Oracle EBS workflow and approval triage agentic AI

An enterprise, policy-aware autonomous workflow management microservice architecture for Oracle E-Business Suite (EBS) Release 12.x. Integrates with Google Cloud Gemini Enterprise, Vertex AI (gemini-3.8-flash), Agent-to-Agent (A2A) Protocol v0.3.0, Model Context Protocol (MCP), and certified Oracle EBS Workflow REST and PL/SQL APIs.

Important

Prerequisites and Required IAM Roles:

• Tools: Python 3.10+, Google Cloud CLI (gcloud), Terraform (v1.5.0+), and Docker.
• Oracle EBS Vision Instance: Active Oracle EBS instance on Google Cloud. Clone and configure from Oracle EBS Framework on Google Cloud.
• Gemini Enterprise Entitlement: Active Gemini Enterprise license on the Google Cloud organization.
• Identity and Access Management (IAM) Roles: Discovery Engine Admin (roles/discoveryengine.admin), Cloud Run Invoker (roles/run.invoker), and Secret Manager Secret Accessor (roles/secretmanager.secretAccessor).
• Enabled Google Cloud APIs: Discovery Engine (discoveryengine.googleapis.com), Vertex AI (aiplatform.googleapis.com), Cloud Run (run.googleapis.com), Secret Manager (secretmanager.googleapis.com), Artifact Registry (artifactregistry.googleapis.com), and Cloud Build (cloudbuild.googleapis.com).

Key capabilities

• Autonomous Multi-Workflow Triage: Inspects, categorizes, and aggregates pending approvals across Expense Reports (APEXP), Purchase Orders (POAPPR), and Requisitions (REQAPPR).
• Policy and risk verification engine: Evaluates corporate policy rules (lodging limits, alcohol spend, delegation thresholds, budget overruns) and assigns risk tiers (LOW, MEDIUM, HIGH, CRITICAL).
• Explicit Human Confirmation Gate: Mutative actions (APPROVE, REJECT) strictly require explicit user confirmation (confirmed=True). Unconfirmed requests receive a detailed ConfirmationProposal with risk alerts and confirmation prompt.
• Certified Oracle EBS Action Layer: ZERO raw SQL updates or direct table mutations. Executes responses strictly using:
  • Path A (Session PL/SQL): apps.fnd_global.apps_initialize + apps.wf_notification.setattrtext + apps.wf_notification.respond
  • Path B (ISG REST): POST /webservices/rest/wf/respond through Oracle Integrated SOA Gateway (ISG).
• A2A Protocol v0.3.0 and MCP compliance: Exposes the agent catalog at /a2a/card and Model Context Protocol database tools at /mcp/.

Architecture and microservices

1. Workflow A2A Gateway Server (src/a2a/a2a_server.py):

   • Cognitive natural language interface using Vertex AI Gemini (gemini-3.8-flash primary, gemini-3.7-flash fallback, gemini-3.5-flash-lite routing).
   • Dynamic intent extraction and Markdown table synthesis.
   • Google Cloud OIDC token validation and authenticated downstream routing.

2. Core Workflow Agent (src/agents/workflow_agent.py):

   • /notifications: Worklist retrieval through Oracle ISG REST and certified PL/SQL API (apps.xx_ebs_wf_agent_pkg).
   • /notifications/{id}: Detailed attribute inspection (amount, requester, justification).
   • /verify-policy: Risk tier calculation and violation identification.
   • /triage: Batch policy evaluation and rich Markdown table rendering.
   • /execute: Certified response execution with human confirmation gate enforcement.
   • /respond: Direct action execution for single notification items.

3. Workflow MCP Server (src/mcp/mcp_server.py):

   • Model Context Protocol tool endpoints for workflow triage, inspection, policy check, and response.

Configuration and environment variables

Variable	Description	Default
ORACLE_HOST	Oracle Database Host / IP	127.0.0.1
ORACLE_PORT	Oracle Database Listener Port	1521
ORACLE_SERVICE_NAME	Oracle Database Service Name	ebsdb
ORACLE_USER	Oracle EBS Schema / Apps User	apps
ORACLE_PASSWORD	Oracle EBS User Password	(from Secret Manager / env)
FND_USER_ID	Default FND User ID	0 (SYSADMIN)
FND_RESP_ID	Default FND Responsibility ID	20420 (System Administrator)
FND_RESP_APPL_ID	Default FND Responsibility App ID	1 (FND / Application Object Library)
WORKFLOW_EXECUTION_PATH	Execution Path: REST or PLSQL	REST
ISG_REST_BASE_URL	Base URL for Integrated SOA Gateway	http://127.0.0.1:8000
ISG_USERNAME	Oracle EBS ISG REST Basic Auth User	ISGADMIN
ISG_PASSWORD	Oracle EBS ISG REST Basic Auth Password	(from Secret Manager / env)
GOOGLE_CLOUD_PROJECT	Google Cloud Project ID	(auto-detected)
GEMINI_MODEL	Primary Foundation Model	gemini-3.8-flash
GEMINI_FALLBACK_MODEL	Fallback Foundation Model	gemini-3.7-flash
GEMINI_FLASH_LITE_MODEL	High-speed routing model	gemini-3.5-flash-lite

Note

Tiered Hybrid Model Architecture: The A2A gateway uses gemini-3.5-flash-lite for fast intent routing and Markdown rendering, and gemini-3.8-flash (gemini-3.7-flash fallback) for policy reasoning and tool orchestration.

