How to add AI agents to Oracle E-Business Suite
Short answer
Build a thin tool layer over EBS read views, Integrated SOA Gateway services and the open interface tables, then let an agent call those tools in a loop. Reads go direct, writes go through an interface table or public API only after a human approves. Run the model on your own GPUs if the data is regulated.
Most Oracle E-Business Suite estates are R12.1.3 or R12.2 on-prem, heavily customised, with a planner or buyer team living in Purchasing, Inventory, ASCP or Discrete Manufacturing forms all day. The pain is rarely missing data. It is the volume of exceptions: late POs, MRP recommendations, stuck interface rows, holds on order lines.
An agent is useful where a person follows the same investigative path every time: open the exception, check supply and demand, look at the supplier, decide, update. The agent does the looking and drafts the decision. A person keeps the click that changes the books.
EBS is a good fit for this pattern because it already has a controlled write path. Open interface tables and public PL/SQL APIs exist precisely so that outside programs can post transactions with the same validation the forms apply.
What teams use it for
MRP exception triage
Read recommendations from MRP_RECOMMENDATIONS (or the ASCP equivalent), group by buyer and urgency, and draft reschedule or expedite actions with the supporting demand shown.
Late PO chasing
Find PO shipments past need-by in PO_LINE_LOCATIONS_ALL, check the supplier's last acknowledgements, and draft a chaser email for the buyer to send.
Interface error resolution
Scan PO_INTERFACE_ERRORS, MTL_INTERFACE_ERRORS and AP_INTERFACE_REJECTIONS, explain each rejection in plain words and propose the corrected row.
Order hold review
List OE_ORDER_HOLDS_ALL entries, summarise why each hold exists (credit, configuration, export), and route to the right owner.
Shortage analysis for a work order
For a WIP discrete job, compare WIP_REQUIREMENT_OPERATIONS to on-hand and incoming supply and report the exact shortage chain.
Non-conformance drafting
Gather lot, supplier and receipt history for a rejected receipt and draft the quality record text for a reviewer.
Oracle E-Business Suite integration surface
- Integrated SOA Gateway (ISG): publishes PL/SQL APIs, open interfaces and business events from the Integration Repository (iRep) as SOAP and REST services. Which services are deployed and enabled depends on your patch level and what your admin has generated.
- Public PL/SQL APIs such as PO, OE_ORDER_PUB and INV transaction APIs: validated write paths, called after setting context with FND_GLOBAL.APPS_INITIALIZE and MO_GLOBAL.SET_POLICY_CONTEXT for multi-org.
- Open interface tables (for example PO_HEADERS_INTERFACE, MTL_TRANSACTIONS_INTERFACE, AP_INVOICES_INTERFACE) processed by concurrent programs: the safest bulk write path.
- Read-only database access through a dedicated reporting schema with views over base tables, ideally on a standby or snapshot so agents never touch the production APPS session load.
- Business Event System and Workflow: raise or subscribe to events (for example PO approval events) to trigger agent runs.
How it works
- 1
Pick one exception queue
Start with a single, high-volume, low-risk queue such as late PO shipments. Write down today's manual steps before building anything.
- 2
Create a read-only data layer
Expose curated SQL views or ISG read services for that queue. Do not hand the agent raw table access or the APPS password.
- 3
Define tools with narrow inputs
Tools like get_po_shipments(buyer, days_late), get_item_supply(item, org), get_supplier_history(vendor). Each returns bounded rows with a hard cap.
- 4
Add a write path through interfaces
Writes become proposals: a row staged in a Netray-owned table, or an interface table row marked pending, never a direct base-table update.
- 5
Insert an approval step
Show the buyer the proposal, the evidence queries and the diff. On approve, a service account with a limited responsibility posts through the public API.
- 6
Log every tool call
Store prompt, tool calls, SQL, rows returned and the approver. Replay is how you debug and how you answer an audit.
- 7
Widen gradually
Add queues one at a time, and keep an eval set of past exceptions with the correct human resolution to score changes.
