Oracle has opened a new route into its Fusion Cloud business software: employees will be able to start work from supported AI assistants such as ChatGPT, while a new set of 25 agentic applications handles longer-running tasks across finance, human resources, supply chain, sales and service.
The company announced Fusion Companion, Fusion Agentic Home and the expanded lineup of Fusion Claw applications on Monday at Oracle AI World in Las Vegas. Oracle describes Companion as an intelligent layer that translates a business objective into the Fusion transactions, workflows, approvals and specialized agents needed to carry it out. The 25 Claw-powered applications are available now across Fusion ERP, HCM, SCM and CX, according to Oracle’s announcement.
The practical change is that a user does not necessarily have to begin inside the correct Oracle module or know which transaction to run. A request made through a supported client can pass to Fusion Companion, which applies the user’s business context and permissions, selects the relevant workflow and routes consequential steps through the organization’s existing controls.
How a ChatGPT request reaches an Oracle workflow
Oracle’s model has three distinct layers. The outside assistant is the conversational starting point. Fusion Companion interprets the objective and coordinates the required Oracle components. The underlying Fusion applications remain the system of record and enforce the user’s access, approvals and business rules.
For example, a finance manager could ask an approved assistant to investigate why the monthly close is behind. Companion could route that objective to the Ledger agentic application, which reconciles entries, investigates exceptions and applies deterministic calculations. Proposed actions still have to remain within the authority and approval path configured by the customer.
Oracle says Companion can also use context from other enterprise systems and coordinate authorized actions across them. That is potentially more consequential than a new chat interface: it puts the assistant in front of a multi-system workflow. The announcement names ChatGPT as an example of a supported AI tool, but it does not provide a complete client list or explain in the main release how each outside client will be connected, licensed or administered.
Fusion Agentic Home provides Oracle’s own front end for the same model. Users can define an objective, delegate it to agents and follow its progress to completion. That gives employees and supervisors a place to inspect active work rather than leaving a long-running task buried in a chat thread.
What the 25 Fusion Claw applications actually do
Fusion Claw is the execution runtime beneath the new applications. Oracle introduced the runtime in September and says it supports long-duration work, continuous replanning and a mix of model reasoning and deterministic computation. The latter matters in payroll, accounting, capacity planning and other jobs where arithmetic and policy checks cannot be left to a language model’s probabilistic output.
The newly available applications are narrower than a general-purpose autonomous worker. Each targets a bounded process and operates against Fusion data and controls. Among the examples Oracle disclosed:
- Ledger analyzes accounting entries, reconciles information, investigates anomalies and prepares approved financial actions.
- Payroll Activity Workspace compares pay, balances and costs with previous periods, then flags material variances for follow-up.
- Workforce Orchestration evaluates redeployment, hiring, development and external-talent options against demand and cost.
- Supply Planning Allocation compares ways to distribute scarce materials and production capacity across competing demand.
- Procurement Spend looks for off-contract purchasing and contract leakage, estimates savings and recommends corrective action.
- CPQ Test-to-Production Setup & Configuration Migration assembles proposed configuration changes and dependencies, flags deployment risks and executes approved migration steps.
- Service Health Check identifies emerging service problems and can prepare changes to knowledge, routing, automation or policy.
- Strategic Field Demand Planning predicts resource needs and coverage gaps while weighing overtime costs.
The supply-chain applications, for example, are designed to evaluate constraints and alternatives before carrying out an authorized plan, rather than merely summarizing a dashboard. Oracle’s supply-chain release also names shipment-load planning, which weighs consolidation, capacity, service commitments and cost before taking an approved action.
Governed execution still needs a precise boundary
Oracle repeatedly describes the applications as governed, but buyers should translate that label into testable controls. A user permission check at the start of a task is not enough for an agent that can replan over hours, call other systems and eventually execute a transaction.
Teams evaluating the features should identify which steps are read-only, which generate recommendations and which can change production records. They should verify whether approval is required at the final transaction or at every material branch, how delegated authority expires, and what happens when underlying data changes after an approval but before execution.
The audit trail also needs to preserve more than a final result. Investigators and business owners will need the original objective, data sources consulted, agent and tool calls, policy decisions, human approvals, retries, configuration versions and the exact production changes made. Oracle says AI Agent Studio includes observability, safety controls and return-on-investment measurement, but customers should confirm the event detail and retention available for their specific workflows.
Rollback is another practical test. Reversing a journal entry, a recruiting communication, a supplier action and a CPQ deployment are different operations. A useful pilot should document the compensation or recovery path for each tool an agent can invoke before expanding its authority.
What Fusion customers should test first
A narrow, high-volume exception process is a better starting point than an open-ended goal spanning several departments. Payroll variance review, contract-leakage detection or close-readiness analysis can be measured against an existing baseline while keeping execution behind human approval.
Before enabling an outside AI client, administrators should test identity propagation end to end. The person authenticated in ChatGPT or another client must arrive in Fusion with the correct individual context; the connection should not collapse many users into an overly privileged integration account. Cross-system context should be minimized to the fields required for the task, with sensitive outputs governed by the destination client’s retention and access policies as well as Oracle’s.
A pilot scorecard should track false escalations, missed exceptions, time to approval, transaction reversals and the number of cases that still require manual reconstruction. Those figures will show whether an agentic application is removing work or simply moving review effort to a new interface.
Oracle’s launch puts a concrete execution layer behind the familiar promise of talking to enterprise software. Its value will depend less on how naturally employees can phrase a request than on whether each customer can keep identity, approval, evidence and recovery intact as the request turns into action.
