That creates a practical integration gap for teams trying to build deterministic AI workflows on top of Fusion ERP analytics.

This guide walks through the exact integration pattern required to connect Oracle AI Agent Studio Workflow Agents with OTBI SOAP APIs using External REST Functions.

Why this matters

Oracle AI Agent Studio gives you workflow nodes or tool(in case of supervisor agents) that can call external systems via REST. OTBI, however, speaks SOAP services— XML envelopes, namespaces. Bridging the two requires a specific pattern that isn’t obvious from Oracle’s Fusion AI documentation.

Once you set this up, your agents can pull live OTBI report data mid-workflow, use it for LLM reasoning, and push results to downstream systems — all in a single deterministic execution chain.

The Core Architecture: Two External REST Functions

Everything hinges on two main External REST Function configurations inside your Data Nodes for Workflow Agents:

1. SessionID External REST This calls the Oracle Analytics SOAP service to authenticate and receive a session token. No session = no report access. This step is always first, always mandatory.

EndPoint URL to be used — /analytics-ws/saw.dll/wsdl/v12?SoapImpl= nQSessionService

2. Run OTBI External REST Using the session ID from step one, this executes any OTBI report by passing the catalog path and session token into the SOAP envelope.

EndPoint URL to be used — /analytics-ws/saw.dll/wsdl/v12?SoapImpl= xmlViewService

Both are configured as External REST Functions inside AI Agent Studio. The node passes its output to the next step in the chain or used directly where needed— this is what makes the pattern composable.

NOTE: As needed, Content-Type can be added as application/soap+xml

The 4-Step Workflow

1. Step 1 — Generate Session ID: Call the Oracle Analytics SOAP endpoint with your Fusion credentials. The response includes a SessionID token you’ll pass forward.

2. Step 2 — Execute OTBI Report: Pass the session token and the exact OTBI catalog path (/Shared/Custom/YourReport/) into the Run OTBI SOAP call.

3. Step 3 — Parse the SOAP Response: The response is an XML document wrapped in a SOAP envelope. Extract the data rows from the relevant nodes. Pay attention to the data while parsing using LLM as columns logic will need to parse data accurately.

4. Step 4 — Use the Data: The parsed data is now available in your workflow. Pass it to an LLM node for analysis, a business object node for updates, or another external system.

If you want to see the node wiring in action before replicating it, this walk-through covers end-to-end: Data Nodes Explained

Critical Technical Notes

These will save you some time:

• Session ID must come first. Wire the SessionID node as a mandatory predecessor to the Run OTBI node.

• Catalog paths must be exact. A single character difference — trailing slash, case mismatch, wrong folder — returns an error. Copy the path directly from your OTBI catalog.

• XML response handling. The SOAP envelope wraps your data in nested elements. You’ll need to navigate the full path to reach the result rows — don’t assume a flat structure. Use LLM to parse this data.

  “If this is useful, subscribe to receive more content like this!

One Underrated Feature Worth Knowing

Anyone who’s tried to copy tabular output from a workflow result knows the drill: click, drag, miss a row, start over, wonder if the formatting survived. Oracle AI Agent Studio’s One Button Copy kills that entirely. One click copies the full output — however many rows — ready to paste wherever you need it. Less friction, faster iteration.

Here's a real example of it in action. I prompted the LLM node to return the OTBI report output in Markdown table format — and when I used One Button Copy and pasted directly into Excel, it landed as a clean, pipe-delimited dataset. Every column header intact, every row preserved. No reformatting, no manual cleanup. Just data, ready to work with.

Prerequisites

Before you configure anything in AI Agent Studio:

• Access to an Oracle Fusion environment

• Valid BI credentials with report execution permissions

• Your target report deployed in the OTBI catalog with the correct permission grants

What this unlocks

Once this pattern is in place, you can build AI agents that can:

• Pull live financial or operational OTBI data on demand

• Feed OTBI analysis or report output directly into LLM reasoning steps

Final Thoughts

As Oracle continues expanding AI Agent Studio capabilities, integration patterns like this become increasingly important for enterprises that still depend on SOAP-based ERP and analytics services.

The key takeaway is simple:

Legacy integration protocols do not prevent modern AI-native orchestration — provided the workflow architecture is designed correctly.

Workflow Agents combined with OTBI SOAP integrations create a powerful bridge between traditional enterprise reporting systems and modern AI-driven automation.

🔔 FOLLOW & SUBSCRIBE

If this was useful, here’s where you can follow along:

• 📺 YouTube — Step-by-step walkthroughs, demos & tutorials: LearnWithArpit — Oracle ERP & AI

• 💼 LinkedIn — Quick tips, updates & community discussion: Arpit Paliwal

• 📬 Substack — Deep-dive posts like this one: Subscribe below to never miss an issue

Subscribe now

• 📬 Substack — Loved deep-dive posts like this one: Share below with your friends

Share

⚠️ DISCLAIMER

All Oracle product names, screenshots, logos, and related imagery referenced or used in this article are the proprietary property of Oracle Corporation. They are used here solely for educational and informational purposes. This content is not affiliated with, endorsed by, or sponsored by Oracle Corporation.

Let’s walk through it feature by feature. For a more detailed video walk-through, check out the YouTube link below.

1. There's Now an "Apps" Section in the Navigation

The bottom navigation bar in AI Agent Studio now has a dedicated Apps tab. And this isn't just a cosmetic reshuffle — it represents a new way of thinking about what you can build inside the studio.

.Previously, the studio was largely about creating and managing agents and agent teams. The Apps section introduces the idea of a full Agentic Application — a composed workspace that brings multiple agents together into a single, purpose-built experience for end users.

What does that look like in practice? The list already comes pre-loaded with examples like Change Task Workspace, Cost Accounting Close Workspace, Design to Source Workspace, Sales Order Command Center, and a handful of HCM-focused workspaces like Manager Concierge and Team Talent Calibration and Review. All of them are published and ready to explore.

You can create your own via the + button or import existing ones — a clear signal that Oracle wants this to be a collaborative, shareable ecosystem rather than a closed set of Oracle-only features.

2. Workflow Agent Teams "Expose to Agentic Apps" Toggle

On the Details tab, there's a new toggle at the bottom: "Expose to Agentic Apps." Flip it on, and the agent team becomes available as a building block inside the Agentic Applications Builder. It's what connects the thing you build to the broader app experience. Small toggle, big unlock.

3. Expanded LLM Selection- Much More Control Over Your LLM

The LLM tab on the New Agent Team form has gotten a meaningful upgrade.

You now choose between two options: use the default model (currently GPT-4.1 Mini) or choose a model yourself. If you choose your own, a Provider dropdown appears — and among the options you’ll find newer models like GPT-5 Mini.

But the more interesting part is the Properties panel that appears below. It gives you three things to configure:

• K — a retrieval parameter (an integer between 0 and 1) - limits the model to choosing from the K most likely tokens

• Max Tokens — configurable up to 32,000

• Reasoning Effort — a dropdown with Minimal, Low, Medium, and High

• or Top p - dynamically adjusts based on probability

That Reasoning Effort control is the one worth pausing on. At Low, the model responds faster and uses fewer resources. At High, it takes more time to think through complex problems. Being able to set this per agent team means you can make intentional trade-offs — speed for simple tasks, depth for high-stakes ones — rather than defaulting to maximum for everything.

4. Security Tab: “For All Employees” Toggle + Role-Based Access

The Security tab on Agent Teams has always let you control who can access an agent by adding specific roles. That's still there, What’s new is the toggle sitting right above it: “For All Employees”.

Flip that on and you don't need to add any roles at all. Every employee in your org gets access to that agent team automatically. Flip it off — which is the default— and access falls back to role-based control, where you explicitly add the roles that should have visibility.

The two modes work cleanly together: toggle on for broad access, toggle off and use the + Add button to pin it to specific roles like COE_AI_AGENT_EXPLORER. You get the flexibility to govern each agent team exactly as tightly — or loosely — as it needs to be.

5. New “Tool” Node Types in the Workflow Agent Teams

Inside the workflow agent team canvas, when you add a new Tool node, the Type drop-down now exposes some new options:

• Deep Link — sends users to a specific page within Fusion Applications

• Chat Attachments Reader (new) — reads files that a user uploads during a chat

• Intent Change Indicator — detects when a user shifts topic mid-conversation

• User Session — accesses context from the current user session

The Chat Attachments Reader is the standout here. When you select it, you get a Conversation Identifier field and two toggles: Parse Files (on by default) and Get raw file content in Base64 format. In plain terms: your agent can now read documents that users drop into the chat — PDFs, spreadsheets, whatever — and actually do something with them as part of the workflow.

Earlier we have to create an Agent with MultiFileProcessor tool to be able to process run time file in Workflow Agent Teams. Now it’s a Tool node in your workflow.

6. "Edit in Playground" Is Now Right There on Agent Teams

This is a small UX change that makes a real difference in day-to-day work.

Published agent teams now show an “Edit in Playground” button directly in the header — right next to Save and Close and Publish. You no longer have to navigate away to test changes on a live agent.

The Playground gives you a sandboxed environment where you can run the agent, see its outputs, trace what steps it took, and adjust prompts — all without affecting what end users actually see until you’re ready to push an update.

It's the kind of thing that sounds small but quietly speeds up the whole build-test-iterate loop considerably.

7. A Proper ROI Dashboard in Monitoring & Evaluation

The Monitoring and Evaluation section has a new Value & Activities tab, and it’s one of the more genuinely useful additions in this release.

At the top, it shows four summary numbers: Total Time Savings (in hours), Total Cost Savings (in USD), Total Usage (interactions count), and Total Agents. To the right, two charts: one for total time saved by your agent teams (in hours) and one for total cost saved (in USD).

Where does it get these numbers? When you create a new Agent Team, there’s a section called “Capture estimated Cost and Time savings.” You enter how many minutes each agent run is expected to save and the estimated cost saved per run.

The system even suggests a default based on your company data (10 minutes is the starting recommendation). From that point on, every time your agent runs, those estimates accumulate in the Value dashboard.

It’s not a perfect measurement system — it’s based on your own estimates — but it gives you something concrete to point to when someone asks whether this AI investment is actually paying off. And having that answer ready matters.

The Activities tab also now has three filter buttons at the top right: Inbound Emails, Approval Emails, and Schedules. This signals that the studio is expanding beyond just real-time chat interactions into email-triggered and scheduled automation — a quiet but important step toward broader workflow coverage. This is specially useful for Triggers for your agent teams.

8. A Connectors Library Inside Credentials

There’s a new tab introduced inside the Credentials section — called Connectors.

The idea is straightforward: instead of manually wiring up knowledge sources every time you build an agent, you now have a library of reusable connectors that agents can pull from. Hit + Add Connector to create your own, or use the five that Oracle has already seeded in there.

