Oracle Fusion Claw pinches AI costs, tightens grip on policies

Oracle is adding a new execution layer to its Fusion Cloud Applications Suite that it says will help enterprises lower the costs of automating complex and long-running business processes with AI agents.

The execution runtime, Fusion Claw, uses a combination of a frontier model, deterministic policies, and governed execution capabilities to plan and carry out a given task or process, instead of relying for each step on AI inference, which can result in repeated model calls as the task progresses. That can help enterprises keep inference and compute costs under control, the company said.

The deterministic policies can include approved algorithms, business rules, calculations, optimization services, reconciliation logic, and transaction-processing capabilities. Enterprises can define or configure them via Fusion’s AI Agent Studio while building the agent or modifying existing agentic applications, Oracle said.

Claw then uses two built-in features, the Enterprise Operating Envelope and the Enterprise Trust Harness, to enforce those policies as controls during task execution, ensuring that the agent operates within its delegated authority and the required governance boundaries, while also generating an auditable record of agent actions and decisions, it further said.

Fusion Claw could help improve an agent’s unit economics

For Manoj Chandra Jha, principal analyst at Nord-IQ research, Fusion Claw’s approach to agentic execution is likely to help enterprises with the unit economics of deploying AI agents to automate business processes.

“By keeping the expensive model for planning and running high-volume work on cheap deterministic computation, Fusion Claw can make agentic costs both lower and more predictable for CIOs,” Jha said.

Beyond cost, Fusion Claw’s approach to separating AI reasoning from deterministic execution can also make agentic automation more reliable and repeatable, which could be “critically important” for CIOs trying to move agentic applications from prototype to production, according to Scott Bickley, advisory fellow at Info-Tech Research Group.

The deterministic controls combined with delegated authority, approvals, escalation rules and an auditable record of decisions and transactions can give enterprises a measure of “safe autonomy,” he said.

There are advantages for developers, too, especially in terms of productivity.

As the new execution layer packages up most of the underlying plumbing stack required to operationalize an agentic application, developers can focus on process design, offloading the mundane work of integrating and orchestrating the underlying components needed to put an agentic application into production, Bickley said.

That, in turn, is likely to result in business processes being automated faster, he added.

Lower AI consumption does not necessarily mean lower costs

However, that reduction in development and orchestration work does not eliminate the work required to govern agentic applications, but rather adds to the governance work required, Jha cautioned: “Development teams must encode policies and approval rules, choose autonomy levels, and verify outcomes.”

That outcome verification will also be important for CIOs in determining whether the promised cost savings and cheaper unit economics actually translate into lower costs, rather than simply lower AI consumption, according to Bickley.

While Oracle’s architecture can reduce the amount of AI inference needed for a task, the actual cost of an agentic application can vary based on factors including the actions it performs, token consumption, model used, testing, human review, exception handling and ongoing operational support, Bickley said.

As a result, CIOs should measure the cost of successfully completing a business outcome, such as resolving an accounting exception or filling a staffing gap, rather than looking only at the cost of AI units or tokens, Bickley added.

The analyst also cautioned CIOs about Oracle’s AI unit policy, which he said could introduce additional considerations into the cost of running Fusion Claw-based applications: “Unused pooled units expire without reimbursement, while Oracle can modify its action and model tables with at least 30 days’ notice.”

CIOs should demand per-outcome pricing in writing and let a bounded pilot build the business case, said Jha.

Adoption will depend on process and governance maturity

Those considerations, combined with the added governance requirements, are also likely to shape which enterprises adopt Fusion Claw first, even among existing Oracle customers.

“Existing Oracle Fusion customers with standardized processes, reliable master data, mature controls and high-volume work with measurable outcomes are likely to be the initial adopters. Finance reconciliation, supply-chain planning, staffing and other constraint-heavy processes are logical candidates because they combine analysis with calculations and repeatable execution,” Bickley said.

On the flip side, adoption will be “slower” in enterprises with heavily customized Fusion estates, fragmented data, multiple ERP platforms, weak segregation of duties, or limited AI governance, he added.

For enterprises willing to try out Fusion Claw, the execution runtime is now available as part of the Suite, Oracle said.

The company is also making 25 pre-built Claw-powered agentic applications available, including Ledger, Workforce Staffing, Shipping Consolidation, and Sales Territory Planning, aimed at helping divisional leaders, such as CFOs, CSOs, supply chain heads, and CHROs, improve productivity and reduce costs.

The company will make the other 21 apps available in October, it said.