Oracle AI Agents: Oracle AI Agent Studio, the Fusion AI Agents List and What They Cost

Oracle's July 14, 2026 announcement says customers can extend "the 1,000-plus AI agents delivered through Fusion Applications and the 22 new Fusion Agentic Applications that were launched earlier this year." That is the number in the press release. The number you can verify is different, and if you are deciding whether a specific agent exists, the verifiable one is the one that matters.

We parsed Oracle's public Fusion AI catalog at oracle.com/applications/fusion-ai on August 23, 2026 and counted every named row. It lists 229 entries: 207 individually named AI agents embedded in Fusion, plus 22 Fusion Agentic Applications. The table above breaks that down by module. The gap between 1,000-plus and 229 is not necessarily marketing inflation, because Oracle counts agent variants and per-process agents that it does not enumerate on the public page. It does mean that if your business case rests on a particular agent, you should confirm it by name against the catalog and your release notes rather than against the headline count.

Two things stand out in the breakdown. Supply Chain Execution is by far the deepest module at 36 named agents, more than the whole of Finance and EPM and Risk combined, which tells you where Oracle put its engineering effort first. And HR Foundation at 31 agents is mostly employee self-service: a Benefits Plan Advisor, a Tax Withholding Guide, a Marital Status Update Assistant. Those are useful, high-volume and low-risk, and they are also the agents least likely to change your cost structure.

Finance, the module most buyers ask about first, carries the smallest count on the list at 11 across Financial Management, EPM and Risk. The five Financial Management agents are the substantive ones: a Payables Agent that ingests invoices from email, portals, EDI and PDFs and matches them to purchase orders and receipts; a Cash Processing Agent that creates receipts from bank statement lines and matches remittance advices; a Ledger Agent that can autocreate adjustment journals; an Expenses Agent that matches receipts to card charges and validates against policy; and a Payments Agent covering early pay, virtual cards and financing options.

Oracle's own wording, repeated verbatim in both the March 24, 2026 and July 14, 2026 announcements, is this: "Available at no additional cost, Oracle AI Agent Studio for Fusion Applications delivers easy-to-use tools, including orchestration, advanced testing, robust validation, and built-in security." There is no per-seat AI fee, no per-conversation charge, and no credit or token meter for the embedded Fusion agents. That is a genuinely different commercial posture from Salesforce, which meters Agentforce at $2.00 per customer-facing conversation or 20 credits per action, and from Microsoft, which runs Copilot Studio on a pooled credit balance.

The catch is what "no additional cost" is additional to. It is additional to a Fusion Cloud Applications subscription, and Oracle does not publish Fusion list prices at all. The figures below come from licensing advisory firms that publish rate cards, not from Oracle, and every Oracle deal is custom quoted. Treat them as the shape of the number, not the number.

Two numbers change the arithmetic more than the rate card does. Oracle Cloud ERP is typically sold with a floor around 25 users, so the smallest realistic deal is not one seat. And discounting is heavy and expected: advisers report 25% to 40% off list for organizations in the 200 to 500 user range and 40% to 55% for 1,000 users and up, driven by deal size, timing and whether SAP or Workday is in the room.

Put those together and the honest summary is that Oracle AI agents are free the way airline seat selection is free once you have bought the ticket. For a company already running Fusion, the marginal cost of switching on the Payables Agent really is zero, and not switching it on is leaving paid-for capability idle. For a company that does not run Fusion, "the agents are included" describes a six-figure annual commitment before a single agent runs.

Oracle AI Agent Studio for Fusion Applications is the development platform Oracle uses to build its own Fusion agents, made available to customers and partners. It is included with a Fusion Cloud subscription at no additional license charge. Oracle's positioning is that you extend the prepackaged agents rather than start from a blank prompt, and that anything you build inherits Fusion security, approvals and audit trails instead of bolting them on afterwards.

The capabilities Oracle lists for AI Agent Studio, taken from its own capability table:

Two items on that list are worth pausing on. Interoperability through MCP and A2A is Oracle committing to the same open agent protocols the rest of the market is standardizing on, which is what makes it realistic to have a third-party agent talk to a Fusion agent rather than replace it. And the agent ROI dashboard is a quietly aggressive product decision: Oracle is offering to measure, per agent, the time and cost it saved. Vendors do not usually ship the instrument that can prove their product did nothing.

On the developer side, Oracle's July 2026 update added an AI Studio Skill so that agentic applications can be built with Visual Studio Code, standard CLIs, Git-based workflows and AI coding assistants including OpenAI Codex and Claude Code, backed by a public GitHub repository of templates and reference architectures. If your team already builds this way, that removes the usual complaint about enterprise low-code studios.

Oracle runs its own agent marketplace, and it sits inside AI Agent Studio rather than as a separate storefront. Oracle describes it as "trusted, validated agent templates from a global network of certified partners deployed within the Oracle Fusion AI Agent Studio," and the sales pitch is deployment "in minutes instead of months."

What makes it different from a general app store is the validation process Oracle documents. Partner templates are "reviewed to meet the same standards as the templates built and provided by Oracle development," specifically for security, functionality, performance, human oversight, feedback and deployment, and they are then supported by the Fusion Applications support team rather than only by the partner. That is a meaningfully higher bar than most agent directories apply, and it is also the reason the catalog is smaller than an open marketplace would be.