Once your connector is added above, Use a new tool type called Connector. You can select this tool type and associate the previously added connector to enable its use within your agent using this tool.

The Bigger Picture

Taken together, these updates sketch out a clear direction for where Oracle AI Agent Studio is heading. It’s moving from a tool that helps you build individual agents into a platform that lets you compose full agentic workflows, measure their impact, and surface them to end users as polished application experiences.

The LLM controls bring precision. The Security can be used for all employee and accordingly cost usage will be charged.The Chat Attachments Reader brings document intelligence. The ROI dashboard brings accountability. And the Agentic Apps layer ties it all together into something that looks less like a developer tool and more like an enterprise AI platform.

If you’ve been waiting to dig in, the tooling has genuinely matured. Now’s a good time to start.

YouTube Video Explained in Detail: Oracle Fusion 26B: AI Agent Studio New Features You Need to Know

Have questions about any of these features or want to share what you're building? Drop a comment — always happy to compare notes with others in the Oracle community.

🔔 FOLLOW & SUBSCRIBE

If this roadmap was useful, here’s where you can follow along:

• 📺 YouTube — Step-by-step walkthroughs, demos & tutorials: LearnWithArpit — Oracle ERP & AI

• 💼 LinkedIn — Quick tips, updates & community discussion: Arpit Paliwal

• 📬 Substack — Deep-dive posts like this one: Subscribe below to never miss an issue

  Subscribe now

• 📬 Substack — Loved deep-dive posts like this one: Share below with your friends

  Share

Oracle announced AI Agent Studio at last year in March 2025 — the first enterprise-wide platform to build, deploy, and manage AI agents across the entire Fusion suite (ERP, HCM, SCM, CX). By March 2026, it had an Agentic Applications Builder, ROI dashboards, contextual memory, and 300+ seeded templates.

Most functional consultants are still unsure—or afraid—of building AI agents. That hesitation will cost them project relevance.

This 7-week, foundation-first roadmap will take you from “I’ve heard of AI agents” to “I can build, evaluate, and deploy real agent teams in Oracle Fusion” — without the overwhelm.

Week 1: Foundation First (Non-Negotiable)

Skipping the foundation is why most AI agent builds fail and you can not build reliable AI agents without understanding how LLMs functions, where they hallucinate, and how prompts shape behavior.

• Prompt Engineering: Learn how to craft effective prompts — the instruction layer that drives every agent's behavior inside Oracle AI Agent Studio.

The Oracle AI Agent Studio "Topics" and "Instructions" are prompt-driven.

Good Resources to check out:

🔗 Tina Huang — Google's 9-Hour Prompt Engineering Course in 20 Minutes

🔗 Check out my post on Topics vs Prompts Comparisons

• LLM Fundamentals: Understand how large language models work — tokenization, context windows, reasoning, and limitations.

🔗 Resource: Andrej Karpathy — Intro to Large Language Models

• Guardrails & Hallucination Control: Oracle AI Agent Studio lets you add guardrails via prompt instructions and Topics. Without knowing why hallucinations happen, you can't design around them.

Quick insight: Oracle supports multiple LLM models — Llama, OpenAI— each with different behavior profiles within Oracle AI Agent Studio. Knowing LLM fundamentals helps you choose and configure the right one. You can also Bring your own model (BYOM) using API Key.

Week 2: Oracle REST APIs & Security Roles

API knowledge and security role mastery are the practical “plumbing” that makes agents actually work in Fusion.

Every tool an AI agent uses in Oracle AI Agent Studio is either a REST API call, a document query, or a business object interaction. If you don’t understand REST, your agents will be blind.

• Oracle REST API Basics:

  • Endpoints, HTTP methods (GET, POST, PUT, DELETE, PATCH), payloads, query parameters

  • JSON structures — agents pass and receive data in JSON

  • Pick your domain focus: ERP (Financials), HCM, SCM, or CX — each has its own API catalog

  • Learn about header parameters within API —optional

• Security Roles — the overlooked killer:

  • Duty Roles → Job Roles → Privileges → Permission Groups

  • Agents inherit Fusion security. If the role powering an agent doesn’t have the right privileges, the agent silently fails or returns empty results.

  • Critical: Permission Groups must be enabled on custom roles for AI Agent Studio access

Week 3: Agent Architecture — How the Pieces Fit Together

Learn about clear mental model of how Oracle AI Agent Studio is structured.

• The core components of AI Agent Studio:

  • Tools — Capabilities agents use to fetch or act on data: Business Objects, Documents (RAG), External REST APIs, Email, MCP tools

  • Topics — Subject-area definitions with step-by-step instructions guiding agent behavior. Topics are reusable across multiple agents

  • Prompts — The instruction layer shaping how agents respond

  • Business Objects — Data entities (APIs) that must be created and linked to a tool before an agent can use them

  • LLM Nodes — The reasoning engine; can be configured per agent

Resources to check out:

🔗 Components of AI Agent Studio

🔗 Check out my YouTube series on how to create each Tools

• Two Agent Team Architectures:

  • Supervisor Architecture: A supervisor agent routes user intent to the right worker agent. The supervisor uses prompts and topics to direct requests across multiple specialized workers. Best for: dynamic routing, conversational interfaces.

  🔗 Resource: Create Custom AI Agents of Type Supervisor

  • Workflow Architecture: Sequential, branching, and parallel execution paths with conditional logic, variable handling, and multiple LLMs. Best for: deterministic, multi-step business processes with predictable outcomes.

🔗 Resource: Create Custom AI Agents of Type Workflow

🔗 Resource: Workflow Agent Complete Guide- YouTube Vedio Series

• Single Agent vs. Multi-Agent:

  • Single agent team: one agent, one tool — for simple, focused tasks

  • Multi-agent teams: complex orchestration where multiple specialized agents collaborate

  • Agent composition: Agent nodes and Workflow nodes let you embed published agents inside larger workflows (system-of-systems)

Key rule: Every agent must have at least one tool. Topics are optional but powerful for guiding complex behaviors.

Week 4: Build Your First AI Agents & Teams

Get hands-on with a clear, progressive build sequence that avoids overwhelm.

• Part 1: The build order matters:

  1. Start with a single agent + single tool (e.g., one REST API call or one Document Tool)

  2. Test it in the Playground, iterate on the prompt (Only from 26B)

  3. Build a single-agent team (wrap the agent in a team so it can be deployed and invoked) and run in debug mode

  4. Add a Supervisor with 2–3 worker agents

  5. Graduate to Workflow agent teams for multi-step logic

• Part 2: Build Supervisor First, Then Workflow:

  • Supervisor teams are more forgiving and faster to build — great for learning routing logic

  • Workflow teams require thinking in steps, conditions, and variables — but unlock powerful deterministic execution

🔗 Resource: Workflow Agent Complete Guide- YouTube Vedio Series

• Part 3: Extending pre-built agents:

  • Oracle ships 300+ seeded agents across ERP, HCM, SCM, CX

  • You can extend any by adding Documents, Tools, Prompts, or APIs without starting from scratch

  • Use “Use Template” or “Copy Template” to deploy a copy you can customize.

Practical build checklist:

• Create Business Object definitions first (if using BO tools)

• Run Process Agent Document schedule process if uploading documents

• Publish each agent before embedding it in a team

• Test routing logic via Debug Mode in Agent Team or Playground

Week 5: Evaluate, Monitor & Measure ROI

Agent evaluation is what separates experimental projects from production-grade deployments — and that Oracle has built-in tools to do it.

• Part 1: Why evaluation is non-negotiable:

  • Most agent pilots fail before production — not because the agent is bad, but because no one measured or run evaluation for it. You need data before you deploy.

  • Oracle AI Agent Studio has a built-in evaluation framework specifically for this with Monitoring & Evaluation Dashboard

🔗 Resource: Monitor & Evaluate docs

🔗 Resource: Built-in observability overview

• Part 2: What to evaluate:

  • Response Correctness — Does the agent answer accurately against a QA dataset you build?

  • Response Time — Is the agent fast enough for real-world use?

  • Token Usage — What is the cost per agent run?

  • RAG Metrics — How effectively does the agent use retrieved document context? (relevant if you’re using the Document tool)

  • Prepare QA set and Compare multiple evaluation runs to track improvements over iterations

• Part 3: Build your QA dataset:

  • Create a set of question-and-answer pairs specific to your domain (Finance, HCM, SCM)

  • Cover edge cases, ambiguous inputs, and error scenarios

  • Re-run evaluation after every significant prompt or tool change

• Part 4: The Playground — Your Iteration Engine:

  • Available in 26B: Edit in Playground to refine prompts and test LLM nodes directly within AI Agent Studio without re-deploying

  • Real-time visibility, debugging, and agent behavior testing

🔗 Resource: Playground in AI Agent Studio

• Part 5: ROI Dashboard (26B onward):

  • New Agent ROI Dashboard shows time saved, cost savings, and productivity gains per agent across workflows and teams

  • This is how you make the business case — go beyond “it works” to “here’s what it saves”

🔗 Resource: Measure AI Impact

Week 6 (Optional) & Week 7: Advanced Protocols + Agentic Applications

Frontier — where Oracle AI Agent Studio is heading and why knowing MCP/A2A and Agentic Applications future-proofs your skills.

Week 6 — MCP, A2A & Cross-Platform Integration (Optional)

• MCP (Model Context Protocol): Open industry standard that lets AI agents securely connect to third-party tools, data sources, and LLMs without custom integration code. Oracle officially supports MCP Tool in AI Agent Studio — enabling agents to reach beyond Fusion.

  • Oracle Integration can also act as an MCP server along with other publicly available servers, letting agents call reusable automation components

  • 🔗 Resource: Oracle Integration as MCP Server

  • 🔗 Resource: MCP support in AI Agent Studio

• A2A (Agent-to-Agent Protocol): Enables cross-agent collaboration — a Finance Agent can pass context directly to a Supply Chain Agent without human intervention. Supported via standardized A2A agent cards in Oracle AI Agent Studio.

  • Resource: 🔗 A2A agent cards

• Chat integrations: Agents can be invoked via enterprise chat platforms (Microsoft Teams, Slack) — bringing agent capabilities directly into daily workflows

🔗 Resource: Microsoft Teams Integration with AI Agent Studio

Week 7 — Oracle Fusion Agentic Applications

• What are Agentic Applications?

