AI-READY INTEGRATION
JD Edwards, Connected to AI. The Right Way.
Most AI tools can connect to JDE in some form. The question is whether that connection is safe, controlled, and actually reliable in production. CIP7's native MCP server is built to answer that question. Task-specific by design, it exposes only what each agent needs, runs every request as the authenticated user, and logs every action end to end.
Prefer to share more detail? Use our full contact form instead.
We care about your data in our privacy policy
THE BASICS
AI agents are capable. But they need context, constraints, and a secure way in.
AI assistants and agents can reason, interpret unstructured information, and make decisions in ways traditional software cannot. But out of the box, they have no knowledge of your business systems, your data, or how JD Edwards works. To do something useful with JDE, an AI needs three things: a structured way to ask questions and trigger actions, enough context to understand what it is working with, and clear constraints so it does not go off and do something you did not intend.
Model Context Protocol (MCP) provides that interface. It gives AI tools a defined, governed connection to an external system. MCP describes what an agent can do, which data it can access, how to interact with that data, and what remains out of bounds. It turns a capable but uninformed AI into a focused agent that can operate safely in your environment.
All three are built into CIP7's native MCP server. No adapters or additional integration layer required.
HONEST BY DESIGN
AI agents are not a replacement for every process. But for the right ones, they are transformative.
One of the most important things to understand about AI agents is where they genuinely add value and where they do not. Not every task is suited to an agent, and deploying AI in the wrong places creates risk, cost, and unpredictability.
AI agents are non-deterministic. Give one the same input twice and it may take a different path to reach the same outcome, or reach a different outcome entirely. That makes them poorly suited to strict, rules-based, high-volume transactional workflows where you need exactly the same result every single time. Automatically approving purchase orders, processing invoices without human oversight, or executing financial transactions based solely on AI judgement are examples of where the risk outweighs the benefit. The technology is impressive, but the consequences of a single wrong decision are too significant.
Where agents add value
Where AI agents genuinely excel is in tasks that look like what a skilled human does: taking variable, unstructured input, reasoning across multiple data sources, applying judgement, and producing a structured output ready for review. These tasks often require hours of manual work because no two inputs are alike and traditional integrations cannot accommodate the variation.
Where agents work well
Unstructured input, structured output
Taking source data that varies every time, such as spreadsheets, site surveys, documents, and legacy exports, and producing a validated, formatted output ready to load into JDE. The AI does the heavy lifting. A human reviews before anything is committed.
Where agents work well
Cross-system reasoning and lookup
Tasks that require a human to look at source data, cross-reference multiple things in JDE, apply business rules, and make a judgement call. The AI can do all of that in seconds, as the authenticated user, with full traceability.
Where agents need guardrails
Not a set-and-forget workflow engine
Agents are not deterministic workflow systems. They should not be left to execute high-stakes transactional tasks without human checkpoints. CIP7 is designed around human-in-the-loop principles for exactly this reason.
HOW IT WORKS
Every agent gets exactly what it needs. Nothing more.
CIP7's native MCP server is built around a principle of tight, intentional context. Rather than exposing a broad interface into JDE and letting the AI decide how to use it, every MCP server you configure in CIP7 is purpose-built for a specific task or agent. The agent can only see and call what you have explicitly defined for it.
Three components make this work together:
01 · PERMITTED ACTIONS
Tools
The functions, queries, and orchestrations the agent is permitted to call. Each tool is scoped to the task and annotated with field-level metadata covering descriptions, required and optional inputs, valid data types, and business context. The AI understands not just what a field is called, but what it means and how to use it correctly.
Cantara handles JDE-specific formatting automatically, including business unit padding, date formatting, and UDC lookups. The agent does not guess. It knows.
02 · SESSION CONTEXT
Skills
Every time an agent session starts, it starts fresh. AI agents do not carry memory from one session to the next. Each run begins with only its base knowledge.
Skills are the session-level context that brings the agent up to speed for the task at hand: guidance on the environment it is operating in, what a successful outcome looks like, what it should prioritise, and what it should avoid. Skills are rewritten each session, keeping the agent consistent and focused every time it runs.
03 · REFERENCE KNOWLEDGE
Resources
Supporting reference data and knowledge sources the agent can consult to inform its decisions. Field definitions, business rules, lookup tables, or any additional context that helps the agent reason correctly about the task.
Resources supplement the tools without expanding what the agent is permitted to do.
DESIGNED WITH INTENT
The AI does the work. Your team makes the call.
Every CIP7 MCP workflow is designed around a deliberate principle: the AI prepares, validates, and structures. Humans review and approve before anything is committed to JDE.
This is not a limitation of the technology. It is a deliberate design choice that reflects how AI agents actually work and where they should sit in an enterprise process. An agent that processes data and produces a reviewed, structured output ready for a human decision is vastly more valuable and far safer than one that executes autonomously without a checkpoint.
Every CIP7 MCP workflow should include a human review stage before data is written to JDE. This is the recommended design pattern and the approach Rinami follows in all implementations. The agent produces its output. A person reviews it. If it looks right, it goes through. If something looks off, you can see exactly what the agent did, what data it referenced, and where it diverged from what you expected. Everything is logged.
AI agentPrepare, validate, structure
Human checkpointReview and approve
JD EdwardsCommit governed output
“The agent does the heavy lifting across thousands of records in seconds. The human makes the final call. That is the right division of labour.”
THE DIFFERENTIATOR
There are other JDE MCP solutions. Here is why they are not the same.
