Plan Your Deployment

Deploy this architecture using the following basic steps:

  • Map the architectural building blocks to Oracle Cloud Infrastructure services
  • Plan the initial implementation with a focus on agent orchestration
  • Enhance the initial implementation by adding agents and integrating advanced LLM reasoning
  • Collect user feedback to drive your continuous improvement process

Map OCI Services

As a first step to planning your deployment, map layers and capabilities in the architecture to the specific OCI products and services you need.

Oracle Cloud Infrastructure provides core building blocks for sidecar implementation, while Oracle Integration adapters provide prebuilt connectivity to systems like SAP. OCI integration services explicitly call out SAP connectivity alongside other SaaS and on-premises sources.

Layer / Capability Responsibilities in Sidecar Primary OCI Products and Services Notes and SAP and Fusion Applications Tie-In
Orchestrator Implementation Host the sidecar orchestrator; expose APIs; optionally model low-code workflows
  • OCI Kubernetes Engine (OKE)
  • OCI Compute
  • OCI API Gateway
  • Oracle Integration (Process Automation)
Orchestrator coordinates calls to SAP/Fusion tools and AI Data Platform data products via well-defined API endpoints.
Agents as Serverless Functions Run lightweight, event-driven agents (enrich, validate, classify, notify)
  • OCI Functions
  • Oracle Integration (SAP, Fusion Applications, and other adapters)
OCI Functions call SAP/Fusion Applications using OCI Functions adapters instead of direct calls, preserving clean core.
Data Storage and Processing Land raw data, curate medallion layers, persist governed “gold” data products; stream data
  • AI Data Platform object storage (OCI Object Storage)
  • Oracle Autonomous AI Database
  • OCI Streaming (Kafka-compatible)
SAP and Fusion Applications data are landed, standardized, and published as governed data products for analytics and AI.
AI and Machine Learning Services Provide LLM endpoints, embeddings, RAG, and analytical insights; power agent reasoning
  • OCI Generative AI service
  • AI Data Platform / Oracle Analytics layer
LLMs use governed AI Data Platform data products to answer questions, explain exceptions, and propose remediation plans.
Networking and Integration Secure connectivity, routing, and identity; connect SAP, Fusion Applications, and other SaaS/on-premises components
  • OCI Virtual Cloud Network (VCN)
  • Private Endpoints
  • OCI Identity and Access Management / policies
  • Oracle Integration (SAP and S/4HANA cloud adapters, Fusion ERP Cloud Adapter)
Provides secure paths for SAP BAPI/RFC/IDoc, S/4HANA OData, and Fusion Applications REST/ERP adapter traffic.
Client Interface Deliver end-user experience: web apps, chatbots, AI assistant sidebars
  • Web apps hosted on OKE/OCI Compute fronted by OCI API Gateway / OCI Load Balancer
  • Chatbot or portal integrations (for example, using Oracle Integration / AI Data Platform UX)
Users interact through an "assistant" UI that can query SAP or Fusion Applications, explain status, and execute approved actions.
Observability and Audit Log, monitor, set alerts, analyze performance and cost management
  • Oracle Cloud Observability and Management Platform services
  • OCI Logging and Oracle Log Analytics
  • OCI Audit
 OCI observability, logging, and analytics services compliment logging and audit capabilities integrated with SAP and Fusion Applications.

Here are some additional details and options for the layers and capabilities:

  • Orchestrator implementation:
    • Containerized microservice (OKE) or compute-based service, fronted by OCI API Gateway.
    • Optionally, process automation flows in Oracle Integration where low-code orchestration is sufficient.
  • Agents as serverless functions:
    • Lightweight agents implemented as OCI Functions where event-driven scaling is desirable (enrich, validate, classify, notify).
    • Agents invoke SAP and Fusion Applications adapters through the integration layer, not directly.
  • Data storage and processing:
    • Oracle AI Data Platform uses OCI Object Storage for raw/landing zones and artifacts.
    • Oracle Autonomous AI Database for governed “gold” data products.
    • Optional streaming/event backbones using OCI’s Kafka-compatible event streaming when near-real-time patterns are required.
  • AI and machine learning services:
    • AI Data Platform’s OCI Generative AI service provides managed LLM endpoints for summarization, embeddings, RAG, and tool-using agent experiences grounded in governed data.
    • Analytics layer for “last mile” insight delivery and feedback to operations.
  • Networking and integration:
    • VCN, private endpoints where applicable, and identity/policy controls.
    • Oracle Integration adapters for SAP (BAPI/RFC/IDoc) and SAP S/4HANA Cloud OData connectivity.
  • Client interface:
    • Web UI, chatbot, or embedded experience inside an internal portal.
    • Common UX pattern: “assistant” sidebar that can answer, explain, and execute actions with approval.
  • Observability and audit:
    • OCI Monitoring for the Oracle Cloud Observability and Management Platform is used to track the performance of the resources in your tenancy. Oracle Log Analytics provides collection, indexing, enrichment, query, visualization, and alerting for logs from any component, including OCI and the SAP and Fusion Applications sources.
    • Leverage the insights from observability and audit functions to control cost, inform your system design decisions, and drive continuous improvement.