  • A new category announced March 2026: outcome-focused applications built from agents — not just apps with AI bolted on

  • Built using the new Agentic Applications Builder (natural language-based, no traditional coding required)

  • Combine Oracle agents, partner agents, and external agents into unified applications

  • Include workflow orchestration, contextual memory (agents remember context across interactions), content intelligence (unstructured + structured data), and LLM multimodal capabilities (text, images, audio, video)

    🔗 Resource: Agentic Applications Builder

• Why this matters for consultants:

  • This is the product-level layer where agents become enterprise applications

  • Knowing how to build agents in Weeks 1–5 directly prepares you to architect Agentic Applications

  • Oracle is positioning this as the future of Fusion — “active applications that run the business”

🏁 CLOSING — Strong Takeaway + CTA

The consultants who build these skills now won’t just be keeping up — they’ll be the ones leading the room when the client asks, “Can we automate this with AI?”

Key takeaway to drive home:

• Oracle AI Agent Studio is not a future tool. It is live, available at no additional license cost, and used in production across ERP, HCM, SCM, and CX today.

• The 7-week roadmap is designed to be sequential and cumulative — each week builds on the last. Skipping Week 1–2 is why most agents break in production.

• Your goal by Week 5 is not just a working agent — it’s an evaluated, production-ready agent with measurable ROI.

🔔 FOLLOW & SUBSCRIBE

If this roadmap was useful, here’s where you can follow along:

• 📺 YouTube — Step-by-step walkthroughs, demos & tutorials: LearnWithArpit — Oracle ERP & AI

• 💼 LinkedIn — Quick tips, updates & community discussion: Arpit Paliwal

• 📬 Substack — Deep-dive posts like this one: Subscribe below to never miss an issue

Subscribe now

• 📬 Substack — Loved deep-dive posts like this one: Share below with your friends

  Share

⚠️ DISCLAIMER

All Oracle product names, screenshots, logos, and related imagery referenced or used in this article are the proprietary property of Oracle Corporation. They are used here solely for educational and informational purposes. This content is not affiliated with, endorsed by, or sponsored by Oracle Corporation.

But the moment you ask a mechanic to swap the engine for a V12 or add a custom turbocharger; that’s a different conversation. It’s a different price tag.

That’s essentially what Oracle is doing with AI Agents in Fusion. Once you understand this analogy, the entire licensing model becomes clear.

First, let’s talk about What your Fusion subscription already gives you?

Let’s take a moment to appreciate what is already there before we talk about what’s “custom”. and what making a mistake here could cost you.

When you’re on Oracle Fusion, you gain access to Oracle-built agent templates - think of them as fully assembled AI employees, ready to deploy. For example, a Benefits Analyst agent or a Supplier Quote assistant. These templates come with job descriptions, tools, and built-in training.

You also enjoy unlimited access to the base LLM- earlier, it’s Llama 3.3, moved to an open-source GPT model in 26A i.e. GPT-OSS— hosted directly on OCI. You can test it, query it, or develop with it without cost.

The premium models—GPT-5 mini and GPT-4.1 mini—do count against your token usage. Here’s an important point: token usage is usage, period. Whether you’re using an Oracle-built template or one you created from scratch, if it’s a premium model, it will be counted. Keep this in mind during extensive testing.

So when does “included” become “custom”?

Here’s a simple way to think about it:

> If you’re tailoring the agent to your organization → Included.

> If you’re expanding what the agent was designed to do → Custom.

Let’s go through some real examples.

✅ The things that are completely fine (included)

Lets assume you have cloned the Benefits Analyst template. Now what?

• You can add your company’s benefit plan names to the prompts. Of course as an example it doesn’t know you as company call it “EmployeeCare Plus.”

• You can upload your own policy documents to the document tool. In fact, you must do this. Oracle doesn’t provide those.

• You can add guardrail topics—like “politely decline if the question is off-topic” or “always maintain a professional tone.” These don’t expand what the agent does; they just improve its behavior.

• You can add an email tool or deep link tool. Oracle recognizes these & considered as minor adjustments.

• You can add custom fields to an existing business object tool.

• You can remove agents, tools, or topics from a template. Subtraction is always allowed.

Think of this as: I’m making this agent function in my environment. That’s included.

🚫 The line you don’t want to cross (custom territory)

Now let’s say someone in your technology team has a new idea.

Just like that, it’s custom.

Why is this the case? Because you’re no longer tailoring it. You’re expanding the scope. You’re asking an agent originally designed and trained for one purpose to now perform two or three tasks it wasn’t built for.

Here’s what else indicates you’re in custom territory:

1. Changing the role and persona of the agent—this defines the agent’s identity. If you rewrite this, you’re effectively creating a different agent.

2. Adding a new business object function—Payable invoice tool was built to *retrieve* invoice records. Adding a “create” or “delete” function? That’s a change in scope.

3. Using REST or MCP tools—these connect to external services Oracle didn’t provide. Clearly custom.

4. Adding a brand-new document tool—lets consider a scenario where one document tool per agent is typically what was designed. Adding a second one for a completely different domain is expanding the scope.

5. Bringing your own LLM—if you want to use Claude or Gemini, it’s fully supported in 26A but is considered a custom SKU. (Your tokens with those vendors are separate—you won’t be charged by Oracle for those.)

6. Non-text output—generating images, audio, video, or podcasts. These use significantly more expensive models and are labelled custom.

   Subscribe now

What about workflow agents?

The same principle applies, just with different terms.

Adding nodes to a workflow agent that involve LLM processing, agent calls, document retrieval, external services, or vector databases is custom.

However, adjusting flow control—like adding a branch to a switch or tweaking a parallel node—is generally fine. You’re not changing what the agent does; you’re changing how it routes.

The Development/Testing safety net

One thing that makes this model more forgiving than it sounds: none of this matters until you hit production.

You can build, break, experiment, and rebuild custom agents all day in development and test environments without any charges. The clock only starts when you deploy.

This is a significant advantage for partners and implementers. You can prototype and validate a custom solution before your customer commits to a license.

How do you pay for it when it is custom? or SKUs?

There are two pricing options:

1. Named user—you pay based on the number of users who have the role that grants access to the agent. Whether they use it or not, if they have access, they count. This is the default model for most agents.

2. All employee—cheaper per unit, but it’s charged against your entire headcount. This option is best for agents you want everyone to access, like a benefits enrollment or a company policy Q&A.

In 26A, you will be able to explicitly mark an agent as intended for all employees, linking it to the all-employee SKU automatically.

The multi-modal question: A handy cheat sheet

This question comes up often, so Oracle put together a table to clarify. Here’s how it breaks down:

The pattern is clear: consuming and retrieving content (even visual content) is included. Generating new media beyond text crosses into custom.

The takeaway (for everyone in the room)

Whether you’re a VP making a build-vs-buy decision, an implementer scoping a project, or a customer wondering why your pilot got flagged, the mental model remains the same:

Tailoring = Included.

Expanding = Custom.

Oracle put significant effort into these templates. They’re scoped, tested, and designed for specific use cases. The moment you push them beyond that—by giving them new tools, new domains, or new identities—you’re creating something new. And new things come with a cost.

That’s not a surprise. That’s how quality software should function.

If you found this useful, subscribe below for more Oracle & AI insights like this each week.

Subscribe now

So What Is an Email Account in Oracle AI Agent Studio?

Once you configure an Email Account inside Oracle AI Agent Studio, you’re essentially telling the system: “Watch this inbox. When something arrives — act on it.”

And what it can act on covers two very different — but equally powerful — use cases.

Two Use Cases. One Email Account Feature

This is the part most people miss when they first encounter this feature. An Email Account in Oracle AI Agent Studio isn’t just for one thing. It supports two distinct workflow scenarios, and understanding the difference is key.

Use Case 1 — Inbound Account (Trigger Workflows Automatically)

This is for fully automated, no-human-needed processing.

You configure an email account as Inbound, and the AI agent monitors that mailbox at a set pooling interval — every 5, 15, or 30 minutes. The moment a new unread email arrives in the designated folder, the agent wakes up, reads the email body, parses any attachments, and kicks off the workflow automatically.

A real example: your payable team receives supplier invoice requests via email — usually a pdf with invoice details. Instead of someone manually logging into Oracle Fusion and creating record or using IDR, the agent does it the moment the email arrives. It can even sends a confirmation back to the original sender when it’s done.

No manual trigger. No batch job to babysit. Just email in, automation out.

Use Case 2 — Approver Account (Human-in-the-Loop Workflows)

This one is for scenarios where a human decision is required before the workflow can proceed.

You configure an email account as an Approver, and assign it to a Human-in-the-Loop node inside your workflow agent. When the workflow reaches that node, it sends an email to the approver. They can approve or reject the request directly from their inbox — no need to log into any system.

Think of it as giving your approvers a frictionless way to participate in enterprise workflows without changing how they work. They just respond to an email, like they always have.

One important rule to remember: the same email account cannot be used for both Inbound and Approver configurations. They serve different purposes and must be set up separately.

How to Set It Up (The Short Version)

You don’t need to be a developer to configure this. Here’s the gist:

Step 1 — Prepare your mailbox. Whether you’re using Gmail or Microsoft, you need to register an application with your email provider and collect three credentials: a Tenant ID, Client ID, and Client Secret. For Google specifically, the Tenant ID maps to your Google Cloud Project ID — something that catches a lot of people off guard the first time.

Step 2 — Connect the mailbox to Oracle AI Agent Studio. Inside the studio, navigate to Credentials → Email Accounts, add a new account, select your provider, choose whether it’s Inbound or Approver, paste in your credentials, and set your pooling interval. This is the secure bridge between your inbox and Oracle.

Step 3 — Build your workflow agent. Create a workflow agent and assign the email account — either as a trigger (for Inbound) or to a Human-in-the-Loop node (for Approver). Then define what the agent should do — parse data, call a business object, send a notification, handle errors.

Step 4 — Test, then publish. Always use debug mode first. For Inbound accounts, the email trigger only activates once the agent is in a Published state. Don’t skip this step.

Why This Matters Beyond Oracle

Even if you’re new to Oracle Fusion, the pattern here is worth paying attention to.

Email is still the most universal business communication channel. It’s how field teams submit requests. It’s how vendors send invoices. It’s how partners share data. Building automation that starts from an email — rather than requiring people to log into a portal or fill out a form — dramatically lowers the barrier to adoption.

The best automation is the kind people don’t have to change their behaviour for. They just send an email, like they always have. The system does the rest.

Share Arpit Paliwal – Oracle, AI/ERP

Key Things to Keep in Mind

A few practical notes before you get started:

• The same email account cannot be used for both Inbound and Approver configurations — always set them up separately

• Don’t assign the same Inbound email account as a trigger to multiple workflow agents — it leads to unpredictable processing

• The agent processes records using the Fusion role security of the user who created the Email Account — so permissions matter

• Once an email is read by the agent, it’s marked as “Read” so it won’t be processed twice

• Always test in debug mode before publishing — for Inbound accounts, you can simulate the trigger using a web-hook test message

Final Thought

Your shared inbox has been a passive communication tool for too long.

