Oracle used its AI World conference in Las Vegas to make a specific bet about enterprise software: employees should stop learning which module or which agent handles a task, and simply state the business goal. The software should then pick the transactions, workflows, approvals and agents, and coordinate them.
The two products behind that pitch are Fusion Companion and Fusion Agentic Home. Companion is reachable through supported AI applications including ChatGPT, and through Agentic Home, which is where Oracle's own interface for delegating work and tracking it to completion lives. Companion is supposed to respect existing identity, permission and approval rules, and to pull in information from other enterprise systems where authorised.
The more interesting engineering is underneath. Long-running or calculation-heavy work goes to Fusion Claw, an execution environment Oracle announced on 29 September that splits the job in two: language models plan and weigh alternatives, while deterministic procedures do the heavy computation and execute defined steps. Models from OpenAI and Gemini are both in the mix. Oracle says 25 Claw-based applications are available out of a portfolio of 75 agent applications announced in September.
The governance claim rests on a single artifact. An "Outcome Receipt" is meant to record which authorities applied, what information was used, which decisions were taken and which transactions were executed, and to pull in the right humans when an approval or exception is required. Companies set the boundaries themselves, from steps that always need confirmation to autonomous execution within explicitly delegated authority.
That is a sensible design: the audit trail, not the demo, is what an IT auditor can actually check. It is also, for now, a description. Oracle has not published the receipt's schema, retention rules or how an external auditor would query it.
There are two more pieces. A public Fusion Setup and Configuration Registry collects skills, scripts and reference workflows describing configurations, dependencies and checks, so AI tools can perform setup work directly in Fusion and then test and validate the result. It is under the Universal Permissive License 1.0, and Oracle's own GitHub repository calls it separately published example and reference material that the regular Fusion product support does not cover. It requires Fusion Applications release 26D with the 16 October 2026 maintenance pack or newer. Separately, an AI Success Navigator offering adds agents for analysis, design, configuration and testing.
The customer proof offered is thin and Oracle is unusually candid about the edges of it. Lexington Health is named as an early user, with effort for certain design and configuration tasks during implementation down 80%. That figure covers those sub-tasks only; testing and go-live were still pending according to the customer report Oracle cites, and no measurement methodology is disclosed.
Pricing is consumption-based. Fusion AI uses "AI Units" at a list price of one US cent per unit. SaaS customers get a monthly quota and can buy more, and consumption depends partly on which model is used — which means the same business goal can cost different amounts depending on how Oracle routes it.
The gap is availability. Oracle calls the 25 Claw-based applications available, but says parts of Companion and Agentic Home will only become usable later. There is no rollout schedule and no list of what customers can switch on today. Companion also overlaps with a separately announced Oracle AI Workspace, which is meant to expose Fusion, NetSuite, databases and cloud services through ChatGPT; how the two fit together is not explained in the announcement.
For anyone running Fusion, the practical question is not whether an agent can start a purchase requisition. It is whether the receipt, the permission model and the approval boundary survive contact with a real month-end close — and that has not been demonstrated in public yet.
