Chatbots were just the beginning.
AI capabilities are growing by leaps and bounds, and we’re entering an era where chatbots are now the starting point, not the end goal. Agentic AI is here and getting more sophisticated by the day.
At iManage ConnectLive 2026, I shared what it actually takes to move from generative AI that answers questions to agentic AI that gets work done. Here's what I covered in my presentation with David Malkinson, Founder at SIGNL — and what it means for your organisation.
Setting the scene
Generative AI changed the way knowledge workers access information. You ask a question, you get an output. Then you, the human, pick it up and do something with it.That model is certainly useful, but the utility of chatbots has plateaued. Buyers aren't asking whether their tools can answer questions anymore. They're asking what jobs those tools can do.
The numbers reflect this shift. Agent investment across legal tech has grown 3x in the last 12 months. The typical lead time to become agent-ready is 18 months — and the gap between organisations moving now and organisations waiting is compounding every quarter. Competitors who move first aren't just ahead; they're building a data advantage that gets harder to close over time.
Standing still is a position. But it's a costly one.
MCP makes agentic AI possible
The Model Context Protocol (MCP) was introduced by Anthropic in November 2024. At its core, it's a standard interface between an AI agent and the tools and data sources it's allowed to reach. Think of it this way: APIs are the foundation — the capabilities that exist in your software and systems. MCP sits on top of that and translates API language into something an agent can natively understand, reason about, and decide when and how to use.
This distinction matters. An API requires a developer to write, deploy, and manage code. MCP opens that capability to any user whose LLM of choice has MCP client support. It democratises access to enterprise systems without sacrificing control.
A few things MCP does well that are worth flagging: permissions are respected at query time, so the agent acts on behalf of the authenticated user. Meaning, it can only see and do what that user is permitted to see and do. And because MCP calls come through governed API endpoints, every action is audited by default (much like any action taken by a human).
To make the art of the possible concrete, consider the following workflow. From a single long-form prompt in Claude, with iManage Work, Microsoft Outlook, Asana, and Zoom MCP services enabled, an agent can read documents, extract key details, draft and send a summary email (with a human validation step), create a follow-up task in Asana, and book a Zoom meeting — all without manual handoffs.
The iManage platform has its own search agent built right in: Ask iManage. When you search with Ask iManage, it provides more than just a list of documents with the necessary keyword in the title. It reads your intent and provides synthesized results that are truly relevant to what you need. It’s a powerful upgrade to traditional search capabilities.
The architecture under every reliable agent
Knowing what’s possible with agents is great — but how do you ensure the ones you employ are effective?
Putting this three-stage architecture in place is essential:
Intent parsing is where natural language becomes structured goals. The agent interprets what the user actually wants to achieve, identifies ambiguity, and surfaces clarifying questions before acting.
Task decomposition is where goals become steps. The agent breaks the objective into discrete, sequenced actions, routes each step to the right tool or service, and manages the dependencies between them.
Loop completion is where the agent executes, checks its own results, adapts when something doesn't go as expected, handles errors without requiring user input, and — critically — knows when to escalate to a human.
Each stage is observable, auditable, and stoppable. MCP is the connective tissue; it’s the mechanism through which agents reach your tools, your data, and your systems of action.
Your institutional knowledge layer is the differentiator
Many vendors skip this part of the agentic AI conversation. They'll sell you the agent, but won't ask whether your data is ready for it.
Disciplined data is what gives agents the precision to act, not just retrieve. Permissions and governance mean that agents inherit access controls and respect them — i.e., security boundaries are enforced at the system level, not interpreted by the agent. And provenance and audit trails mean every action is traceable to its source documents, producing outputs your organisation can stand behind.
The iManage platform is built around matter-centric context. Documents, email, and tasks organised around the matter, so an agent inherits a complete picture rather than a disconnected search result. That structure matters enormously. An agent operating over well-organised, consistently-tagged matter data is fundamentally more capable than one operating over a flat file system or an unstructured inbox.
The AI delegation curve
The final piece of adopting agentic AI is knowing where you are on the AI delegation curve — and being honest about it.
We see five levels of adoption:
Level 1 — Assisted. AI in a tab. Users copy-paste into a chat interface. Every prompt is self-contained. No integration with documents or systems. This is where many organisations started.
Level 2 — Augmented. AI in the workflow. Copilots and assistants deployed organisation-wide. Document summarisation and drafting support in tools like Microsoft Word. AI suggestions reviewed before applied. Most organisations are here (whether they know it or not).
Level 3 — Orchestrated. AI with tools. Agents connected to document and data platforms, running multi-step workflows with human checkpoints at consequential moments. MCP or API-based integrations. This is where most organisations think they are.
Level 4 — Delegated. AI with autonomy. Defined classes of agent-autonomous actions run unattended, inside governed boundaries with fallbacks. Ethical walls enforced on agents, not by convention. Agent-to-agent delegation chains with governance checkpoints. Very few organisations are genuinely here.
Level 5 — Integrated. AI as infrastructure. Agents managed like associates. Governance is operations. This is the horizon worth planning to.
Roughly 80% of organisations are at Level 2 but tell themselves they're at Level 3. The reason for this gap? Governance.
There's also a structural challenge at Level 4. Today's governance frameworks assume a human initiates every action. When agents start triggering themselves autonomously — one agent handing off to another — the policy model breaks. That's not a reason to avoid agentic AI. It's a reason to build the governance infrastructure before you need it, not after.
Our advice: review your information architecture before you buy another AI capability.
The right foundation for agentic AI
The organisations pulling ahead on the AI delegation curve don't just store knowledge. They govern it, connect it, and activate it in context across every decision. iManage is the knowledge work platform that makes that possible.
To learn more about how we can support your agentic AI journey, get in touch with us today.
Paul Walker is Global Solutions Director at iManage, where he helps define and deliver solutions that bring advanced technology and AI into practical use for law firms, corporate legal teams, and other knowledge-intensive organizations. With more than 20 years’ experience across legal practice, professional services, and enterprise technology, Paul works at the intersection of data, compliance, and knowledge management. His background includes senior roles at PwC, Slaughter and May, Autonomy, and HP, giving him a unique perspective on how to turn emerging technologies into tools that streamline workflows, strengthen governance, and unlock institutional insight.