With Oracle AI Agent Studio’s Email Account feature, it becomes an active automation gateway — triggering workflows, routing approvals, processing attachments, and responding to senders, all without any manual steps in between. Whether it’s supplier onboarding, invoice processing, data validation, or multi-step approval workflows — if it starts with an email, Oracle AI Agent Studio can automate it end to end.

The setup is simpler than you’d expect. The impact is bigger than you’d think.

Subscribe now

Want to see this configured from scratch? I’ve put together a step-by-step video walkthrough covering Gmail OAuth setup, the full Email Account configuration in Oracle AI Agent Studio, and a live demo inside Oracle Fusion. Check it out on my YouTube channel using the link below.

YouTube Video: Email Account Config Vedio

Found this useful? Subscribe for more deep dives into Oracle Fusion, AI Agent Studio, and enterprise automation.

Your finance team arrives, grabs their coffee, and sits down expecting a clean start to the week. Over the weekend, an AI agent was supposed to run — processing a batch of invoices, updating supplier records, handling the payables queue that had been building up since Friday.

Nobody watched it run. Nobody needed to. That was the whole point of automating it.

But somewhere between Saturday night and Sunday morning, something went wrong. A field was missing. An API call timed out. The agent stopped mid-way through — silently, without telling anyone.

And now it’s Monday. The records aren’t updated. The invoices aren’t processed. And your team is about to spend the first two hours of their week figuring out what happened, when it happened, and how much needs to be redone.

Sound familiar?

This is the moment that Success and Error Handling Emails in Oracle AI Agent Studio were built for.

Let’s Rewind That Monday Morning

Same scenario. Same AI agent. Same weekend run.

Except this time, the agent is configured with an Error Handling Email.

The moment it fails — at 2:47am on Sunday, at the exact step where the API call timed out — an email lands in the support team’s inbox. The subject line reads: Payables Processing Agent — Error.

Inside the email: the name of the agent, the step where it failed, the exact error message, a workflow trace ID for deeper investigation, and a recommended action — what to do next to resolve it and rerun the process.

By the time your team arrives on Monday morning, the issue has already been flagged, investigated, and in many cases — fixed. The records are processed. The week starts cleanly.

That’s the difference a single configuration makes.

But What About When Things Go Right?

Now imagine the same agent runs perfectly. Every record processed. Every invoice created. Every update applied — exactly as expected.

Without a Success Email, nobody knows. The agent ran silently in the background and finished silently in the background. Your business users have no confirmation. Your manager has no visibility. And if someone asks “did the Saturday run complete?” — the answer is a login, a search through logs, and a few minutes of uncertainty.

With a Success Email configured, your team wakes up to something like this in their inbox:

“Payables Processing Agent completed successfully at 3:12am. 47 invoices processed. 0 errors. Summary attached.”

Thirty seconds of reading. Full peace of mind. No login required.

That is what good automation feels like from the outside.

Two Emails. Two Very Different Purposes.

Now that you’ve felt the difference, let’s understand how each one works inside Oracle AI Agent Studio.

The Success Email is attached to the end of your workflow as a communication node. When your agent reaches the final step and everything has gone well, this node fires and sends a summary to whoever needs to know. You control the recipient, the subject line, the content — and you can pull in dynamic details like the agent name, completion time, and output summary directly from the workflow context.

Keep it short. Keep it meaningful. Your business users don’t need a technical breakdown — they need a clear confirmation and the numbers that matter.

The Error Handling Email works differently. It lives in the settings of your agent team, not inside the workflow itself. This is your safety net — if anything fails anywhere in the workflow and there’s no specific handler to catch it, this email fires automatically. You define who receives it, what the subject says, and how much detail to include.

And here’s the part most people don’t think to add: a recommended action. When you’re building the agent, you know better than anyone what could go wrong and what someone should do about it. Write that down in the email template. When your support team receives the error at 2am, they’ll thank you for it.

There Are Actually Two Levels of Error Handling

Here’s something worth knowing as your workflows get more sophisticated.

The global error handler is the one we just talked about — it sits at the agent team level and catches anything that isn’t handled elsewhere. Think of it as the last line of defense. Always configure this one. Always.

But Oracle AI Agent Studio also gives you node-level error handling — the ability to define what happens when a specific step in your workflow fails. Maybe one node handles a complex API call that needs a very specific error response. Maybe another node processes user input that could fail in a particular way. Node-level handlers let you catch those scenarios precisely, capture custom variables, and surface the right information in your error email.

Start with the global handler. Add node-level handling as your workflows grow. Together, they give you complete visibility at every layer of your automation.

For success notifications at the end, Use Send Email Node

Four Things to Get Right From the Start

Before you go and configure these, here are four things worth keeping in mind:

Send the error email to the right people. Your business users care about outcomes. Your tech team cares about errors. These are different audiences with different needs. Configure your recipient lists to match — don’t send a stack trace to a finance manager, and don’t send a “completed successfully” summary only to IT.

Keep success emails scannable. The goal isn’t to impress anyone with detail. It’s to give someone thirty seconds of reading that replaces five minutes of manual checking. A clean subject line, a one-line confirmation, and the key numbers. That’s it.

Make every email actionable. Whether it’s a success or an error, the person reading it should know what — if anything — to do next. For success emails, that might mean linking to the output. For error emails, that means a recommended action that actually helps someone resolve the issue faster.

Always configure error emails for critical processes. If an agent runs automatically and its output affects your business, there is no good reason not to have an error email in place. It takes ten minutes to set up. The alternative — discovering a failure three days later — takes significantly longer to recover from.

Want to see this configured from scratch? I’ve put together a full video walk-through showing exactly how to set up both Success and Error Handling Emails in Oracle AI Agent Studio — including a live demo of how both emails look when they land in your inbox. Check it out on my YouTube channel using the link below.

YouTube Video: Success & Error Handling Vedio

If this resonated, share it with someone on your team who’s running automation without notifications in place — the share button is right below. 🙂

Share

Subscribe for more practical Oracle AI Agent Studio content, every week.

Subscribe now

Here’s a breakdown of everything that shipped.

AI Agent Studio: The Foundation Gets Stronger

Before diving into the domain-specific agents, it’s worth noting that the underlying platform — AI Agent Studio — received four meaningful upgrades that raise the ceiling for what’s possible.

Build Agentic Apps (New)

You can now create multi-agent applications directly in Agent Studio. Instead of relying on a single AI to handle everything, you compose purpose-built agents — a Sales agent, a Finance agent, an Inventory agent — that collaborate behind the scenes to power a unified user experience. Each agent maintains deep, domain-specific context. The result: more accurate outputs, better reasoning, and a foundation that scales as your needs grow.

Playground: Edit and Test Faster (New)

This one is big for developers and implementers. The new Playground lets you isolate and test individual nodes — supervisor agents, worker agents, nodes — without running the full workflow from start to finish. You can tweak prompts, adjust parameters, validate outputs in real time, and track every change through Run History. What used to take multiple full-cycle runs can now be iterated in seconds.

Workflow Agent Enhancements

Three new node types extend what’s possible in workflow agents: a Tool node (call tools like Deep Link, Chat Attachments Reader, or User Session), a While node (loop execution until a condition is met), and a Wait node (pause until an action completes or a set number of iterations is reached). These additions unlock genuinely complex, multi-step agentic patterns that weren’t possible before.

Predefined Topics for Guardrails (New)

Oracle now ships three Oracle-managed topics you can plug into any agent: Professional Voice & Tone, Age-Neutral Output, and Gender-Neutral Output. These act as built-in guardrails, ensuring your agents communicate in ways that are appropriate, inclusive, and consistent — without having to write custom instructions from scratch.

Share Arpit Paliwal – Oracle, AI/ERP

Expenses Agent: End-to-End Automation

The Expenses Agent received three meaningful upgrades in 26B, each targeting a different friction point in the expense lifecycle.

Email-Based Expense Completion (New)

Employees can now forward receipts directly to the Expenses Agent via email. The agent processes the receipt, creates the expense, and — if anything is missing (attendees, cost center, justification) — sends a follow-up email asking for the specific details it needs. The employee replies, the agent fills in the gaps, and the expense is submitted automatically once everything checks out.

This eliminates the loop of logging into the UI, hunting for a receipt, manually filling fields, and chasing down missing information. For organizations with high-volume expense submissions, this is a material productivity gain.

Auto Itemization and Spend Authorization

Hotel accommodation receipts are now auto-itemized by default — the agent breaks out individual charges line by line, both in the email summary and in the desktop application. Additionally, employees can now attach pre-approved spend authorizations directly to expenses through the agent workflow, supporting organizations that require pre-approval before spending occurs.

Policy Inquiry Enhancements

The expense policy inquiry capability — which lets employees ask natural language questions about corporate travel and expense policies — has moved into Agent Studio with a refined prompt for higher-quality responses. Managers can upload policy documents (PDF, TXT, or Markdown), and the agent delivers instant, contextual answers to employee questions before expenses are even incurred. Fewer policy violations. Less back-and-forth.

Ledger Agent: AI-Powered General Ledger

This is a brand-new agent in 26B, and it’s arguably the most ambitious launch in this release. The Ledger Agent brings proactive, conversational AI to the General Ledger — an area of finance that has historically required navigating multiple screens, reports, and tools to get a clear picture.

Proactive Monitoring

Accountants can set up monitoring prompts — using a catalog of predefined templates or their own custom variations — to continuously scan for conditions like accounting process exceptions, manual journals posted without approvals, or revenue anomalies. When a condition is met, the Ledger Agent surfaces an insight automatically. You define the frequency. The agent does the watching.

Natural Language Balance Inquiry

Instead of navigating to multiple reports, accountants can simply ask: “What are the total expenses for cost center 1200 this quarter compared to last year?” or “Show me journals posted to account 5100 in March.” The agent understands financial entities, generates the right query, and responds with contextualized answers. Follow-up questions are supported conversationally.

Root-Cause Insight Explanations

When an anomaly or exception surfaces, the Ledger Agent doesn’t just flag it — it explains it. You get a rich, contextual summary of what happened, correlated data, and suggested next steps. This reduces the time accountants spend investigating and assembles the “why” that previously required pulling data from multiple sources.

Payable Agent: Touch-less Invoice Processing

The Payables Agent is Oracle’s most comprehensive agentic launch to date on the financials side. It re-imagines the entire invoice lifecycle — from capture to payment readiness — with automation at every stage.

Touchless Invoice Ingestion

Invoices can be received from multiple sources and in multiple formats. The agent uses LLMs to recognize fields automatically, and adaptive training handles edge cases. A new Streams view gives real-time visibility into ingestion status across all incoming invoice flows.

Compliance and Control via Natural Language

Here’s where it gets interesting: administrators can write Standard Operating Procedures in plain language, and the agent converts them into structured defaulting rules and anomaly detection patterns. No coding. No configuration forms. Just describe what you want the system to enforce, and it creates the rules.

