OracleERP PlatformUnited States

JD Edwards World + IBM i

AI for JD Edwards World on IBM i, without a EnterpriseOne migration

Short answer

JD Edwards World running on IBM i can get natural-language question answering and document AI today by reading DB2 for i physical files through a read replica or CDC feed, without touching green-screen programs or committing to a EnterpriseOne migration. The AI layer sits beside the box, on a Linux or Windows host with a GPU, and never writes back to World tables directly.

ERP: JD Edwards World, JD Edwards A9.x, JD Edwards WorldSoftware
Industries: Manufacturing, Distribution, Aerospace, Electronics
Written for: IT Director

Discuss your JD Edwards World environment See use cases

JD Edwards World is still running the plant. It has been running it since before most of the staff who maintain it were hired, and the green-screen sessions on 5250 emulators still process orders, close work orders, and post the general ledger correctly, every day. The problem is not reliability. The problem is that nobody outside the small group who know the RPG and CL programs can get a straight answer out of it, and every AI vendor pitch assumes EnterpriseOne, or assumes cloud, or assumes both.

The honest starting point is that JD Edwards World is a legacy platform Oracle no longer actively develops, and most sites on it are there because a EnterpriseOne or a different-ERP migration has been deferred, not because World is strategic. That changes what an AI project should look like. It should not require touching the CNC/RPG programs, it should not require a database migration, and it should be small enough to prove value before anyone commits budget to the bigger EnterpriseOne conversation.

What does work is reading DB2 for i directly. World's physical and logical files (F4211 sales order detail, F4102 item branch, F0911 GL detail, F0901 account master, and the rest of the 4-xx and 0-xx file ranges) are ordinary DB2 for i tables under the covers, reachable over ODBC/JDBC via IBM i Access, or via journal-based change data capture if you want near-real-time without polling load. An AI layer built on that connection can answer questions and summarize documents without a single change to a World program object.

IT directors evaluating this usually have two real constraints: IBM i capacity is not to be touched (the partition is sized for the ERP workload, not for a GPU), and any new system has to survive the IT director who set it up eventually leaving. Both point the same direction: run the AI stack off-box, on commodity Linux with a GPU (or CPU for smaller models), pointed read-only at IBM i, documented plainly enough that whoever inherits it can operate it.

What usually gets in the way

The problems we hear most from it director teams running JD Edwards World.

Only a few people can query World data directly

Answering "what shipped against this sales order" means someone who knows F42 file relationships and DDS gets on a 5250 session or writes an RPG query. That person is often close to retirement.

No modern reporting layer over DB2 for i

World predates web-friendly reporting. Business users get static reports on a schedule or wait for IT, instead of asking a question and getting an answer.

Vendor AI pitches assume EnterpriseOne or cloud

Most JDE-branded AI and Fusion AI material targets EnterpriseOne or Oracle Cloud customers; World shops are told to migrate first, which is a multi-year, multi-million-dollar answer to a question that needed a report.

Tribal knowledge lives in RPG and CL, not documentation

Custom programs, table relationships, and business rules exist as code comments (if that) written decades ago. New hires take months to become productive.

IBM i capacity is precious and not meant for AI workloads

The partition is licensed and sized for the ERP job; nobody wants to add GPU inference or a vector database on the same LPAR.

Where AI earns its place in JD Edwards World

Each use case names the ERP objects it reads or writes, so your ERP team can judge the integration effort before anyone commits budget.

Natural-language sales order and shipment status

Let a planner or CSR ask "which lines on SO 84213 have not shipped" or "what's the open order balance for customer 40021" in plain English, answered from F4211/F42119 via a read replica.

Touches: F4211 (sales order detail), F42119 (sales order history), F4201 (sales order header), F0101 (address book)

Outcome: Cuts routine order-status lookups from a request to IT or a 5250 query down to a self-service question with a cited answer.