Deployment options

Before running either deployment method:

1. Create terraform/terraform.tfvars (gitignored and auto-loaded by Terraform and ./scripts/deploy_stack.sh) with your project, VPC/subnet, and Oracle settings mapped to the table above and terraform/variables.tf (such as project_id, region, vpc_network_name, oracle_host, and isg_rest_base_url).

2. Export database and ISG passwords in your shell so credentials stay out of Terraform state and Secret Manager versions are not left empty:

   export ORACLE_PASSWORD="YOUR_APPS_DB_PASSWORD"   # EBS Vision default: apps
   export ISG_PASSWORD="YOUR_ISGADMIN_PASSWORD"     # EBS Vision default: Welcome1
   

Method A: Local development and Identity-Aware Proxy (IAP) tunneling

Establish Identity-Aware Proxy (IAP) TCP tunnels to the private Oracle database listener:

# Terminal 1: Oracle Database Listener Tunnel (Port 1521)
gcloud compute start-iap-tunnel oracle-vision 1521 \
  --project=YOUR_GOOGLE_CLOUD_PROJECT_ID \
  --zone=YOUR_ZONE \
  --local-host-port=localhost:1521

# Terminal 2: Start microservices locally
source .venv/bin/activate
uvicorn src.a2a.a2a_server:app --port 8080 --reload
uvicorn src.agents.workflow_agent:app --port 8001 --reload
uvicorn src.mcp.mcp_server:app --port 8000 --reload

Method B: Google Cloud Run deployment using Terraform

Option 1: Automated script (recommended)

# Reads terraform/terraform.tfvars, populates Secret Manager from ORACLE_PASSWORD
# and ISG_PASSWORD, bootstraps EBS DB, builds image, and deploys Cloud Run services
./scripts/deploy_stack.sh

# Infrastructure teardown
./scripts/destroy_stack.sh

Option 2: Direct Terraform apply

cd terraform

# 1. Initialize Terraform and create APIs, secret containers, and Artifact Registry
terraform init
terraform apply \
  -target=google_project_service.enabled_apis \
  -target=google_secret_manager_secret.db_password \
  -target=google_secret_manager_secret.isg_password \
  -target=google_artifact_registry_repository.agent_repo \
  -target=google_artifact_registry_repository_iam_member.cloudbuild_ar_writer \
  -target=google_project_iam_member.cloudbuild_logging

# 2. Populate Secret Manager versions (required before Cloud Run can start)
printf "%s" "${ORACLE_PASSWORD}" | gcloud secrets versions add oracle-ebs-workflow-db-password-dev --project=YOUR_PROJECT_ID --data-file=-
printf "%s" "${ISG_PASSWORD}" | gcloud secrets versions add oracle-ebs-workflow-isg-password-dev --project=YOUR_PROJECT_ID --data-file=-

# 3. Build and push the container image, then deploy Cloud Run services
gcloud builds submit .. --project=YOUR_PROJECT_ID \
  --tag="us-central1-docker.pkg.dev/YOUR_PROJECT_ID/oracle-ebs-workflow-repo-dev/oracle-ebs-workflow-app:latest"
terraform apply \
  -var="container_image=us-central1-docker.pkg.dev/YOUR_PROJECT_ID/oracle-ebs-workflow-repo-dev/oracle-ebs-workflow-app:latest"

Configuring and accessing the agent in Gemini Enterprise

1. In Gemini Enterprise Apps, click Create App, enter application details, and select region.
2. In Overview (or Settings > Authentication / Resource access control), configure Google Identity or Workforce Identity Federation and assign target Google Workspace groups or user emails.
3. Click Agents > + Add agent, paste the rendered Agent Card JSON from Cloud Storage (gcloud storage cat "$(terraform -chdir=terraform output -raw a2a_agent_card_gcs_uri)") or /.well-known/agent.json (/a2a/card), click Preview Agent Details, and click Next (Display Name: Oracle EBS Workflow AI Assistant).
4. In Agent authorization and Tool authorizations, keep the default Basic HTTP authentication (and optional PKCE verification enabled) checkboxes or leave external OAuth credentials unconfigured, and click Next > Create (Cloud Run OIDC authentication is handled by Google-managed Discovery Engine and Cloud AI Companion service agents).
5. Click Go to Gemini Enterprise from Overview (or open Gemini Enterprise) and select Oracle EBS Workflow AI Assistant under Agents from your organization.