Oracle's July 2026 announcement said the marketplace "is expanding to support a catalog of agentic applications in addition to the existing portfolio of AI agents," and that the partner bench behind it had grown to more than 80,000 certified experts trained in AI Agent Studio, up from 63,000-plus reported in March 2026. That is roughly 27% growth in the certified population in under four months, which is a reasonable proxy for how hard Oracle is pushing the ecosystem.

The constraint is the obvious one: everything in the Oracle AI Agent Marketplace runs inside Fusion, for Fusion customers. It is an excellent answer if your systems of record are Oracle. It is not an answer if your finance data lives in NetSuite or QuickBooks and your support desk is Zendesk, which is the case for most mid-market companies. A broader AI agent marketplace that connects across vendors covers that case instead.

Oracle now sells two things under the agent banner and uses precise language for each, so it is worth getting right before you sit down with a rep.

An AI agent is a single embedded capability inside a Fusion module. The Ledger Agent answers a natural language question about a balance and can draft an adjustment journal. Scope is one job.

A Fusion Agentic Application is, in Oracle's words, "a new class of enterprise application: outcome-driven systems backed by teams of specialized AI agents that reason, coordinate, and decide, then execute work through Fusion business objects, workflows, tools, policies, approvals, and logged actions." Scope is a business outcome, and several agents work under it. Oracle launched 22 of these in March 2026: 12 for ERP, 7 for HCM and 3 for CX.

The 22 are named after the outcome rather than the module, which makes them easy to identify. On the ERP side: Batch Process Manufacturing Workspace, Collectors Workspace, Cost Accounting Close Workspace, Design to Source Workspace, Logistics Execution Command Center, Maintenance Operations Workspace, Production Shift Operations Workspace, Product Readiness Workspace, Sales Order Command Center, Security Command Center, Sourcing Command Center and Warehouse Operations Workspace. On the HCM side: Career Advancement Command Center, Hiring Workspace for Store Managers, Manager Concierge Workspace, My Help Workspace for Employees, Team Learning and Development Workspace for Managers, Team Talent Calibration and Review Workspace and Workforce Operations Command Center. On CX: Contract Compliance Workspace, Cross-Sell Program Workspace and Sales Command Center.

The practical test when a vendor or partner says "Oracle has an agent for that": ask whether they mean a named embedded agent that ships in your module today, or an agentic application, or something a partner would build for you in AI Agent Studio. All three are real, and they carry very different implementation timelines.

Oracle's pricing model is the outlier here, and in the buyer's favor if you are already an Oracle shop. Nobody else in this comparison gives you the agents and the builder with no incremental meter.

The honest reading of that table is that the comparison is decided before you get to it. If your systems of record are Fusion, Oracle's model is the cheapest of the four at the margin and the argument is over. If they are SAP, the same is true of Joule. Cross-suite agent shopping only becomes a real decision for the processes that sit between systems, and for companies that do not run a full suite at all.

Where the models genuinely diverge is risk. Salesforce charges per conversation, so a support deflection program that works costs more as it scales, which is defensible but needs modelling. Microsoft's credits expire monthly and Microsoft disables custom agents at 125% of prepaid capacity. Oracle's flat inclusion means an agent that gets ten times busier costs the same, which is the easiest of the four to budget and the reason enablement, not procurement, is the bottleneck on the Oracle side.

If you are running the same evaluation on the SAP side, our SAP Joule agents page covers the Joule Agents and Joule Assistants split and the AI Unit rate card, and Agentforce pricing works through the Flex Credits versus Conversations break-even.

This trips up more evaluations than anything else on the page. Oracle sells agents from two different parts of the company, they are priced on completely different models, and both are called Oracle AI agents in conversation.

The reason this matters commercially: when Microsoft documented the six outside platforms that sync into the Agent 365 registry, the Oracle entry is Oracle Generative AI Agents, the OCI service, not the Fusion agents. If your Microsoft admin tells you Oracle agents are already visible in your registry, they are talking about the OCI product. Our Microsoft Agent 365 page has the connection requirements, including the least-privilege IAM policy Oracle recommends for that sync.

Every embedded agent inherits the boundary of the suite it lives in. Oracle is explicit that Fusion Agentic Applications "run natively in Oracle Fusion Applications, inherit Fusion security and governance controls, act against Fusion business objects and workflows." That is the whole design, and it is why the governance story is strong. It is also the limit.

Three gaps come up repeatedly in real evaluations:

• Processes that end outside Fusion. The Payables Agent handles the invoice beautifully once it is in Fusion. If your suppliers email PDFs to a shared inbox that nobody has connected, the agent never sees them. Document capture into the system of record is a separate problem, and pulling structured line items out of supplier invoices before they hit the ledger is usually where the manual hours actually are.
• Companies that run part of the stack elsewhere. Fusion for HR and NetSuite for finance is a common shape, and it is exactly the shape embedded agents handle worst, because neither suite's agents can see the other's data. Our AI agents for NetSuite page covers the finance half of that split.
• Teams below the Fusion threshold. A 40-person distributor is not buying a 25-seat Fusion ERP contract to get a stockout advisor. Connected agents that reach the tools they already run, on a flat subscription, are the realistic option.

None of that is an argument against Oracle's agents. If you hold the license, use them, and use AI Agent Studio to extend them rather than building parallel automation outside the audit trail. The argument is only against treating "Oracle has 1,000 agents" as a complete answer to a process that crosses the suite boundary. For the crossings, a marketplace agent that connects through APIs with scoped tokens, with a person approving anything financial, closes the gap without a second ERP contract.