Anomaly checks run automatically at invoice creation. If something looks off — duplicate invoice, unusual amount, mismatched supplier — the invoice is flagged and held before it can move toward payment.

Share Arpit Paliwal – Oracle, AI/ERP

Unified Invoice List

All invoice management tasks, exceptions, and queries now live in a single centralized view. Smart filters, natural language queries, and direct action buttons let teams work through exceptions without switching between screens or applications.

Payments Agent: From Tactical to Strategic

The Payments Agent transforms how organizations think about outgoing payments. Rather than treating payment as a scheduling function, the agent introduces real-time decision support — helping teams evaluate financial options before committing to a payment approach.

The agent is composed of three task-specific assistants:

Payment Options Assistant

Given a set of outstanding payable installments, this assistant evaluates which payment programs — dynamic discounting, virtual credit cards — offer the best financial benefit relative to your working capital goals. It calculates ROI comparisons, explains the trade-offs, and helps you select the right approach transaction by transaction.

Supplier Offers Assistant

Once an opportunity is identified, the Supplier Offers Assistant takes over. It drafts personalized early-payment offers, communicates them to suppliers, and tracks responses through the full lifecycle. Accepted offers automatically update installments so the negotiated terms carry through to the next payment run.

Payment Execution Assistant

This assistant guides users through scheduling, submitting, and monitoring payment process requests — all via natural language. It surfaces exceptions, delays, acknowledgment failures, and follow-up actions so issues are caught early rather than discovered after the fact.

An Insights dashboard ties all three together, giving teams a real-time view of what requires attention and routing them directly to the relevant assistant.

Partner E-Invoicing: Thomson Reuters Integration

Rounding out the release is a significant compliance capability: a pre-configured, turnkey integration with Thomson Reuters for global e-invoicing.

Organizations operating across multiple countries face a fragmented e-invoicing landscape — different standards, different government platforms, different service providers. This integration handles the complexity. Customers can receive e-invoices from suppliers worldwide through Thomson Reuters’ network, with automated acknowledgments for both processed and rejected invoices.

In Release 26B, support covers eight PEPPOL network countries operating under a model without clearance: Australia, Belgium, Denmark, Finland, the Netherlands, Norway, Sweden, and the United Kingdom.

No custom development. No complex integration testing. Onboarding is turnkey, and it sits squarely within Oracle’s broader Touchless Payables initiative.

What This All Means

Taken individually, each of these features is useful. Taken together, they signal something more significant: Oracle is building a finance function where the AI doesn’t just answer questions — it monitors, acts, escalates, and executes.

That's the direction Release 26B is pointing. And based on what's in the pipeline (expanded program support, standing offers, broader PEPPOL coverage, Payables insights view), 26C will push it further.

NOTE: All features referenced in this post are documented from Oracle’s official 26B What’s New release notes.

What this guide covers

• Creating Dropbox OAuth credentials

• Setting the correct redirect URI for Oracle Cloud AI Agent Studio

• Enabling the required Dropbox scopes

• Access Dropbox in Oracle AI Agent Studio

⏱️ Total setup time: under 5 minutes

Subscribe now

Step 1: Identify required credentials (Client ID & Secret)

1. Open Agent Studio

2. Navigate to Chat Experience

3. Select Drop Box

4. Click Add New Account

You’ll see two required fields:

• Client ID

• Client Secret

These will be generated from Dropbox Developers Console.

Step 2: Create a Dropbox Application

1. Go to Dropbox Developer Portal

2. Click New App

3. Choose API and then provide access for the API

4. Name your App such as ‘OracleAgentDemo’

5. Click Create App

Step 3: Configure OAuth Redirect URL

1. Enter Redirect URL and Click Add

https://<your-oracle-domain>/api/rwdinfra/trap/2/oauth/callback

⚠️ Important:

The redirect URI must match exactly — this is the most common cause of 401: invalid_client errors.

Step 4: Configure Scope

1. Go to Permission tab from top

2. Select the correct Scope

3. Click Submit at the bottom

Step 5: Configure Oracle Cloud AI Agent Studio

A popup will display:

• Client ID

• Client Secret

1. Copy the App Key and paste it in Client ID into Agent Studio

2. Copy the App Secret and paste it in Client Secret into Agent Studio

Subscribe now

Step 6: Enable Dropbox File for Chat Experience

This step is required for file access.

1. Return to Oracle AI Agent Studio

2. Go to Agent Teams Chat Experience

3. Click third party file upload

4. Enable for Dropbox credentials

Step 7: Verify integration

1. Go back to Oracle Cloud AI Agent Studio Teams

2. Look for the attachment icon with Agent Team Chat

3. Select Dropbox

4. Grant the requested permissions when prompted

🎯 Done!

Your AI agent can now securely access Dropbox files directly from the chat interface.

Final result

You’ve successfully:

• Set up OAuth credentials in Dropbox

• Configured redirect URIs correctly

• Enabled scopes within Dropbox

• Connected Dropbox to Oracle Cloud AI Agent Studio

⏱️ Total time: ~5 minutes

To watch this in details, you can also check out YouTube video using link: Enable Dropbox Integration in 5 minutes

Share Arpit Paliwal – Oracle, AI/ERP

• Error 401: invalid_client

• Redirect URI mismatch

• File access failures even after successful login

What this guide covers

• Creating Google Cloud OAuth credentials

• Setting the correct redirect URI for Oracle Cloud AI Agent Studio

• Enabling the required Google APIs

• Fixing the most common authentication and access errors

⏱️ Total setup time: under 5 minutes

Subscribe now

Step 1: Identify required credentials (Client ID & Secret)

1. Open Agent Studio

2. Navigate to Chat Experience

3. Select Google Drive

4. Click Add New Account

You’ll see two required fields:

• Client ID

• Client Secret

These will be generated from Google Cloud Console.

Step 2: Create a Google Cloud project

1. Go to Google Cloud Console

2. Click New Project

3. Enter a project name (example: Demo Oracle)

4. Click Create

5. Ensure the new project is selected in the top project dropdown

Step 3: Configure OAuth consent screen

1. Open the ☰ → APIs & Services

2. Click OAuth Consent Screen

3. Select Create

Fill in the required fields:

• App name (example: Demo Oracle)

• User support email

• Developer contact email

All other fields are optional.

Click Save and Continue.

Step 4: Create OAuth client credentials

1. Go to APIs & Services → Credentials

2. Click Create Credentials

3. Select OAuth Client ID

4. Choose Web Application

5. Enter a name (example: DemoOracle)

Add Authorized Redirect URI

Paste the redirect URI provided by your Oracle Cloud environment, for example:

https://<your-oracle-domain>/api/rwdinfra/trap/2/oauth/callback

⚠️ Important:
The redirect URI must match exactly — this is the most common cause of 401: invalid_client errors.

Click Create.

Step 5: Configure Oracle Cloud AI Agent Studio

A popup will display:

• Client ID

• Client Secret

1. Copy the Client ID → paste it into Agent Studio

2. Copy the Client Secret → paste it into Agent Studio

Subscribe now

Step 6: Enable Google Drive API

This step is required for file access.

1. Return to Google Cloud Console

2. Go to Enabled APIs & Services

3. Click Enable APIs and Services

4. Search for Google Drive API

5. Select it and click Enable

Step 7: Publish the OAuth within Cloud Console

This step is required for to not receive client id issue.

1. Return to Google Cloud Console and then go to Audience and Publish

Step 7: Verify integration

1. Go back to Oracle Cloud AI Agent Studio

2. Refresh the page

3. Look for the attachment icon with Agent Team Chat

4. Select Google Drive

5. Grant the requested permissions when prompted

🎯 Done!

Your AI agent can now securely access Google Drive files directly from the chat interface.

Final result

You’ve successfully:

• Set up OAuth credentials in Google Cloud

• Configured redirect URIs correctly

• Enabled Google Drive API

• Connected Google Drive to Oracle Cloud AI Agent Studio

⏱️ Total time: ~5 minutes

To watch this in details, you can also check out YouTube video using link: Enable Google Drive Integration in Under 5 minutes

Share Arpit Paliwal – Oracle, AI/ERP

You type in a question, and boom—you get a detailed, coherent answer. You ask for a poem, and it just writes one. It feels like magic. But here’s the thing: it’s not magic. It’s elegant engineering built on a surprisingly simple core principle with Large Language Models (LLMs).

Let me pull back the curtain.

The Big Secret: It’s All Just Word Prediction

At their heart, large language models (LLMs) are incredibly sophisticated guessing machines. Their one and only job? Predict the most likely next word in a sentence.

That’s it. Seriously.

No consciousness. No understanding. No “thinking” in the way humans do. Just statistical pattern recognition at an enormous scale.

But here’s where it gets genius: every task gets turned into a game of word prediction.

Example 1: Sentiment Analysis

You don’t ask the AI: “Is this sentence positive or negative?”

Instead, you frame it like a fill-in-the-blank:

The movie was absolutely amazing. This review is _____.

The model has read billions of movie reviews during training. So when it sees that sentence, it thinks: “What’s the most statistically likely word to come next?”

Answer: “positive”

It’s not feeling anything. It’s just predicting.

Example 2: Question Answering

Same trick for Q&A. You set it up in question-answer format:

Q: Who wrote "On the Origin of Species"?
A: Charles _____

The model has essentially read the entire internet, including Wikipedia. So it predicts: “Darwin”

Then it looks at the whole sentence again and predicts the next word: “wrote”... and so on.

It’s just a chain reaction of predictions, one word at a time.

Subscribe now

How Does It Get So Good At This?

The answer is pre-training, and the key ingredient is a mind-boggling amount of text.

The model is forced to read a huge chunk of the internet—web pages, books, scientific articles, even computer code from Stack overflow, GitHub. Its only job, billions and billions of times, is to guess the next word.

• Every time it gets it right → it reinforces that pattern

• Every time it’s wrong → it needs to adjusts

Think of it as the world’s most intense flashcard session, except instead of a deck of cards, it’s processing hundreds of billions of words.

It’s a giant slice of all the text humanity has ever put online. And the model has “read” all of it.

But Wait—If It Just Picks the Most Likely Word, Wouldn’t Responses Be Boring?

If the model always picked the #1 most probable word, every response would be repetitive and predictable. That’s where sampling comes in.

Think of it like rolling a weighted die. Words with higher probability are like bigger sides of the die—more likely to come up. But there’s still a chance it lands on something less likely, more interesting.

The problem? If you let it roll that die forever, eventually it’ll land on a really weird word that sends the whole response into nonsense.

So we need control. We need knobs to turn.

The Three Knobs That Control AI Creativity

Whether you’re using ChatGPT, building workflow agents with LLM in Oracle AI Agent Studio, you’ll encounter three critical parameters:

1. Temperature: The Risk Tolerance Knob

Temperature controls randomness.

