AI agents are powerful because they can reason, decide and call tools. That is also why they need boundaries. In JD Edwards, the best boundary is not a raw database connection and not a screen-scraping shortcut. It is an Orchestration: named, typed, approved and auditable.
Do not ask an AI agent to improvise your ERP integration. Give it the business services you already trust.
Why raw access is the wrong start
JD Edwards data is not just data. It is business logic, security, semantics, dates, decimals, codes, relationships and process. Raw access may answer a question, but it can also bypass the very rules that make the answer trustworthy.
For read-only analytics there are controlled patterns, including JDE-aware ODBC. For agentic work that changes or triggers business processes, raw access is the wrong shape. The right shape is an approved action.
Orchestrations are agent tools
An Orchestration already has what an AI tool needs: a name, a description, a schema, required inputs, optional inputs, predictable outputs and a security context. That is why it maps so naturally to agent workflows.
- Discovery. The agent can see which Orchestrations are available to the current user.
- Introspection. It can learn the input schema instead of guessing.
- Execution. It can call the approved operation and receive a structured result.
The BrainStorm model
BrainStorm is built around this idea. The user asks a question or requests an outcome. The system can retrieve supporting documentation, cite sources, identify the right Orchestration and run it under the user’s own JDE identity. Conversation becomes the interface, but JDE remains the control plane.
MCP as the connector
AI MCP Servers expose those JDE operations to agents as small, typed tools. The agent does not get an open shell. It does not get arbitrary JDE reach. It gets a curated tool surface: enough power to help, not enough freedom to surprise.
The right question
The first question should not be “Can AI access JDE?” The better question is: “Which JDE actions should this user, in this context, be allowed to request conversationally?” Once you ask the question that way, Orchestrations stop being a technical feature and become an AI governance strategy.
Start with the Orchestrations you already trust. Then expose them as typed MCP tools or through BrainStorm, instead of inventing a parallel integration layer.