An AI agent can be capable enough to do the job and still be too expensive to put to work. Give it a million accounting records, and the economics become harder to ignore. If a language model reasons through each entry individually, processing costs accumulate while the company waits for results.

Oracle’s answer is Fusion Claw, a new runtime designed to make complex AI work more economical within its enterprise applications. Inspired by OpenClaw’s idea of giving an agent its own computing ecosystem, it provides an isolated environment where agents can work through specialist assignments under customer-defined rules and permissions. AI develops a plan and assembles the tools needed to execute it, then hands the work to deterministic computation, avoiding a separate call to the language model for every calculation.

The IT giant is betting that separating AI reasoning from execution will make it economical to automate work its applications still leave to specialists. How much work can companies confidently delegate before an expert needs to step in and reconsider the plan?

Oracle CEO Mike Sicilia sees an opportunity to make enterprise software more useful.

“Today, people still spend considerable time pulling data out of systems, analyzing it, deciding what to do, and then going back into those systems to take action,” he tells me in a written statement. “With Fusion Claw, agentic AI can do more of that work on its own while keeping people in the loop to the extent the customer chooses.”

Sicilia explains that Fusion gives agents a practical place to execute this work because it already holds customers’ financial records and business rules. Existing access controls restrict what agents can use, while customers determine which actions they can take independently and which require human approval. Those boundaries allow AI to “reliably handle more work,” he says, “freeing resources and giving people more time to solve problems, serve customers, and drive innovation.”

For Oracle, the promise of completing more work comes with an opportunity to strengthen growth in its applications business. In its fiscal first quarter of 2027, Fusion Applications grew 14% year over year. Broader cloud applications revenue rose 10% to $4.2 billion, while cloud infrastructure grew 121% to $7.4 billion.

“Fusion Claw builds on the work our agentic applications already perform toward a business outcome and extends that capability into specialist work that people have traditionally carried out outside our system,” Chris Leone, Oracle’s executive vice president of applications development, tells me in an exclusive interview.

Customers are already using more of the company’s embedded AI. Usage grew 42% sequentially during the quarter, while production AI-agent activity nearly doubled, according to Oracle. Those figures predate Fusion Claw, giving the company evidence of increased activity without yet establishing demand for the new runtime. However, the company spent $55.7 billion on capital expenditures in fiscal 2026, and sold $20 billion of stock to fund further expansion. Shares fell after Oracle recently issued a force majeure notice tied to its Project Jupiter data center in New Mexico.

Applications give the company an AI growth path that relies far less on bringing the next data center online.

Agentic AI Expands Into Specialist Business Tasks

Fusion is Oracle’s cloud business suite and system of record. Its agentic applications, introduced in March, organize work around a business outcome. The company unveiled 25 new Claw-powered applications that cover ledger reconciliation and workforce staffing, alongside shipping consolidation and account territory growth plan. Oracle describes its broader portfolio as 75 agentic applications, and customers can build their own through Oracle AI Agent Studio.

To illustrate the difference, Leone described a hypothetical hospital manager aiming for 95% utilization across five hospitals. The manager could authorize nurses to move between locations and permit 25% overtime, giving Claw a set of constraints within which to develop a staffing plan.

“In the past, solving that problem would have required us to build an entirely new module. Now we can give the problem to Claw, with the appropriate controls around it, and have it work toward the outcome the customer wants,” Leone says. “That gives us a different kind of runtime execution capability.”

Oracle also promises continuous replanning as conditions change. Planning software already handles constrained optimization. Claw’s larger claim concerns how much work follows the plan without someone manually advancing it. The company’s July builder announcement promised no separate runtimes. Leone says “Claw remains inside Fusion on Oracle Cloud Infrastructure, adding an isolated execution container without requiring a separate customer environment.”

Fusion Claw Separates AI Reasoning From Execution

Claw uses frontier AI models, including Google Gemini and models from OpenAI, to plan an assignment. “What distinguishes Fusion Claw from other approaches is that we take an orchestration agent and place it inside an isolated runtime container,” Leone says. “When it receives the outcome we want it to achieve, it uses agentic AI to reason through the assignment and develop the plan it will execute.”

The agent then assembles the tools needed to carry out that plan and hands them to the compute layer. “We then execute those tools deterministically. It becomes a codified artifact that we can run against large bodies of data to execute the work economically and precisely toward the outcome the user requested,” he says.