• Low temperature (0.1-0.3): Conservative. Picks the most obvious words. Focused, factual, but maybe boring.

• High temperature (1.5+): Creative. Takes risks on unusual words. Diverse and imaginative, but prone to hallucinations.

2. Top-P (Nucleus Sampling): The Menu Filter

Top-P is a probability threshold.

The model sorts all possible next words by probability, then draws a line. It says: “We’re only keeping the smallest group of words whose probabilities add up to P%.”

Everything outside that group? Gone.

• Low Top-P (0.5): Small, safe menu. Very predictable.

• High Top-P (0.95): Much bigger menu. More variety and surprise.

Pro tip: Top-P is often more effective than temperature for controlling creativity without sacrificing coherence.

3. Top-K: The Hard Limit

Top-K is the blunt instrument. It asks: “What are the top K most likely words?”

If K = 10, the model throws away every option except the top 10. Nothing else even gets considered.

• Low Top-K (5-10): Extremely disciplined. Safe, focused, but sometimes repetitive.

• High Top-K (50-100): Opens things up, but with a hard ceiling.

Think of it this way:

• Temperature → Changes how risky the model feels

• Top-P → Decides how wide the menu is

• Top-K → Locks the door

Seeing It In Action: The Playground

To really understand these parameters, you need to experiment. I have built a live playground (link for code in the references).

The Bottom Line

Large language models aren’t magic. They’re probability engines trained on massive amounts of text, predicting one word at a time.

But with the right understanding of Temperature, Top-P, and Top-K, you can control exactly how they behave—balancing creativity and precision to get the results you need.

Want to experiment yourself? Check out the playground in the video description and start tuning those knobs.

Reference Links

Follow me on LinkedIn: https://www.linkedin.com/in/arpitpaliwal11/

Follow me on Youtube: https://www.youtube.com/@LearnWithArpitOracleERPAI

For Playground code: https://github.com/arpitpaliwl/Oracle-Cloud/tree/main/LLM%20Playground

#LLM #AI #MachineLearning #Oracle #ChatGPT #Claude #ArtificialIntelligence #TechExplained

If you only remember three things from this post:

1. FAW is legacy and got replaced with FDI in June 2024

2. FDI has evolved into Oracle Fusion AI Data Platform, extending the same trusted foundation with deeper AI, agents, and automation.—Oracle just extended it in December 2025

3. If you’re on FDI, you are already on the foundation of Fusion AI Data Platform and will benefit from its AI enhancements as they roll out.

Still confused? Keep reading. I’ll break it all down.

Why This Matters

I kept seeing people (including myself) get confused about these three platforms. Are they different products? Do they replace each other? Which one should i use?

After diving deep into Oracle’s documentation and their recent announcements, here’s the clarity we all need.

The Evolution Timeline

Understanding the history helps everything make sense:

📊 ~Late 2019 - June 2024: FAW (Fusion Analytics Warehouse) Era

↓

🧠 June 2024 - December 2025: FDI (Fusion Data Intelligence) → Rebranded from FAW

↓

🤖 December 2025 - Present: Oracle Fusion AI Data Platform→ Evolution of FDI with AI agents, automation, and AI-first architecture

The Two Rebranding Explained

First Rebrand: June 2024

FAW (Fusion Analytics Warehouse) → FDI (Fusion Data Intelligence)

Oracle announced on June 5, 2024 that “Oracle Fusion Analytics Warehouse is now Oracle Fusion Data Intelligence.”

What changed:

• Name only - zero impact on services or functionality

• No action required from customers

• Better reflected the platform’s advanced analytics and intelligence capabilities

• All existing setups and data continued operating seamlessly

Oracle’s official statement: “This is a rebranding initiative, and there’ll be no impact on your current services. All functionalities and features you rely on will remain unchanged.”

Evolution: December 2025

FDI (Fusion Data Intelligence) → Oracle Fusion AI Data Platform

Oracle announced on December 4, 2025 that “Oracle Fusion Data Intelligence (FDI) is now Oracle Fusion AI Data Platform.”

What changed:

• Significant AI capability enhancements

• Strategic shift from “data intelligence” to “AI-powered action”

• Introduction of no-code AI agent creation for business users

• Built explicitly on Oracle AI Data Platform foundation

• Enhanced automation and conversational AI features

Oracle’s official statement: “This is more than a new name—it signals our commitment to deliver AI where it creates the most value: on your trusted data, directly inside the business processes your teams use every day.”

The Complete Comparison

Here's everything side-by-side so you can see exactly what's different (and what's the same):

📊 Platform Basics

🎯 Purpose

⚙️ Tech Stack & Architecture

🤖 AI & Intelligence

📈 Content & Features

🔒 Governance & Future

Oracle Fusion AI Data Platform is not a replacement customers must migrate to. It is the evolution of Fusion Data Intelligence, built on the same trusted foundation, now enhanced with AI agents, automation, conversational experiences, and support for structured and unstructured data. Customers using FDI are already positioned to benefit from these capabilities as they are progressively enabled.

What’s Next

Oracle Fusion AI Data Platform represents Oracle’s commitment to making sophisticated, trustworthy AI simple and useful for every team, every day. In the coming months, expect:

• Enhanced AI-powered features that will deepen automation and intelligence

• Expanded agent-building capabilities for more use cases

• New personalization options tailored to your role and workflows

• Continued focus on security, privacy, data governance, and business value

Resources

📖 Official Oracle Blogs:

• Introducing Oracle Fusion AI Data Platform

• FAW is now FDI

📖 Follow along with me on LinkedIn and YouTube below:

• https://www.linkedin.com/in/arpitpaliwal11/

• https://www.youtube.com/@LearnWithArpitOracleERPAI

#Oracle #FusionApplications #AI #DataAnalytics #EnterpriseAI #DigitalTransformation #OracleAI

Subscribe now

With Oracle Fusion AI Agent 26A, file uploads are now seamlessly integrated into your AI agent interactions. Whether it’s an invoice you need analyzed, a contract you need reviewed, or images you need processed, you can simply attach and continue your conversation.

Two Powerful Upload Options

Oracle gives you flexibility with two upload methods:

1. Local File Upload: Upload files directly from your computer’s storage. Perfect for quick document sharing and ad-hoc processing needs.

2. Third-Party Cloud Storage Integration: Connect your Google Drive, Microsoft OneDrive, or Dropbox directly to Oracle Cloud. Access your cloud files without leaving the conversation.

Real-World Use Case: Document Processing

I recently built a “Document Assistant” agent to test this functionality. Here’s what it can do:

• Accept file uploads during live conversations

• Analyze and extract information from PDFs, images, and other documents

• Provide structured insights based on uploaded content

• Process files from both local storage and cloud drives

In my testing with an invoice PDF, the agent successfully extracted header details and line items, organizing the information logically and accurately—all within seconds of uploading. Watch the vedio from the link below.

How to Set It Up

Setting up this functionality involves a few key steps:

Configure Cloud Connections If you want to use cloud storage, you’ll need to set up connections to your chosen providers. Each platform (Google Drive, OneDrive, Dropbox) has its own configuration process.

Configure Your Agent Team and Enable Upload Capabilities In the Chat Experience tab, enable either or both local file upload and third-party cloud storage options based on your needs.

Define Your Agent’s Role: Set up a clear persona and prompt that tells your agent how to handle file uploads. Be specific about what the agent should do when files are shared.

Add the Run Time Processing Tool: Oracle provides a seeded runtime processing tool that handles attachment processing. This is the engine that makes everything work.

Key Benefits

• Seamless Workflow: No more context switching or conversation interruptions

• Flexibility: Choose between local and cloud storage based on your needs

• Real-Time Processing: Get instant insights from uploaded documents

• Accurate Extraction: The processing tool reliably extracts structured information from various file types

• Enhanced Context: Files become part of the conversation context, making agent responses more relevant

Best Practices

Based on my testing, here are some recommendations:

• Be explicit in your agent’s prompt about handling files to prevent hallucinations

• Use clear descriptions so users understand what your agent can process

• Test both upload methods to ensure smooth functionality

• Set up cloud connections in advance if you plan to use that feature

This feature represents a significant step forward in making AI agents more practical and user-friendly. By eliminating the friction of file sharing, Oracle has made it possible to have truly continuous, productive conversations with AI agents.

Watch the Full Demo

I’ve created a complete walkthrough video showing exactly how to set up and test this functionality. You’ll see the agent in action processing both local files and cloud-stored documents.

Youtube: How to Upload Files to Oracle AI Agents in Real-Time | Oracle AI Agent 26A Demo

Have you tried the new file upload feature? What use cases are you exploring? Share your experiences in the comments below.

Subscribe to stay updated on Oracle Fusion Cloud features and best practices.

Subscribe now

So what exactly is the difference — and more importantly, which one should you actually use when building real-world integrations, AI agents, or data pipelines across ERP, SCM, HCM, and Financials?

Let’s clarify what Oracle has really built and how the new “Other Data Source Application” resource introduced in Release 26A fits into this model

Two API Architectures, One SaaS Platform

Oracle Fusion today exposes two API layers:

• Monolith REST (SaaS APIs): The original Fusion REST layer

• BOSS (Business Object Spectra Service)A newer, OCI-native business object abstraction

The difference is not cosmetic — it is architectural.

Monolithic (SaaS) REST APIs: The original Fusion REST layer

• Paths like /fscmRestApi/resources/..,/hcmRestApi/resources/..

• Built directly into the core Fusion application

• Versioned per product family (Financials, SCM, HCM)

• Power almost all production integrations today

BOSS (Business Object Spectra Service): The newer, OCI-native layer

• Paths containing /api/boss/

• A cloud-native data service sitting on top of Fusion

• Uniform resource model across products

• Powers modern Oracle experiences (Redwood UI, APEX apps, Data Integration Services, AI related services). These are nothing but custom or extended objects

Think of it this way:

• Monolithic APIs = “Execute transactions in the SAAS”

• BOSS = “Query and extract business data intelligently”

Why This Matters for AI, APEX, and Agents

When you build:

• AI Agents

• APEX applications

• OCI-native apps

• Intelligent dashboards

You need:

• Unstructured data extraction

• ERP Core objects

• Data Integration Services

• Filtered Data Extraction

BOSS supports that natively. Monolith APIs do not.

That is why APEX explicitly distinguishes:

• BOSS REST Data Source

• SaaS (Monolith) REST Data Source

Same authentication. Different backends.

So… Can You Use BOSS for Financials?

This is where confusion happens.

Today:

• Financials, SCM, and most ERP objects directly do not expose BOSS endpoints

• They only expose /fscmRestApi/resources/...

Which means:

For Invoices, Payments, Suppliers, Assets, GL transactions etc — must use Monolith REST APIs.

