Oracle Fusion Claw: What a Governed Agent Runtime Inside Your ERP Means for Finance and Operations Teams

Most enterprise AI agent announcements promise the same thing: software that does the work instead of just describing it. What is rarer is a vendor putting the guardrails at the center of the pitch. On September 29, 2026, Oracle did exactly that with Fusion Claw, a runtime that sits inside its Fusion ERP applications and splits “thinking” from “doing.”

In this post we unpack what Oracle actually announced, why the reasoning-versus-execution split matters for cost and risk, how its governance model works, and what finance, HR and operations leaders should ask before switching on autonomous agents in systems of record.

What Oracle Announced

Oracle describes Fusion Claw as “a governed agentic execution runtime for Oracle Fusion Agentic Applications”. It launched with 25 Claw-powered applications, bringing Oracle’s Fusion Agentic Applications portfolio to 75, and Oracle says all 25 are available now.

The target is longer-running business work rather than one-off chat answers: reconciling large volumes of accounting entries and investigating exceptions before period close, consolidating shipments, building workforce staffing plans, or planning sales territories. Oracle’s SVP of Fusion Applications product management, Natalia Rachelson, told SiliconANGLE, “We did not previously have solutions for advanced optimization.”

It is also part of a broader pattern we have covered: agents becoming a default feature of enterprise platforms. If you want the wider context, see our piece on why AI agents are becoming a default feature in enterprise software.

The Core Idea: Separate Reasoning From Execution

Models plan, conventional software computes

The architectural claim at the heart of Claw is simple. A frontier language model handles reasoning and planning, and then execution passes to deterministic enterprise computation for high-volume, repeatable transactions. Per SiliconANGLE, the model is invoked only when judgment is needed, so work over large datasets runs outside it.

That matters for two reasons:

Cost. