# How to add AI agents to Oracle Fusion Cloud ERP > Source: https://netray.co/integrations/oracle-fusion-ai-agents > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: Integrations ## Short answer Build agents as a thin tool layer over Fusion REST resources and BI Publisher or OTBI reports for reads, and route every write through a human approval step into REST or FBDI. Oracle ships its own embedded agents, so the gap worth filling is cross-module, cross-system work that those do not cover. ## System Oracle Fusion Cloud ERP ## Capability ai-agents ## Intro Teams on Oracle Fusion Cloud ERP and Supply Chain Management already get quarterly feature drops, many with AI features switched on per module. What they still do by hand is the connective work: chasing late supplier acknowledgements, reconciling a receipt mismatch against an invoice hold, or explaining why a work order is short on a component that shows as available in a different organization. An agent is a loop that reads state, decides the next step, calls a tool and stops for approval before it changes anything. Fusion is SaaS, so the agent never touches the database. It works through the same REST resources, reports and import services a good integration developer would use, which also means Fusion's own security roles and data security policies stay in force. Be honest about scope. If the job is drafting a requisition description or summarising a supplier inside a Fusion page, the vendor's embedded features are probably enough. Custom agents earn their keep when the work spans Fusion plus a shop-floor system, email, a quality system or an on-prem ERP in another plant. ## Use Cases ### Supplier acknowledgement chasing **Description:** Query open purchase order schedules past their promised date, check which suppliers have not acknowledged, draft a follow-up and propose a promised-date update for a buyer to approve. ### Receipt and invoice hold triage **Description:** For payables invoices on hold, pull the matching purchase order, receipt and tolerance, classify the hold reason and prepare the fix: a receipt correction request, a tolerance exception or a supplier credit memo query. ### Shortage explanation for work orders **Description:** When a manufacturing work order shows a component shortage, trace on-hand by subinventory and organization, open supply, and reservations, then explain the real constraint in plain language. ### Expedite and reschedule proposals **Description:** Read planning recommendations and supply requests from Supply Planning outputs, rank by customer impact and propose expedite, push-out or cancel actions as a queue for planners. ### Month-end accrual review **Description:** Compare uninvoiced receipts and open purchase orders to accrual entries, flag outliers with the supporting lines and draft journal descriptions for a controller to review. ## System Surface REST resources under the fscmRestApi, hcmRestApi and crmRestApi roots, versioned in the path (for example 11.13.18.05). Use q filters, fields, limit and offset, and onlyData=true to keep payloads small. Resource names and available attributes vary by release, so read the REST API catalog for your pod. ERP Integration Service (SOAP) and the equivalent REST import endpoints for FBDI (file-based data import): you upload a zipped CSV to UCM or Object Storage, then submit the Enterprise Scheduler job. This is the supported bulk write path. BI Publisher reports and OTBI subject areas for read-heavy questions that are awkward through REST, called through the report service or scheduled to deliver files. BI Cloud Connector (BICC) extracts of view objects to OCI Object Storage for bulk, incremental reads into your own store. Good for agents that need history or cross-record search. Business events and Oracle Integration (OIC) adapters for event triggers, such as a purchase order approval or a receipt, so agents react instead of polling. ## How It Works ### Create a dedicated integration user and roles **Detail:** Give the agent a service user with the narrowest job roles and data security contexts that cover its reads, and a second, separate identity for the approved-write path. Do not reuse a human buyer account. ### Define a small tool catalog **Detail:** Start with read tools: get_purchase_order, list_open_schedules, get_onhand, get_work_order, get_invoice_holds. Each tool wraps one REST call or report and returns a trimmed, typed result. ### Mirror slow or large data **Detail:** Use BICC or scheduled BI Publisher output to land history in a store the agent can search. Live REST is for confirming the current record, not for scanning a million lines. ### Add proposal tools, not write tools **Detail:** Write tools create a proposal record with the intended REST payload or FBDI row set, the evidence and the reason. Nothing reaches Fusion at this stage. ### Insert the approval step **Detail:** A named approver sees the proposal, edits it and approves. Only then does a separate executor call REST or submit the FBDI job, using the write identity. ### Log every call **Detail:** Record the prompt, tool calls, REST URLs, response codes and