There is no BOSS alternative for these today.

So in AI Agent Studio:

• Financials objects = Monolith resource

• HCM/Common objects = sometimes BOSS

• ERP Core objects/OCI services and deep filters: BOSS resource

Release 26A: Other Data Source Application

Oracle 26A introduces a new resource type: Other Data Source Application. This allows OAuth-2.0 + JWT secured API which require IDM setup and then creds need to be updated.

What's your experience with BOSS vs monolithic APIs? Have you run into integration challenges when trying to use one vs the other? Drop a comment below.

Subscribe now

Do you:

• Tab over to documentation (and lose your flow)?

• Search oracle communities (and hope someone asked the same question)?

• Ping a colleague (who’s probably just as busy)?

Oracle just eliminated all three.

Introducing “Ask Oracle” — Your AI Assistant Inside Agent Studio

With the 26A release, Oracle AI Agent Studio now includes Ask Oracle, a conversational AI agent embedded directly into your workspace.

It’s not just a chatbot. It’s a assistant that understands the context of what you’re building and can answer questions ranging from:

• Basic: “What is an AI Agent Team?”

• Technical: “How do I use Vector DB in a Workflow Agent?”

• Specific: “Can I control chunk size in the RAG document tool?”

And it responds in natural language—with step-by-step guidance, examples, and best practices.

How It Works

1. You’re Building → You Have a Question

Whether you’re creating agents, agent teams, tools, topics, deep links, or business objects—the Ask Oracle button appears right in your window and guided journey banner.

2. Click “Ask Oracle” → Ask Your Question

A panel opens immediately. No new tab. No leaving your workspace. Just type your question in plain English.

3. Get Instant, AI-Powered Answers

The assistant searches Oracle’s documentation, synthesizes the information, and delivers:

• Clear explanations

• Step-by-step instructions

• Code samples and parameter details

• Relevant links for deeper exploration

Subscribe now

Real Examples from the Demo

I tested Ask Oracle with three questions while building an agent:

Q1: “What is an AI Agent Team?”

Response: A detailed explanation of agent teams as deployable collections of agents, tools, and topics that collaborate toward shared objectives—complete with key attributes, use cases, and practical examples.

Q2: “How to add Deep Links to an Agent?”

Response: A user-friendly, step-by-step guide including:

• Navigation instructions

• Required parameters

• Sample URL templates

• Security and access control best practices

Q3: “How can I use Vector DB in a Workflow Agent?”

Response: Technical breakdown covering:

• What Vector DB is

• How to provision it

• Integration patterns for workflow agents

• Testing and monitoring checklists

Each response felt like having a knowledgeable colleague sitting beside you.

Important Setup Note

To ensure Ask Oracle works optimally, make sure these two scheduled processes run daily:

1. Index AI Agent Studio Assistant Documents

2. Index AI Agent Studio Assistant Objects and Attributes

Without these, you might see “No information found” responses. Once you run them, log out and back in to see the full functionality.

Watch the Full Demo

I’ve recorded a complete walkthrough showing Ask Oracle in action, including:

• Where it appears across different screens

• Live Q&A examples

• Troubleshooting tips

• Setup requirements

[Watch on YouTube →]

Try it out. Ask Oracle a question. See how it changes your workflow.

The challenge? Each country has its own format, platform, timeline, and technical requirements. Germany wants XRechnung or ZUGFeRD. Italy demands FatturaPA through their SDI system. India requires JSON format with IRN validation. Saudi Arabia mandates Phase 2 integration with ZATCA including QR codes.

For multinational organizations, this creates a compliance nightmare. Or does it?

In this post, I’ll break down the global e-invoicing landscape and explain how Oracle Cloud’s OpenPEPPOL model or Oracle Cloud’s Access Point model offers a approach to managing this complexity.

The Global E-Invoicing Landscape: A Regional Breakdown

Europe: The Rapid Acceleration

• Europe is experiencing an e-invoicing revolution driven by the EU’s VAT in the Digital Age (ViDA) initiative, which will mandate e-invoicing for cross-border B2B transactions progressively from 2028-2030. But individual countries aren’t waiting.

• Germany just made B2B e-invoicing mandatory in January 2026 , requiring businesses to accept e-invoices immediately (XRechnung or ZUGFeRD 2.0.1+ formats), with sending requirements phasing in progressively. Germany uses a peer-to-peer model—no central clearance system.

• France is rolling out mandatory B2B e-invoicing, with large companies going first, followed by mid-size and SMEs (2027). They’re using the Chorus Pro platform and Factur-X hybrid PDF/XML format.

• Italy has been ahead of the curve since 2019, mandating B2B and B2C e-invoicing through their SDI (Sistema di Interscambio) central platform using FatturaPA XML format. Every invoice must be passed through the government system.

• Spain, Poland, Belgium, & Romania are all implementing mandatory systems, each with their own platforms: VeriFactu, KSeF, Peppol-based, and RO e-Factura respectively.

• The UK remains an outlier with no mandatory B2B e-invoicing as such, though their Making Tax Digital (MTD) initiative requires digital record-keeping.

Latin America: The Pioneers

• Brazil led the way with Nota Fiscal Eletrônica (NF-e) and require real-time authorization from SEFAZ.

• Mexico’s CFDI system (now on version 4.0) requires all invoices to be certified by SAT through a PAC (Authorized Certification Provider).

• Chile, Argentina, Colombia, & Peru all have mature mandatory e-invoicing systems with real-time validation by their respective tax authorities (SII, AFIP, DIAN, SUNAT).

The Latin American model of pre-clearance—where the government approves an invoice before it’s valid—has become the blueprint many other regions are following.

Asia Pacific: Following the Trend

• India has made e-invoicing mandatory for businesses above revenue thresholds (currently ₹5 crores turnover). The Invoice Registration Portal (IRP) system uses JSON format and returns an Invoice Reference Number (IRN) for validation.

• Malaysia is rolling out mandatory e-invoicing with large companies first, using the MyInvois portal with JSON or XML format.

• Singapore & Australia have taken a voluntary but encouraged approach, promoting the PEPPOL network. However, it’s mandatory for government suppliers.

• South Korea requires all VAT-registered businesses to use the National Tax Service e-invoicing system with real-time reporting.

Middle East & Africa: Emerging Requirements

• Saudi Arabia (KSA) has implemented one of the most stringent systems. ZATCA e-invoicing is mandatory with Phase 1 (e-invoice generation) completed in 2021 and Phase 2 (real-time integration) rolling out in 2023-2024. They require XML format (UBL 2.1 with KSA extensions) and QR codes on all invoices.

• Egypt has had mandatory e-invoicing since 2020 with progressive rollout through the Egyptian Tax Authority platform using JSON format.

• UAE has announced e-invoicing, but the implementation timeline is still pending.

North America: The Exception

• The United States and Canada remain outliers with no federal mandatory B2B e-invoicing requirements. However, government suppliers in both countries must use the Peppol network, and industry-specific requirements exist (healthcare, government contracts).

The Challenge: How Do You Manage 40+ Different Requirements?

If your company operates in multiple countries, you’re facing:

• 20+ different XML/JSON formats: FatturaPA, XRechnung, ZUGFeRD, Factur-X, UBL, CFDI, NF-e, and more

• 15+ government platforms: SDI, Chorus Pro, IRP, ZATCA, SEFAZ, SAT, and others

• Different models: Some countries use central clearance (Italy, Brazil), others use peer-to-peer (Germany), and some use Peppol networks (Belgium, Singapore)

• Varying timelines: Some are already mandatory, others are rolling out over 2025-2027

• Digital signatures: Most jurisdictions require qualified electronic signatures or digital certificates

• Archiving requirements: Typically 5-10 years of digital storage

• Continuous updates: Regulations change frequently (Mexico moved from CFDI 3.3 to 4.0, India keeps lowering revenue thresholds)

This is where Oracle Cloud’s OpenPEPPOL model becomes transformative.

The Oracle Cloud Solution Example: The OpenPEPPOL Model

What is PEPPOL?

PEPPOL (Pan-European Public Procurement On-Line) is not an API. It’s a secure e-document exchange network using Access Points (APs) and BIS XML standards. Think of it as the “postal service” for electronic business documents—a standardized infrastructure that enables different systems to communicate.

The genius of PEPPOL is interoperability. Instead of building direct integrations to dozens of government platforms, you connect to the PEPPOL network through a certified Access Point, and that Access Point handles the routing to various destinations whether it can be government authorities or the ERP systems.

How Oracle Cloud Implements This

Oracle Cloud doesn’t connect directly to the PEPPOL network. Instead, Oracle Cloud integrate via a certified PEPPOL Access Point or service provider like Avalara, GHX or E2open etc. Popular Access Point options include:

• OpenText/ Pagero/ Tungsten/ Comarch/ Basware

Here’s the architecture:

Oracle ERP Cloud ↔ Access Point ↔ Central Hub/Government Platform ↔ Receiver

The sender (your company using Oracle Cloud) communicates with a service provider (the Access Point), which connects to the central hub or government platform, which then delivers to the receiver.

The Power of 500+ Data Elements using UBL 2.1

Oracle Cloud ERP includes 500+ data elements available for inclusion in UBL messages. This is critical because PEPPOL requires 128 elements, and different countries add their own extensions.

For example, Germany mandates formats like XRechnung and ZUGFeRD 2.0.1+. XRechnung is already in XML format, making it easily integrated using UBL 2.1 or other XML formats provided by Oracle like OAG and cXML.

Oracle’s extensive data element library means you can accommodate:

• Industry-specific requirements

• Country-specific extensions

• Custom business needs

• Future regulatory additions

The Three-Tier Architecture

The Oracle Cloud e-invoicing solution operates on a clean three-tier model:

1. ERP CLOUD: Your Oracle Cloud ERP system where invoices are created and managed

2. ACCESS POINT: The certified Access Point (your choice of provider) that handles format transformation, validation, and routing

3. HUBS: Government platforms, tax authorities, and other receivers

The Bottom Line: Future-Proofing Your Compliance

E-invoicing is not going away. The trend globally is toward more countries mandating it, more stringent requirements, and deeper system integration. According to industry analysts, by 2028, over 60 countries will have mandatory e-invoicing requirements.

The question isn’t whether you need e-invoicing compliance. The question is how you’ll manage it.

This overview sets the foundation. In the coming weeks, I’ll publish in-depth implementation for each country—covering everything from Access Point integration to handle uses cases in Oracle Cloud. These guides will save you months of research. Subscribe below to get them all.

Subscribe now

The Evolution of My Invoice Automation Journey

This isn’t my first rodeo with invoice automation. Over the past few months, I’ve explored various approaches:

• n8n Workflow Automation: Built a system that processes PDFs and images to automatically create invoices in Oracle Cloud

• Oracle AI Agent Studio: Developed a workflow AI agent that handles spreadsheet-based invoice processing

Each iteration taught me something new about the challenges of enterprise automation—from handling diverse file formats to ensuring data accuracy before committing to ERP systems.