GL and account inquiry for finance

Ground an assistant on F0911/F0901 so controllers can ask about account balances, journal entry detail, and period-over-period variance without a custom report request.

Touches: F0911 (account ledger), F0901 (account master), F0006 (business unit master)

Outcome: Reduces ad hoc reporting requests to the ERP team; answers come with the underlying journal entries shown, not just a number.

Inventory and item master lookups

Warehouse and purchasing staff ask about on-hand quantity, item branch/plant setup, and lot/serial detail without learning F41/F42 file layouts.

Touches: F4102 (item branch), F41021 (item location), F4108 (item cost), F42 sales files

Outcome: Faster answers on hand-inventory and allocation questions, especially useful for staff who are not IBM i literate.

Work order and routing document summarization

Summarize open work orders, parts shortages, and routing steps for a shop supervisor from World manufacturing files, in plain language rather than a print spool.

Touches: F4801 (work order master), F3111 (routing), F3411 (work order parts list)

Outcome: Faster shift handoffs; supervisors get a plain-English summary of what is blocking a work order instead of reading a raw spool file.

PDF and document intake for AP and receiving

OCR and classify vendor invoices, packing slips, and certificates of conformance, then propose a match against open PO lines for a clerk to confirm, without writing to World.

Touches: F4311 (PO detail), F0411 (AP ledger)

Outcome: Cuts manual keying and matching time on routine invoices; exceptions still route to a person.

Tribal-knowledge capture for RPG/CL programs

Point a code-aware model at exported RPG, RPGLE, and CL source to generate plain-English descriptions of what a program does, as a documentation aid ahead of any modernization decision.

Touches: QRPGLESRC/QCLSRC source members, program-to-file cross-reference

Outcome: Builds a documentation baseline in weeks rather than depending on the one person who remembers why the program branches the way it does.

Read-only agent for EnterpriseOne migration scoping

Use an agent to profile actual File usage, custom program call patterns, and data volumes across World, producing an evidence-based input to a future EnterpriseOne or replacement decision.

Touches: cross-reference of program-to-file usage, F98 object management files

Outcome: Gives the migration business case real usage data instead of guesswork, without committing to the migration yet.

Reference architecture

The AI stack runs entirely off the IBM i partition, on a Linux (or Windows) host with GPU or CPU inference, reading DB2 for i through a read replica or journal-based CDC feed and never writing back to World tables.

1
IBM i connectivity
ODBC/JDBC via IBM i Access Client Solutions, or DB2 Mirror/journal-based CDC for near-real-time replication of F4211, F0911, F4102 and related files to an off-box replica.
2
Semantic layer
A mapping table that translates World's 4-6 character file and field names (SDDOC, SDAN8, SDLITM) into business terms an LLM and a human both understand, built once with input from whoever still knows the file layouts.
3
Model serving
Open-weight models (Llama, Qwen, Mistral class) served via vLLM or Ollama on a GPU box sized to query volume, entirely separate from the IBM i partition.
4
Retrieval and agents
Text-to-SQL against the read replica for structured questions, RAG over scanned documents and program source for unstructured ones, with read-only default and explicit approval for any proposed write.
5
Governance and audit
Every generated query and every answer logged with the source file and record identifiers, so a controller or auditor can trace an AI answer back to the F09/F42 rows it came from.

Integration notes for your ERP team

Use IBM i Access Client Solutions (ODBC/JDBC) for the initial connection; for higher volumes or near-real-time needs, use journal-based CDC (IBM's own DB2 Mirror, or a CDC tool that reads the journal receivers) instead of polling.
Build the semantic layer with whoever on staff still knows the F4xxx and F09xx file layouts; this is the single highest-leverage step and the one most projects underinvest in.
Keep the connection strictly read-only at the database user level (a profile with SELECT-only authority on the relevant libraries); write-back, if ever needed, goes through JDE World's own APIs or is proposed for human entry, never a direct table write.
Confirm CCSID/EBCDIC-to-Unicode handling on any text extraction; World data often carries legacy code pages that need explicit conversion in the ETL step.
Size the read replica refresh to the use case: sales order status can tolerate a 15-minute lag; a same-day AP matching workflow needs closer to real-time.
For source-code documentation use cases, export RPG/RPGLE/CL source members to a flat file store rather than granting the model any access to the IBM i program libraries themselves.
Plan for World's field-length and abbreviation conventions in prompt design: field names like SDDOC or SDAN8 need to be translated to "document number" and "customer number" before an LLM will reason about them reliably.

