Ask JD Edwards
Ask about orders, invoices, balances or items in your own words. The answer comes from real queries against your JD Edwards data, run as you, and the queries used are shown alongside it. If a call fails, the platform says so plainly rather than reporting an empty result.
- Every figure comes from a query, not from the model's memory
- Exports are built on the server from the database, with the true row count
- Paste or drop in scans, images and PDFs and have them read
- Search and history across your assistants, agents and past chats
How you would verify it. Ask a question you already know the answer to, then open the queries listed under the reply and compare them with JD Edwards.
Do the work
The platform can create and update sales orders, vouchers, journal entries, address book records, work orders and price records, signed in as the person asking. It uses the sanctioned JD Edwards routes: application stacks, form services, business functions and orchestrations. After a change it reads the record back and tells you whether it could confirm it.
- Changes run under the person's own JD Edwards security
- A trust badge shows read only, change made, or change not verified
- Forms and orchestrations are discovered before use, not guessed
- Reports and batch jobs can be run by name, with failures explained
How you would verify it. Make a test change in a non-production environment and look for the trust badge, then open the record in JD Edwards.
Module experts
There are 24 JD Edwards subject experts plus a front desk that points you to the right one. Each comes with quick actions, shipped skills and an optional switch to ground answers in the Oracle documentation. Experts can be edited in the product, and your changes carry through upgrades.
- Covers areas such as general ledger, payables, receivables, sales, procurement, inventory and manufacturing
- More than 40 shipped skills carry the rules the experts follow
- A Platform Expert answers questions about your own instance, such as usage and spend
- Module audits can be run as a report from the expert page
How you would verify it. Open the experts list and try the same question with and without the documentation switch.
Build your own
Assistants chat; agents follow workflow steps. Your people can build either from a wizard, a plain description, a template or a design document. Assistants and agents can write to JD Edwards, and a read only switch is there when you want one that only looks. Rules are written as editable skills that can be drafted, promoted and rolled back.
- Each agent can be pinned to the AI model that suits its job
- A coach suggests fixes from real runs
- Configuration exports as a portable bundle, so a test build can move to production
- Owners can edit their own agents; admins keep control of shared ones
How you would verify it. Build a small assistant from a one-paragraph description, set it to read only, and test it in the builder.
Run unattended
Run an agent or a pipeline on a schedule or once, and have the result emailed with its cost. Give an assistant its own email address: people on an allowlist send a message or a photo and get a reply. Each run can act as its owner, as the sender, or as a named user, and any run can be forced read only.
- Scheduled and one-off runs with results emailed
- Email in, reply out, including photos and scans
- Read only mode strips every write tool for a given run
- Your own mail service can be used instead of ours
How you would verify it. Set up a scheduled read only run to your own mailbox and check the reply and its cost line.
Rules as code
Where a task needs no judgement, the work runs as code instead of model reasoning. Maths and comparisons run as tested, sandboxed rules. Whole jobs can run as pipelines with no model in the loop: query, compute, write and report, with writes verified and gated.
- A rule that fires is told apart from a rule that crashed
- Large row sets pass from step to step by reference, not through the model
- A step that skipped its work is failed, not reported as success
- Pipelines can be scheduled or started from email
How you would verify it. Run the same pipeline twice on the same data and compare the output and the trace.
Documents and reports
Agents can produce Word, PowerPoint and Excel files and HTML reports from live JD Edwards data. The server builds each file from a structured specification, so a file cannot carry macros, links or model-written formulas. Whole multi-section documents can be driven section by section, with a provenance appendix.
- Files are saved in the creator's own folder, with sharing and file-type policy set by an administrator
- Every generated document states that it was written by AI
- Long documents are assembled by the engine, one section at a time
- Charts are drawn on the server from the data
How you would verify it. Ask an agent for a short workbook, open it, and check the figures against the queries it lists.
Knows your estate
Answers can draw on vectorised Oracle JD Edwards guides, with cited passages, and on your own documents. JDE-IQ learns your estate: custom tables, what your category codes mean, how fields are scaled and how they are really used. It is switched on one module at a time, so you can start with a pilot.
- Cited passages from the documentation and your own files
- A base JD Edwards model your team can download, extend and upload
- A readiness view per module shows what each one still needs
- A report of words your people use that the model does not yet know
How you would verify it. Switch on one pilot module and ask a question that depends on one of your own category codes.
Your own AI tools
The platform is also available from your own AI tools. The connection layer lets a client ask the experts, run multi-step agents and search the documentation, while the experts, skills, JDE-IQ, knowledge and governance sit above it. There are around a hundred tools in three access tiers: user, full and write.
- Each person can generate their own personal token
- Chatting with an expert on a personal token acts in JD Edwards as that person
- The write tier can build agents and pipelines over the connection
- A plugin for Claude Code carries the platform's guidance
How you would verify it. Connect a personal token to your AI tool and ask an expert a question from there.
Tests and drift
Turn a past conversation into a repeatable test. Checks cover the tools used, cost, and the identifiers in the answer, and a lower-cost model scores whether the answer is broadly the same kind as before. A drift view compares recent runs with a baseline you set. A run that could not finish is reported as errored, never as a failure.
- Generic starter tests ship for the experts
- Tests run read only, as the platform's own system user
- Cost caps stop a test run before it overspends
- Off by default, because each run is a billed model call
How you would verify it. Save one real question as a test, run it, and then pin the result as your baseline.
Run and govern
Sign-in uses your Microsoft enterprise application, and JD Edwards must recognise the person. Administrators see usage, cost per conversation and a return-on-investment view per agent, and can replay the journey of any query. The interface runs in nine languages and can carry your own name, logo and colours.
- Per-person interface and reply language
- Optional spend budgets, rate limits and a kill switch
- Optional activity export to your monitoring tools, structure only
- One instance can route people to named regional JD Edwards servers
How you would verify it. Open the usage and cost screens after a pilot week and compare them with your AI provider's bill.