Why Google Antigravity?

Google Antigravity represents a new frontier in AI-powered application development. For this experimental use case, I wanted to test its capabilities in a real-world enterprise scenario: processing invoices from multiple file formats and integrating with Oracle Cloud’s Payables module.

The goal was simple: create a universal invoice processing interface that handles any input format while maintaining the human oversight necessary for enterprise operations.

The Solution: A Four-Step Intelligent Workflow

Here’s how the system works:

Step 1: Universal File Upload: Users can upload invoices in any format—Excel, PDF documents, or images (JPG, PNG). The system accepts whatever format vendors send, eliminating the need for manual data entry or format conversion.

Step 2: AI-Powered Data Extraction: Google Antigravity’s webpage used Gemini AI model in backend to extracts critical invoice fields.

Step 3: Human-in-the-Loop Validation: Before any data touches Oracle Cloud, the system presents extracted information for human review. Users can:

• Verify all extracted fields

• Correct any errors or omissions

• Add missing information

This validation step is crucial for enterprise environments where accuracy isn’t negotiable.

Step 4: Oracle Cloud Integration: Once validated, the system creates the invoice directly in Oracle Cloud Payables using the proper API structure.

Watch the Demo below:

Conclusion

The intersection of AI and enterprise systems is where real transformation happens. By combining Google Antigravity’s intelligent extraction with Oracle Cloud’s robust ERP capabilities, we can reimagine business processes that have remained largely unchanged for decades.

This experimental use case demonstrates that the future of enterprise automation isn’t about replacing human judgment—it’s about augmenting it with intelligent systems that handle the tedious work while humans focus on validation and decision-making.

What automation challenges are you tackling in your organization? I’d love to hear about your experiences in the comments.

Connect With Me

If you’re interested in enterprise automation, Oracle Cloud, or AI based integration, let’s connect on LinkedIn Arpit Paliwal.

Your AI agent just ran and processed bulk queries/transactions. Which ones hallucinated? What did it cost ? Which ones error out?

If you’re checking logs manually, you’ve already lost.

Why Testing AI Agents Is Different

Traditional testing assumes deterministic outputs and predetermined execution paths. AI agents break both rules:

• Same input → different outputs

• Dynamic tool selection

• “Correctness” is contextual

You can’t unit test your way out of this.

Enter METRO

Metrics, Evaluation, Tracing, Reporting, Observability

Oracle built agent governance directly into Fusion AI Agent Studio—not as a bolt-on, but as a native framework.

Four pillars:

1. Evaluation (Test Before Deployment to PROD)

• Create Q&A evaluation sets

• Run automated correctness checks using “LLM as a Judge”

• Compare runs to measure improvement

2. Monitoring (Dashboard for Production)

• ⚡ Performance: P99 Latency, Error Rate

• 💰 Cost: Token usage (Input/Output)

• 🎯 Quality: Median Correctness, Safety Flags

3. Tracing (Execution Timeline)

When things break, see exactly what happened:

• Step-by-step agent actions

• Tool invocations with inputs/outputs

• Full LLM context and prompts

• Memory state at each step

4. Metrics (Core Measurements)

Error rate, latency, token counts, correctness scores—all agent-specific, not generic CPU stats.

Why This Matters for Workflow Agents

Remember the Google DeepMind multi-agent scaling white paper and my Evaluation series (Links at the end)?

Key findings:

• Multi-agent systems use 3-5× more tokens than single agents

• Oracle’s Workflow agent orchestration beats naive scaling

• Token efficiency is the deployment constraint

METRO lets you measure coordination overhead. You can trace:

• Which agent consumed the most tokens

• Where latency bottlenecks occur

• Which tool calls are redundant

  ---

  Best Practices in 60 Seconds

  ✅ Define success criteria before deployment
  ✅ Create at least 1 scenario per tool function
  ✅ Test with 5-10 questions per scenario
  ✅ Schedule daily metric aggregation

  Oracle’s rule: Evaluation isn’t optional—it’s core to the agent lifecycle.

• To get started Run the process: Navigator > Tools > Scheduled Processes > "Aggregate AI Agent Usage and Metrics"

  ---

Example Screenshots from the different Screen within METRO for Oracle Fusion AI Agent Studio:

Monitoring Details across different time periods

Agent Tracing for the given session from the above screen

Evaluation Screen

Evaluation Run Screen

Example of Error when check from Agent Tracing

Look out for below LinkedIn Posts for more details:

Part 1: Oracle Cloud AI Agent Evaluation Part 1

Part 2: Oracle Cloud AI Agent Evaluation Part 2

Part 3: Oracle Cloud AI Agent Evaluation Part 3

Google Deep Mind White Paper: Comment on Document for Oracle

If you’ve ever worked in Oracle Cloud finance, you know the drill. An invoice arrives via email. Someone opens the PDF or may directly send to IDR. Depending upon the approach, there are still some manual process performed in the ERP system either due to IDR limitations or other issues during the process which at the end increases the backlog.

This process can take hours. Multiply that by dozens or hundreds of invoices per week, and you have a significant productivity drain. It introduces errors—duplicate invoices, wrong amounts, mismatched codes—leading to payment delays and compliance issues.

What if I told you this could be done in seconds, automatically, with near-perfect accuracy?

The Vision: Snap, Send, Done

Imagine a world where creating AP invoices in Oracle Cloud ERP is as simple as snapping a photo. You receive an invoice on your phone, snap a picture (or forward a PDF), send it to a bot, and within seconds, the invoice is validated, data extracted, and created in your ERP system—no manual entry, no spreadsheets, no scripts.

That’s exactly what I built, and I want to show you how.

The Solution: Three Technologies, One Workflow

1. n8n - Workflow automation that orchestrates everything
2. Google Gemini Flash 2.0 - AI that extracts and validates invoice data from images/PDFs
3. Oracle Fusion Cloud ERP - The destination for automated invoice creation
4. Telegram Bot - Simple interface users already know

How It Works

𝟏. User sends an image of an invoice to Telegram Messenger Bot.

𝟐. The workflow uses AI-powered image processing to extract every key detail (supplier, amount, dates, line items) and structures into JSON Format.

𝟑. This solution uses Google Gemini for intelligent extraction, n8n for orchestration, and Oracle Fusion Cloud for instant AP invoice creation.

4. Data is validated, mapped, and seamlessly sent to Oracle Fusion Cloud for automated AP invoice creation—saving time and avoiding human error.

🧠 𝐖𝐡𝐲 𝐭𝐡𝐢𝐬 𝐦𝐚𝐭𝐭𝐞𝐫𝐬:

𝐒𝐩𝐞𝐞𝐝: What once took hours is now done in seconds.
𝐀𝐜𝐜𝐮𝐫𝐚𝐜𝐲: Eliminates manual mistakes and duplicate efforts.
𝐅𝐥𝐞𝐱𝐢𝐛𝐢𝐥𝐢𝐭𝐲: While my demo shows image-based processing, the same workflow can process PDF invoices with zero code changes.

The combination of cloud ERPs, workflow automation, and generative AI is just getting started.

See It in Action [🎥 Watch the Demo Video]

Let’s Connect

Drop a comment, connect on LinkedIn, or subscribe for more posts on Oracle Cloud ERP automation and AI.

Resources Used:

• n8n: https://n8n.io

• Google Gemini API: https://ai.google.dev

• Oracle Fusion Cloud Docs

Share this with finance and technology colleagues who could benefit from AI-powered automation.

#FinanceAutomation #AI #OracleCloud #APAutomation #ERP #n8n #Googlegemini #Oraclefusion #oraclepayablescloud

The Real Problem

When an agent is just a couple of prompts + a few tools—with no structure, no guardrails, no lifecycle—it will fail in enterprise environments.

For Finance, SCM, HR, and CX functions, reliable AI agents need more than a workflow diagram. They need three things:

1. Clear Ownership

What actions will this agent perform? Which business teams will it access?

Ownership enables stakeholders to trust an agent enough to move it from testing to production.

2. Boundaries That Protect Core Systems

Before building, decide:

• Which security roles it needs

• Which actions it can take

• Which actions always require human approval

This prevents risky automation over critical ERP processes.

3. A Lifecycle, Not a One-Time Build

Treat agents like products, not projects:

Roadmap → releases → monitoring → continuous improvement

Start with testing, observe behavior, then expand scope based on evidence.

A Framework That Works

Here is a five-step approach for building governable AI agents:

The Bottom Line

The difference between a demo agent and a production agent is structure. The teams that succeed will treat agent design with the same rigor as traditional enterprise applications.

Note: This reflects what I’ve learned so far. I’d love to hear different approaches or lessons learned from others building enterprise AI agents.

#AIAgents #EnterpriseAI #OracleCloud #Automation #OraclecloudERP #DigitalTransformation #AIinFinance #ERPOracleCloud #AIAgentStudio

We already know how to work with Tools and Business Objects—they’re foundational.

But when it came to writing prompts, I initially found myself putting everything into them just to make the agent work. The role of Topics wasn’t immediately clear.

Here’s what I’ve learned.

🛠️ Prompt: Define the Agent’s Core Identity
Every agent needs a clear purpose. The prompt is where that purpose is defined.

This includes:

• Role & Persona: “You are a Payables Support Agent.”

• Scope: What it can and cannot do.

• Input/output rules: Expected data to fetch, constraints, validation logic.

• Examples: Sample interactions, tone, reasoning steps.
  This is your agent’s DNA—specific, focused, and non-negotiable.
  
  🔁 Topic: Reuse What Should Never Change
  A Topic isn’t an alternative to a prompt—it’s a force multiplier.

  Topics are best used for logic that should remain consistent across agents, such as:

• Policies: Example “Always use professional tone,” “Never share confidential data.”

• Basic Error handling: Example “If step 1 fails, retry once, then return error.”

• Formatting rules: “Return JSON without code block wrappers.” or “Format into HTML table for list of data which has more than 2 rows.”

• Conversation patterns: How to start, follow up, or close a session.

🧩 How Topics Are Applied

To use a Topic inside an agent, reference it directly in the prompt using an expression such as:
$topic.<topic_code>
This allows the agent to inherit shared behavior without bloating its core prompt.

✅ Key Insight
Prompt = Agent-specific logic
Topic = Organization-wide consistency for handling conversation scenarios

This separation has changed how I build agents.

I now spend less time debugging repeated behavior and more time improving actual capabilities.

#OracleAI #AIagents #PromptEngineering #OracleFusion #AIAgentStudio #oraclefusion #oraclaiagentstudio ]#oracleaiagenttopics