Note

Production security and least-privilege credentials:

• Gemini Enterprise Identity: End-user access is governed by Google Identity (or Workforce Identity Federation) and per-app IAM resource controls in Overview / Settings (the agent wizard's Basic HTTP authentication and PKCE checkboxes apply only to optional external OAuth 2.0 token exchanges).
• Cloud Run OIDC IAM: Public access (allUsers) is disabled. Services authenticate with Google Cloud OIDC tokens (gemini-ebs-workflow-sa-dev and ebs-workflow-agent-sa-dev) and route database traffic over Direct Virtual Private Cloud (VPC) Egress.
• Oracle EBS service account: Defaults (ISGADMIN, FND_USER_ID=0 / SYSADMIN) target EBS 12.2 Vision demo instances. For production, use a dedicated least-privilege FND_USER and workflow responsibility (ISG_USERNAME, ISG_PASSWORD, FND_USER_ID, FND_RESP_ID).

Enabling Oracle Integrated SOA Gateway (ISG) on EBS Vision

To execute workflow responses through Path B (REST) against a live Oracle EBS 12.2 instance:

1. Register the ISGADMIN password in FND_USER and FND_VAULT ('ISG' module) and grant the UMX|FND_IREP_ADMIN role.
2. Verify that the oafm_server1 WebLogic managed server is running on the EBS application tier (sudo su - oracle and source EBSapps.env run).
3. In Integrated SOA Gateway > Integration Repository, deploy the WF_NOTIFICATION (Workflow Notification System) PL/SQL interface as a REST service with alias wf (RESPOND, SETATTRTEXT, GETATTRTEXT, and GETINFO) and create user grants.
4. Set WORKFLOW_EXECUTION_PATH=REST and point ISG_REST_BASE_URL to your EBS web entry point (http://<ebs-host>:8000).

Important

How REST (Path B) and PLSQL (Path A) route queries and actions:

• PLSQL (Path A): Runs 100% of queries and APPROVE/REJECT actions directly through apps.wf_notification (apps_initialize) over port 1521 with zero ISG HTTP calls.
• REST (Path B, default):
  • Actions (APPROVE / REJECT): Calls ISG REST POST /webservices/rest/wf/setattrtext/ and /respond/, falling back to PL/SQL if ISG is unreachable or if the workflow requires an initialized FND session to transition to CLOSED.
  • Queries (/notifications, /triage): Seeded WF_NOTIFICATION methods only accept a single NID over HTTP POST (no GET /notifications collection exists in standard EBS 12.2 ISG), so queries fall back to PL/SQL and cache that state. To serve queries over REST, optionally expose apps.xx_ebs_wf_agent_pkg (src/workflow/sql/xx_ebs_wf_agent_pkg_spec.sql) through Oracle REST Data Services (ORDS) or register

  it in ISG with `irep_parser.pl` and `FNDLOAD`.

For complete SQL scripts, command-line steps, and resolution of common deployment errors (ISGADMIN vault configuration, node_info.txt permissions, FND_DIAGNOSTICS, and FND_SOA_SERVICE_OPERATIONS_PK constraint cleanup), see the Oracle EBS ISG setup and troubleshooting guide.

Recommended prompts for testing

• Worklist and notifications triage:
  • "List pending approvals for user JSMITH"
  • "Show my pending notifications"
  • "Triage pending notifications for JSMITH"
  • "Summarize workflow items needing my approval"
• Item inspection:
  • "Inspect workflow notification 10294"
  • "What are the details of notification 99482?"
  • "Show line item breakdown for purchase order 99120"
• Policy compliance and risk evaluation:
  • "Verify policy compliance for notification 10294"
  • "Check expense policy violations for notification 8472"
  • "Evaluate approval risk for PO notification 99120"
• Human-in-the-loop action execution:
  • "Approve notification 10294" (Triggers Human Confirmation Gate returning an advisory Proposal with risk status)
  • "Confirm approval for notification 10294 with comments 'Approved per project travel plan'" (Executes certified apps.wf_notification.respond)
  • "Reject notification 99120 due to budget overrun and unapproved vendor" (Executes certified rejection)

Example API invocations

# 1. Multi-item workflow triage
curl -X GET "http://localhost:8080/triage?recipient=JSMITH&limit=10"

# 2. Request approval proposal (confirmed: false triggers Human Confirmation Gate)
curl -X POST "http://localhost:8080/execute" -H "Content-Type: application/json" \
  -d '{"notification_ids": [10294], "action": "APPROVE", "comments": "Standard travel expense", "confirmed": false}'

# 3. Execute confirmed approval (confirmed: true invokes apps.wf_notification.respond)
curl -X POST "http://localhost:8080/execute" -H "Content-Type: application/json" \
  -d '{"notification_ids": [10294], "action": "APPROVE", "comments": "Confirmed within budget", "confirmed": true}'

Running automated tests

Run test suite verification against the integration suite:

./scripts/run_tests.sh