.jpg)
At the end of every quarter, a controller sits down with thousands of accounting entries. Most will reconcile cleanly, but a handful of them won’t. Each of those exceptions requires an investigation, a judgment, and a correction inside the system of record, all with a paper trail. Delegating that work to AI raises a practical question: Which decisions can the software make, and which still require a person? The answer depends on the data AI can use, the actions it is authorized to take, and how its work can be checked.
“For decades, we built enterprise software to help people do work. We’re entering an era where software increasingly does the work and delivers business outcomes, while people define the objectives, establish the policies, and retain control,” said Chris Leone, executive vice president, Applications Development, Oracle. “Our Fusion Agentic Applications were an important first step, and Fusion Claw takes that to a whole new level.”
Work That Resists Automation
The reconciliation scenario demonstrates why a certain type of enterprise work has remained manual: It requires reasoning and precision at the same time; it lives inside the transactional system; and it needs to leave evidence behind. Most AI deployments sit alongside the enterprise system, without native access to the data, workflows, permissions, and approvals needed to act. The result is a familiar pattern in enterprise AI: plenty of pilots, but too few production deployments.
“The hard part of enterprise AI isn’t getting a model to recommend what to do,” Leone said. “It’s getting from that recommendation to actually doing the work, against authoritative business data, at enterprise scale, with the security, controls and economics an enterprise requires."
Oracle Fusion Agentic Applications were built to help close that gap. They bring together coordinated teams of specialized AI agents that operate with shared context to coordinate end-to-end business processes around outcomes. The portfolio now spans 75 Agentic Applications. One major question was how far that model could extend into the hardest work within those processes.
Enterprise AI’s Division of Labor
Through new Fusion Claw-powered agentic applications, work like accounting reconciliation, workforce staffing analysis and plans, shipping consolidation, and sales territory planning can be addressed. These 25 new Claw-powered Agentic Applications are autocapable, meaning they can execute authorized workflows autonomously within configured enterprise controls and applicable approval requirements.
Fusion Claw-powered Agentic Applications use frontier AI models to understand objectives, reason through complex problems, and develop execution plans. Fusion Claw then uses deterministic enterprise computation to execute those plans precisely across massive transactional data sets, without requiring the AI model to reason through every individual record.
“We use AI where reasoning is valuable: understanding the objective, developing the plan, and adapting it as conditions change. We use deterministic enterprise computation where precision and scale matter: reconciliation, modeling, optimization, business rules, and transaction processing," Leone said. “You can’t economically ask a frontier model to reason through a million ledger entries one at a time. The model determines how to solve the problem, and Fusion Claw can execute that plan precisely across enterprise-scale data.”
The split is intentional. It reserves AI reasoning for the parts of the job that require judgment and hands the heavy, repetitive computation to systems built for precision at scale. That’s what puts specialist-grade work, such as reconciliation across huge data sets, simulation and modeling, optimization, or continuous re-planning, within reach of the system’s capabilities.
Throughout, Oracle Fusion Applications remain the authoritative system for enterprise data, business state, controls, approvals, and transactions. Fusion Claw adds execution depth on top of that foundation rather than replacing it.
Autonomy Is Delegated Instead of Assumed
If software is going to execute work, someone must decide what it’s allowed to do. Oracle addresses this through Know Your Agent (KYA), a framework for delegated authority across Fusion AI. The fundamental design principle is that AI should not have more authority than the person or organization delegating the work, subject to applicable legal and regulatory requirements. Access to data or the ability to perform an action does not automatically grant authority to execute it.
In Oracle’s design, that authority is spelled out in advance. Fusion Claw operates inside an Enterprise Operating Envelope, the policies, permissions, risk thresholds, decision rights, and approval requirements configured by the organization. An Outcome Trust Harness is designed to apply those configured limits on each run, governing the system’s identity, capabilities, data, and actions.
Each Claw Outcome runs in an isolated execution environment with controlled access to enterprise data, capabilities, and network destinations, providing additional protection while the work is performed.
“Auto doesn’t mean unconstrained autonomy. People define what the system is authorized to do. Within that authority, AI executes, subject to applicable oversight and approval requirements. If approval is required, it seeks approval. Outside that authority, it escalates,” said Leone. “The goal is to make enterprise controls part of the architecture that enables greater autonomy.”
Every Claw Outcome also produces an Outcome Receipt, an auditable record of the objective, authority, evidence, decisions, actions, transactions, and result. “We don’t ask customers to take the system’s word for it,” Leone said. “If AI is going to move from recommending work to doing work, that evidence has to be part of the execution.”
Judgment Where It Matters Most
For business and technology leaders, the question is no longer about whether AI can help get work done—it can—but instead which work to delegate, under what authority, with what human oversight, and with what evidence. In the reconciliation example, the routine entries keep moving while the controller focuses on the exceptions, trade-offs, and decisions where judgment materially changes the outcome.
That’s the shift Fusion Claw is designed to help facilitate: More specialist-grade work can move into controlled execution, while people continue to own the outcome and the authority behind it, remaining accountable for the result and providing oversight where required.