The JDE partner ecosystem has produced a number of MCP implementations in the past year. Most of them work by wrapping the AIS server or Orchestrator and exposing orchestration discovery endpoints. The AI then searches for orchestrations by keyword and picks the one it thinks is most relevant. In a controlled demo environment with a handful of orchestrations, that looks fine. In a production environment with dozens of orchestrations built up over years, it creates real problems.
Broad by defaultGeneric JDE MCP Wrappers
Constrained by designCIP7 Native MCP Server
Discovery
The agent searches for orchestrations by keyword and picks one. In production with many orchestrations, the agent will not always pick correctly. There is no guarantee it will take the same path twice.
Explicit scope
You define exactly which orchestrations and functions are available to the agent. It cannot discover or call anything outside what you have configured. The agent stays on task every time.
Shared identity
Most implementations run as a service account. The AI inherits the permissions of that account, which often provides far more access than any individual user should have. JDE row-level and function-level security is bypassed entirely.
User identity
Every request executes as the authenticated user. JDE row-level and function-level security is fully enforced. If the user cannot do it in JDE, neither can the AI.
Raw fields
Tools expose the raw JDE field names with no additional context. The agent sees a field called NUMB and makes assumptions about what it is and how to use it. Wrong assumptions mean wrong data.
Business context
Every tool publishes field descriptions, required and optional status, valid data types, and business context directly to the agent. Cantara handles JDE data type formatting automatically. The agent knows what every field means before it touches it.
Broad context
Broad context and unrestricted access means more decisions the agent makes for itself. More agent decisions means more opportunity for unpredictable behaviour that is difficult to trace or explain.
Traceable action
Tight, task-specific context keeps the agent focused. Every action is logged. Every decision is traceable. You can see exactly what the agent called, what JDE returned, and what the agent did next.
USE CASES
What this looks like when it runs.
The scenarios where CIP7's native MCP server adds the most value share a common characteristic. The source information varies every time, no traditional integration can handle the variation, and the cost of solving it with manual labour is significant. Here are two examples of what that looks like in practice.
USE CASE 01
Asset Data Migration and Validation
Turning unstructured site data into JDE-ready asset records.
Organisations with large asset registers, particularly those in infrastructure, resources, and heavy industry, face a problem that has no clean traditional software solution. When assets are acquired, moved, or newly commissioned, the source information comes in whatever form it comes in: site survey spreadsheets, inspection reports, legacy system exports, handwritten forms digitised to CSV. No two sites structure their data the same way.
Loading that data into JDE requires a person to go through every record and cross-reference it against live JDE data, including business units, company codes, address book records, UDC category code lists, and manufacturer codes. They must resolve every lookup, apply formatting rules, generate structured field values from free-form text, and produce a validated load-ready file before anything can go near the Z table. For a large asset register, that is days of skilled manual work per site.
With CIP7's native MCP server, an AI agent takes that unstructured source file, validates every record against live JDE reference data, resolves lookups across business units and UDC tables, generates correctly formatted field values, including multi-line description fields built from long free-form text, and produces a structured spreadsheet ready for human review. The agent runs as the authenticated user. Every lookup it makes against JDE is logged. Once the output is reviewed and approved, it is ready to load.
What changes is not the process. What changes is who does the thousands of individual lookups and formatting decisions that process requires. The agent does it in minutes. The human reviews the result.
USE CASE 02
Operational Queries and Daily Summaries
Answering JDE questions without running a report.
Teams that depend on JDE data day to day, including maintenance supervisors, operations managers, and procurement leads, often spend significant time running queries, pulling reports, or waiting for someone else to pull them. What work orders were created yesterday? Which purchase orders are sitting in my approval queue this morning? What is the status of assets in this business unit?
An AI agent connected to CIP7's native MCP server can answer these questions in natural language, pulling from live JDE data through a governed, auditable interface. The agent runs as the user asking the question, so it only sees what that user is permitted to see. Every query is logged. And because the MCP server is task-specific, the agent cannot go beyond the data and functions it has been configured to access.
SECURITY AND GOVERNANCE
Every request is authenticated, governed, and logged. No exceptions.
Enterprise AI integration that cannot be audited is not enterprise AI integration. CIP7's native MCP server is built on the same authentication, policy, and audit infrastructure that underpins the rest of the Cantara platform. Every AI-initiated action passes through the same governed path as every other request.
User-Level Security, Always
Every request executes as the authenticated user, not a shared service account. JDE row-level and function-level security is fully enforced for every AI-initiated action. If the user does not have permission to perform an action in JDE, the AI cannot perform it either. There is no elevated access, no bypass, and no way for the agent to escalate its own permissions.
Full Transaction Audit Trail
CIP7 logs every AI-initiated action with the authenticated user identity, timestamp, tool called, parameters passed, and JDE response received. You can see exactly what the agent did, in what order, and what JDE returned at each step. For organisations subject to SOX compliance, internal audit requirements, or external regulatory review, this is not optional. It is built in by default.
No AIS Exposure to the Internet
CIP7's native MCP server sits in front of JDE. Your AIS server is never exposed directly to the internet. All AI-initiated requests pass through CIP7's full authentication and policy layer before they reach any JDE service. The attack surface does not expand when you add AI.
Identity Provider Integration
CIP7 connects directly to your existing corporate identity provider, including Microsoft Entra ID, Okta, or any OAuth 2.0 compatible IdP. The same identity governance that controls access across your organisation controls access for your AI agents. No separate user management, no shadow accounts, no service account sprawl.
NATIVE TO CIP7 · AVAILABLE NOW
Ready to connect JDE to your AI stack?
CIP7's native MCP server is available now. Talk to the Rinami team about what a task-specific, authenticated, fully governed AI integration looks like for your JDE environment.
