI models can plug into existing systems using a consistent protocol, minimising friction and IT overhead. It eliminates the need for custom integrations between AI tools and enterprise systems, streamlining deployment and reducing time-to-value.

MCP is also the first step toward deploying truly agentic AI at work, since it allows models to safely plug into your secure data.

I recently co-hosted a webinar with Nikki Shaver, CEO & co-founder, Legaltech Hub, entitled Innovation update: The knowledge foundation behind agentic AI. During the presentation, we explored why MCP is generating so much interest across the legal ecosystem, how iManage is approaching it, and what it means for the future of work. Read on for more. 

Better than an integration 

At iManage, we first started looking at MCP around Q1 of 2025, after a few of our clients asked us about leveraging it on our platform. Once our engineering teams started building it, we quickly saw its potential.

The big lightbulb moment for me was during prep for ILTACON last year. I was in a meeting room in Chicago when the engineering team showed MCP to me and some of my product management colleagues.

We started trying some different use cases, and the looks on some people’s faces as we realised what was possible were truly legendary. It’s one of those moments that I’ll always remember. 

On the surface, MCP can sound deceptively simple. Technically speaking, it’s a wrapper around APIs that allows AI systems to discover and interact with tools and data using natural language. The real shift is conceptual.

Historically, accessing data across systems required rigid, pre-built integrations. Those integrations took time, engineering effort, and ongoing maintenance. Users would go to their IT teams, IT would build and test an integration, and months would pass. By the time something was built, the original use case often no longer mattered.

MCP changes that dynamic. It allows AI to interact directly with systems of record — like document management, knowledge platforms, task systems, or CRMs — without moving data and without requiring users to understand APIs at all. You can ask questions in plain language, and the AI can securely retrieve, analyse, or act on information where it already lives.

This democratisation of the API is what makes MCP so powerful. 

The importance of data staying put 

To really understand the value of MCP, it’s helpful to consider the way things work without it. If you want to leverage AI technology, you are forced to copy your data out of trusted sources and paste it into the external AI system. This is something that many organisations allow because they want their teams to get the benefit of AI, but it presents a number of security concerns.

At iManage, our approach is grounded in a simple principle: your data should stay in your system of record. With MCP, documents never leave the iManage platform, so security policies, information barriers, and ethical walls are respected by default. Only the minimum required text layer is shared with an AI model, and even that depends on how customers configure their AI environment. This means firms can adopt AI without compromising the structures they rely on to manage risk, compliance, and professional responsibility.

Agentic AI: Enormous potential, important questions

A big part of our discussion focused on agentic AI — the idea that AI systems will increasingly do more than just answer questions. They’ll take initiative, execute tasks, and operate with some degree of autonomy.

That potential is exciting, but it also raises hard questions.

In my mind, there are three broad modes of AI use, listed below from least to most challenging from a security standpoint:

1. Human-led interactions, where the AI acts strictly on behalf of a user.
2. Human-in-the-loop workflows, where AI completes steps but requires approval.
3. Fully agentic flows, where systems run autonomously in the background.

It’s that third category where governance, security context, and liability really become complex. What permissions should an agent have? How do you prevent unintended actions? How do you preserve ethical walls or client restrictions when no human is directly involved?

These are questions for the entire industry, and I can’t claim to have the answers for them yet (although I do have some ideas). What matters now is that firms recognise these questions and start factoring them into their AI strategies. 

MCP and beyond 

The beauty of MCP is that it allows you to use multiple AI tools while maintaining your data safely in the iManage platform. While agentic AI is still being developed, our adoption of MCP ensures that iManage clients will be able to take advantage of it in the future. 

There’s much more to be found in the full webinar recording, including discussions of the changing nature of the legal tech stack and the evolving definition of the matter file. Listen to the webinar on-demand now to hear it all.