the approver. Fusion's own audit trails cover the change itself, your log covers why it was proposed. ## Example Flows ### Late supplier follow-up **Trigger:** Nightly schedule reads purchase order schedules whose promised date has passed **Steps:** - Agent lists schedules past promise with no acknowledgement - It groups by supplier and pulls the buyer, item and need-by date - It drafts one email per supplier listing the lines - It proposes new promised dates as a pending update for the buyer **Result:** Buyer reviews one queue in the morning instead of opening forty purchase orders. ### Invoice hold explanation **Trigger:** An invoice lands on a matching hold in Payables **Steps:** - Agent fetches the invoice, its lines and the matched receipt - It compares quantity and price against the purchase order and tolerance - It classifies the cause and drafts the resolution note **Result:** AP sees the cause and the evidence, and a receiving correction request goes to the warehouse when needed. ## On Prem Fusion itself is Oracle-hosted, so there is no on-prem Fusion database to point a model at. What can stay private is everything around it: the agent runtime, the retrieval index and the language model. Pull data through REST, BICC or BI Publisher into a store inside your own network or sovereign cloud, serve an open-weight model there with vLLM or Ollama, and send only the final, approved transaction back to Fusion. If your contract or regulator restricts what leaves the Oracle tenancy, extract only the fields the agent needs and mask the rest at the boundary. ## Security - Use OAuth 2.0 through your identity domain with short-lived tokens, rather than stored basic-auth passwords, wherever your pod supports it. - Rely on Fusion data security: the agent's user should only see the business units and ledgers it needs. - Separate read and write identities, and keep write credentials out of the model's reach entirely. - Redact supplier banking details and personal data before anything is sent to a model outside your boundary. - Keep an immutable log of proposals, approvals and executed payloads for audit. ## Options ### Oracle's embedded AI agents and Oracle Integration **Fit:** Work that stays inside Fusion pages and standard processes, with Oracle-managed models. **Effort:** Configuration by your Fusion team, subject to licensing and release availability. ### Build in-house on REST and OIC **Fit:** A strong integration team that wants control of the tool layer and approvals. **Effort:** Weeks for a read-only pilot, longer for governed write-back. ### Partner build with Netray **Fit:** Cross-system agents or a private model requirement, with Fusion as one source among several. **Effort:** Scoped pilot on a single workflow first, then widen. ## Pitfalls - Paging through large REST collections is slow and throttled. Mirror bulk data instead of scanning live. - REST attribute names and resource versions change between releases. Pin the version and test after each quarterly update. - Direct REST writes skip some process validations you get through the UI or FBDI. Test the exact payloads in a non-production pod first. - Agents that read free text in descriptions or attachments can be steered by it. Treat that text as untrusted input. - Letting an agent post to a closed period or approve its own work breaks segregation of duties. Enforce it in the executor. ## Faq ### Can an AI agent write directly to Oracle Fusion? **Answer:** Technically yes, through REST or FBDI, but you should not allow it without review. Have the agent produce a proposal with evidence, let a named person approve, and have a separate executor submit the transaction under a controlled identity. ### Do I need Oracle Integration to build agents? **Answer:** No. Direct REST calls and BI Publisher work fine for a pilot. Oracle Integration helps when you need adapters, event handling, retries and monitoring, which production usually does. ### Where should the language model run? **Answer:** That depends on your data rules. A hosted model is quickest to start. If supplier, pricing or defense program data must not leave your boundary, run an open-weight model on your own GPUs and keep only approved transactions flowing back to Fusion. ### How is this different from Fusion's built-in AI? **Answer:** Built-in features are tied to specific Fusion pages and Oracle-hosted models. A custom agent can span Fusion and other systems, use your own model, and apply your own approval rules, at the cost of building and maintaining it. ### How do I limit what the agent can see? **Answer:** Create a dedicated service user with only the needed job roles and data access sets. The agent inherits those limits automatically, so it cannot query business units or ledgers outside them. ## Search terms oracle fusion ai agents, ai agents for oracle fusion cloud erp, oracle fusion agentic ai, oracle fusion rest api automation, build ai agent oracle fusion scm, oracle fusion procurement agent, oracle fusion custom ai agent, oracle fusion ai workflow automation, oracle fusion fbdi automation --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json