Implement Process Automation and Agent Orchestration

The first phase of implementation focuses on producing a working sidecar backbone: consistent tool contracts, reliable connectivity, and repeatable workflows, before turning on full autonomous reasoning.

These are the outputs you plan and implement in phase one:

  • Tool registry for SAP/Fusion Applications/AIDP actions (read/write/curate).
  • Deterministic orchestration scripts (known sequences for top use-cases).
  • End-to-end observability and audit logs.

Here are some example tool definitions.

SAP tools:

  • sap.getPurchaseOrder(poNumber)
  • sap.getInvoiceStatus(invoiceId)
  • sap.createOrUpdateVendor(vendorPayload)
  • sap.postIdoc(idocType, payload)
  • sap.callBapi(bapiName, params) (governed allow-list)

Fusion Applications tools:

  • fusion.getSupplier(supplierId)
  • fusion.createInvoice(invoicePayload)
  • fusion.submitFBDI(jobName, fileRef) (where applicable via ERP adapter patterns)
  • fusion.queryRest(resource, params)

AIDP tools:

  • aidp.publishDataProduct(name, version)
  • aidp.runPipeline(pipelineId)
  • aidp.searchCatalog(term)
  • aidp.ragAnswer(question, dataProductRefs)

At this phase, the orchestrator maintains cross-system context:

  • SAP objects retrieved via adapter calls (BAPI/RFC/IDoc or S/4HANA Cloud OData).
  • Fusion Applications objects retrieved via REST/APIs/adapters.
  • Governed data products from AIDP’s gold layer.

Implement observability and logging. In this pattern you treat the sidecar as a production integration product, using:

  • Tool invocation logs (inputs/outputs redacted as needed).
  • End-to-end tracing from orchestrator through adapters to the data platform.
  • Business identifiers for “follow the transaction” debugging.
  • Data lineage for curated data products (especially those used by AI agents).

Implement Autonomous Agentic Behavior

In phase two, you implement more autonomy.

The LLM can:

  • Plan multi-step workflows (“what do I need to check next?”)
  • Call tools
  • Generate business-ready narratives grounded in governed enterprise data via AIDP

When you implement LLM-driven agent reasoning, instead of fixed sequences, the agent can dynamically decide:

  • Whether to query SAP first or Fusion Applications first
  • Which data products to retrieve
  • Whether to open an exception case or propose a remediation plan
  • Which write-back action is appropriate (often gated by approval)

AIDP explicitly positions its platform for agentic experiences built on governed data and integrated AI services. For example, the LLM can:

  1. Retrieve an invoice status from SAP
  2. Check supplier data in Fusion Applications
  3. Propose remediation using AIDP RAG pipeline

The following additional agents and tools are possible (with example use cases given):

  • Close Assistant Agent (SAP plus Fusion EPM): explains variances, identifies missing postings, and suggests journal adjustments.
  • Supplier Onboarding Agent: harmonizes supplier data, checks duplicates, and creates supplier records in the target system(s).
  • Order Promising Agent: pulls demand/supply signals, identifies constraints, and recommends reallocation or expedites.

Collect Feedback for Continuous Improvement

Outcomes should feed back into the system:

  • “Was the recommendation accepted?”
  • “Did the remediation resolve the exception?”
  • “Did the mapping rules reduce future failures?”

These outcomes become labeled signals for improving matching logic, exception playbooks, and (where appropriate) ML models and prompts - closing the loop on both automation quality and business trust.