Example flows
Expedite a late PO
Trigger: Nightly job finds shipments more than 3 days past need-by for a critical item.
- Agent pulls the shipment, PO line and the work orders consuming that item.
- It checks on-hand and other open supply in the same org.
- It estimates the production impact by date.
- It drafts an expedite note and a promise-date change proposal.
- Buyer approves or edits.
Result: Buyer sees a ranked list with evidence instead of building the same analysis by hand.
Fix a rejected interface batch
Trigger: A concurrent program finishes with rejected rows in the interface errors table.
- Agent reads the error codes and the staged rows.
- It looks up the missing reference data (for example an inactive UOM or an unmapped category).
- It proposes corrected rows and the reference-data fix.
- A functional analyst approves.
- The import program is resubmitted.
Result: Rejections get a plain-language reason and a ready fix.
Keeping data on-prem
EBS data is usually the reason this has to stay inside the network. Serve an open-weight model such as a Llama, Qwen or gpt-oss class model with vLLM or Ollama on your own GPU servers, in the same data centre or VPC as the EBS database tier. The agent service queries a standby or reporting schema over a private link, holds its audit log on local storage, and has no outbound internet route. Embeddings for any document retrieval (supplier contracts, work instructions) are computed locally as well. Plan GPU memory for the longest tool-output context you allow, not for the shortest question.
Security checklist
- Dedicated database user and EBS application user for the agent, with a narrow responsibility; never reuse a human or APPS credential.
- Respect operating-unit and organisation security: set the policy context per request so the agent only sees what the requesting user could see.
- Read-only by default; every write is a staged proposal with a named human approver.
- Log the exact SQL and rows returned for each answer, and retain it per your audit policy.
- Mask or exclude columns with personal or pricing data that the use case does not need.
Ways to get it done
Oracle's embedded AI and Workflow
Effort: Low, if available
Cloud applications get more built-in AI than classic R12 on-prem. For EBS, check what your release and patch level actually include before relying on it.
Build in-house on ISG and a read replica
Effort: 6 to 12 weeks for a first queue
You have PL/SQL and integration skills and want full control.
Partner build with Netray
Effort: 4 to 10 weeks for a first queue
You want a private on-prem agent with approval gates and an evaluation set, delivered alongside your EBS functional team.
Pitfalls
- !Letting the model write SQL against base tables without a row cap, which can pin the production database.
- !Forgetting multi-org context, so answers silently mix operating units.
- !Calling public APIs without APPS_INITIALIZE, producing confusing validation failures or audit entries under the wrong user.
- !Skipping a regression set: a model or prompt change shifts behaviour and nobody notices for weeks.
- !Automating a broken process. If buyers disagree on the correct action today, the agent will too.
Frequently asked questions
Can an AI agent write to Oracle E-Business Suite safely?
Yes, if it never updates base tables directly. Stage proposals, show them to a person, and post through open interface tables or public PL/SQL APIs so the normal EBS validation and audit apply. Start read-only and add one write type at a time.
Do I need Integrated SOA Gateway?
Not strictly. ISG gives you managed SOAP and REST services and is a clean option when enabled. Many teams use a reporting schema for reads and open interface tables for writes. The right mix depends on your release, security stance and DBA preferences.
Which release does this work on?
R12.1.3 and R12.2 both have open interfaces and public APIs. ISG capabilities and REST support differ by release and patch level, so confirm in the Integration Repository on your instance.
Can this run without sending data to a cloud AI vendor?
Yes. Use an open-weight model on your own GPUs with vLLM or Ollama, and keep the agent service, vector store and logs inside your network. Accuracy depends on the model size and how well your tools are designed.
What is a sensible first agent?
Late PO shipment triage or interface error explanation. Both have clear inputs, a measurable outcome and low blast radius. Avoid starting with anything that posts financial transactions.
More for Oracle E-Business Suite
Want AI agents working on your Oracle E-Business Suite?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment