Oracle gives Fusion agents long-running planning jobs that can change the system of record

Oracle's own example starts with a sick call. An employee phones in, and a Fusion Claw agent reopens the staffing plan on its own ^[6]. It weighs qualifications, availability, labor rules and cost targets, runs the calculations and, if authorized, writes the new schedule into the system of record ^[6]. The manager who used to rebuild that rota by hand now has one decision left: whether to see the plan before it goes live.

The pitch is optimization. "We did not previously have solutions for advanced optimization," said Natalia Rachelson, Oracle's senior vice president of Fusion Applications and product management ^[5]. She said Oracle's earlier agents pursued focused goals such as improving renewals, while Claw is meant for long-running work where the plan changes repeatedly ^[4]. The 25 Claw apps join 50 existing Fusion agentic applications ^[3][1].

What Claw actually does is narrower than the pitch, and easier to test. A language model reasons and plans, and conventional software does the calculations and processes the transactions ^[2]. Claw runs in an isolated environment that keeps the model from updating Fusion business objects directly, Rachelson said ^[12]. For a controller, that split decides what can be checked. The totals in a reconciliation come from ordinary code ^[7]. The choice of which exception to chase, and which fix to propose, comes from a model. In the first release that model is from Gemini or OpenAI ^[8].

Oracle expects comfort to grow with use. The company expects customers to authorize more work without individual approvals over time ^[14]. "Honestly, any task can be in full auto mode," Rachelson said ^[15], adding that she expects organizations to start with more human review ^[16]. Review mode assumes someone reads the plan. A queue that staff approve without opening it gives the same control as delegated execution, plus a click. The evidence for Oracle's gradual path is thin: Oracle did not provide measured cost savings or evidence of how the apps perform in production, according to SiliconANGLE ^[17].

The meter follows the work. Customers pay for AI units as work completes ^[18]. An agent that reopens a plan on every absence completes more work than one that plans once a week, and I'd expect the bill to rise with how often plans reopen.

Oracle's 25 apps include tools for accounting exceptions, staffing plans, shipment consolidation and sales territory planning ^[9]. To set a mode for each, sort them on two axes. One is whether the outcome turns mostly on calculation, the part Oracle keeps outside the model, or on the model's judgment about what to try ^[7]. The other is whether a wrong write can be undone before anyone is paid or billed. The same app can land in different boxes at two companies, depending on labor contracts and the close calendar.

Calculation-heavy and reversible: delegated execution, with a weekly sample of Outcome Receipts. Judgment-heavy and reversible: review mode until the receipts show the model choosing what your planners would choose. Calculation-heavy and irreversible: review mode with quick approvals, since a reviewer can check the numbers. Judgment-heavy and irreversible: review mode until production numbers exist, from Oracle or from your own pilot.