Deployment options

Air-gapped on-prem

Sites with no appetite for any data leaving the building, or defense/aerospace suppliers on World who need to keep ITAR-controlled technical data on premises.

GPU or CPU inference server on the same LAN as the IBM i partition, replica database on the same network segment, no outbound internet path for the model.

Private or sovereign cloud

Sites open to cloud economics but not to a public multi-tenant AI API touching ERP data.

Model and replica run in a dedicated VPC or single-tenant cloud instance; IBM i stays on-prem, connected over a site-to-site VPN to the replica.

Hybrid

Sites that want to pilot fast on a laptop-class GPU or CPU-only box before committing rack space or capital.

Small quantized model for the pilot, same architecture, scaled up (or moved to a dedicated GPU server) once the use case is proven.

Compliance and data control

How the architecture supports your obligations. Certification and accountability stay with your organisation; the design keeps the evidence straightforward.

ITAR / export control

For defense-adjacent JDE World shops, technical data never crosses to a public API; the model and the data stay on hardware the company controls, with access scoped to US persons where required.

SOX / financial controls

Read-only access to F09/F0911 for GL inquiry; any AI-proposed adjustment routes through the existing journal entry approval workflow rather than writing directly.

Data retention

Read replica retention and logging policy set to match your existing IBM i backup and retention schedule; no new offsite copy of ERP data created without sign-off.

Change control

AI project runs entirely outside the World object library (no changes to CNC/RPG programs), so it does not touch existing change management or SOX ITGC scope for the ERP.

Where Netray fits

ERPray

ERPray's ERP-agnostic connector architecture fits a JDE World read replica the same way it fits any relational ERP: point it at the semantic layer, keep it read-only, and it answers questions with the underlying query shown.

Custom build

Source-code documentation, migration-scoping agents, and OCR-based document intake for World are typically custom builds, since they touch program source and document workflows outside a packaged connector.

How an engagement runs

Phase 1 . 2-3 weeks

Discovery

-File-to-business-term mapping for the target use case (sales, GL, or inventory)
-Confirmed connectivity path (ODBC replica vs. journal CDC)
-Read-only access and security review with IT and IBM i admin
-Pilot scope and success criteria

Phase 2 . 6-8 weeks

Pilot

-Working replica and semantic layer for the pilot data domain
-Model serving stood up on a GPU or CPU host separate from the IBM i partition
-5-10 real user questions answered end to end with cited sources
-Pilot review with IT director and business sponsor

Phase 3 . 8-12 weeks

Production

-Hardened replica refresh (CDC if the pilot used polling)
-Role-based access aligned to existing World security groups
-Audit logging of queries and answers
-Runbook for whoever operates IBM i to maintain the connector

Phase 4 . ongoing

Scale

-Additional file domains (inventory, work orders, AP) added to the semantic layer
-Document intake use cases layered on top of the same infrastructure
-Quarterly review of query patterns to prioritize the next use case

Questions to ask any vendor, including us

A short list that separates real JD Edwards World AI work from a chatbot demo.

Does your approach require any change to World program objects, or is it strictly read-only against the database?
How do you handle journal-based CDC versus polling, and what is the realistic data latency for each?
Who builds the file-to-business-term semantic mapping, and how much of my staff's time does that take?
Where does the model run relative to my IBM i partition, and does any data leave my network?
What happens to this project if we later decide to migrate to EnterpriseOne, or to a different ERP entirely?
Can you show me a real example of a text-to-SQL query your system generated against F4211 or F0911, not a demo screenshot?
How do you handle EBCDIC/CCSID text conversion in the pipeline?
What is your rollback plan if the read replica falls behind or the connection drops during a business-critical period?