Leone contrasts this approach with the Model Context Protocol (MCP), which connects agents to tools, and application programming interfaces, which let software systems communicate. He argues that those approaches have not given users the same combination of governed execution and deterministic computation at scale.

Claw also records the results of completed runs. When a customer approves a successful plan, a similar assignment can draw on that experience instead of starting its reasoning from scratch. “One can save economic costs and reasoning costs — we can cut those down by 50%-plus,” he claims. However, Oracle provided neither supporting benchmarks nor Claw-specific pricing, leaving the effect on customer bills unclear.

Leone’s reference points run past enterprise software. He cites Muse, Meta’s new consumer runtime, where an orchestration agent sits inside a container and takes in workloads to execute. To him, the split between reasoning and execution looks familiar, but he’s quick to say the governance does not. “That same architecture, and how they’ve separated reasoning and execution, is very similar — not the same — but it was all inspired by OpenClaw and where OpenClaw started, where you give a computer to an agent,” he says.

Enterprise AI Guardrails Still Need Human Judgment

Oracle refers to customers’ governing framework as an ‘Enterprise Operating Envelope’, which captures operating policies and delegated authority, including approval requirements and escalation boundaries. An Outcome Trust Harness applies those controls to each run and an Outcome Receipt records the supporting evidence and resulting actions so customers can inspect what happened. During a financial close, Leone claims Claw can examine a million ledger records for discrepancies and assemble the evidence a controller needs to decide whether the books are ready to close. It can also create and execute transactions within the authority the customer grants.

“No agent gets access to anything that is not part of our role-based access control,” Leone says. Customers can stage actions for review in research mode or authorize full automatic execution within defined boundaries. But a plan can follow every rule and still lead to a poor business decision. Combining shipments might save money while delaying an order a customer urgently needs.

I asked Leone how Claw recognizes trade-offs that require human judgment. “If it reaches one of the defined boundaries, it will escalate and ask a human to step in and decide how it should proceed,” he explained. “We also have a research mode, where it performs the same workload and returns with the actions it proposes to take, staged for human review.”

An audit trail can help explain a decision without proving it made business sense. Leone expects experts to spend more time defining the policies agents follow, while remaining accountable for the results. That leaves companies responsible for keeping those rules current and recognizing circumstances that would have led an experienced employee to stop and reconsider.

The Race For Autonomous Enterprise

While Oracle is betting that Fusion’s business data and built-in controls will give Claw an advantage, rivals are making similar claims about their own platforms’ ability to turn AI reasoning into completed work.

SAP offers the most direct comparison. Its Autonomous Close Assistant targets work that overlaps with Claw’s Ledger application, including financial reconciliation and the resolution of accounting exceptions. The platform’s broader Autonomous Suite includes more than 50 Joule Assistants coordinating over 200 specialized agents. Both companies’ applications already hold the business records agents need, along with the rules governing how employees can change them. For finance leaders, the question is how effectively each vendor can turn that existing knowledge into work they feel comfortable allowing agents to complete.

Workday approaches the opportunity through its strength in human resources and finance, while Oracle can extend its argument into supply-chain operations through applications such as Shipping Consolidation. Likewise, ServiceNow and Microsoft emphasize coordinating agents and workflows across systems. Oracle places greater emphasis on execution inside Fusion, where the underlying records and access controls already reside.

That emphasis on connected business context helps explain Leone’s concern about fragmented enterprise systems. An agent might recommend a new product offering, he says, but the recommendation delivers little value if the company then needs months to connect the applications required to sell it.

“I think organizations will increasingly consolidate their systems of record and underlying architecture around broader business processes. That consolidation would allow them to execute more of the work across those processes autonomously,” he says. “This is the direction enterprise systems will have to take if companies want to achieve the levels of productivity they are looking for from AI.”

That vision also aligns with Oracle’s commercial interests. If closer integration allows agents to complete more work, customers may have a stronger reason to consolidate their operations around Fusion.

“Connecting agents that can make decisions and tell you what to do addresses about 10% of the problem. The other 90% of the productivity opportunity lies in doing the work and executing it at scale. We haven’t reached that point yet because people haven’t moved to the next level of execution,” Leone says. “Execution is the key, and that’s what we’re focused on.”

Fusion Claw could give Oracle’s applications a larger role in running the business, provided the cost of supervising agents does not consume the savings from using them. Oracle can supply the execution engine, but customers still have to decide what good execution looks like—and when the system should stop.