Frequently asked questions

Can I add AI to JD Edwards World without migrating to EnterpriseOne?

Yes. JD Edwards World's data lives in ordinary DB2 for i tables, reachable read-only via ODBC/JDBC or journal-based CDC. An AI layer can query and summarize that data on a separate host without touching World's RPG programs or requiring an EnterpriseOne migration.

Does this AI approach change or write to JD Edwards World data?

Not by default. The standard architecture is strictly read-only: a replica or CDC feed off the IBM i partition, with any proposed action reviewed by a person before it is entered into World through normal channels. Write-back is possible later but is a deliberate, separately scoped decision.

Will this put extra load on my IBM i partition?

Model inference and vector storage run on a separate Linux or Windows host, not on the IBM i LPAR. The only load on IBM i is the replication feed itself, which can be tuned (polling interval, or journal-based CDC) to stay well within normal database activity levels.

How do you handle the cryptic file and field names in World, like F4211 or SDAN8?

A semantic layer translates World's file and field names into business terms before the model ever sees them. Building that mapping with someone who knows the file layouts is the most important step in the project, more important than model choice.

Is this relevant if we are already planning to move off World?

Often more relevant, not less. An AI layer built on read access can also profile actual file and program usage, giving your EnterpriseOne or replacement-ERP business case real usage data instead of guesses, while you still get value from World in the meantime.

Can this read scanned documents like vendor invoices and match them to World PO data?

Yes. Document AI (OCR plus classification) can extract invoice fields and propose a match against F4311 PO detail via the same read replica, with a person confirming the match before it is keyed into World.

What about the RPG and CL programs nobody remembers the details of?

A code-aware model can read exported RPG/RPGLE/CL source and generate plain-English descriptions of what each program does, useful as a documentation baseline for onboarding new staff or scoping a future migration, without executing or modifying the code.

Related guides

JD Edwards + AI agents

AI for JD Edwards EnterpriseOne, Built on Orchestrator and AIS

Add AI to JD Edwards E1 9.2 using Orchestrator, AIS, and BSSV rather than screen-scraping. Ground a private LLM on F4211, F4111, and F0411 data.

Oracle EBS + private AI

AI for Oracle E-Business Suite, Without Leaving On-Prem

Add AI to Oracle E-Business Suite 12.2 without moving off-prem. Query concurrent programs, interface tables, and AP/PO data with a private LLM. See how.

NetSuite Manufacturing + AI agents

AI agents for NetSuite manufacturing operations

AI agents for NetSuite manufacturing: WIP tracking, routing exceptions, and work order status grounded in SuiteQL, with human approval on anything that writes back.

ERP migration + on-prem AI

AI-assisted data migration for ERP implementations

AI speeds ERP data migration by profiling legacy data, proposing field mappings, and flagging cleansing issues before cutover, with a human validating every rule.

Infor XA on IBM i + on-prem AI

AI for Infor XA on IBM i: On-Prem, No Replatforming Required

Add AI to Infor XA on IBM i (AS/400) without a migration project. Read DB2 for i directly, ground answers in live XA data, keep everything on-prem.

RAG + SQL + permissions

A Private LLM Grounded on Your ERP Data

How a private LLM answers questions on your ERP data: RAG plus text-to-SQL, role-based permissions inherited from the ERP, and where each fits.

Plan it with numbers

Free Tool

Talk it through with an engineer who knows JD Edwards World

Bring one real question your team cannot answer from the ERP today. We will map the data path, the model, and where it runs, and tell you honestly if AI is the wrong tool for it.

Start the conversation See ERPray

All AI for ERP guides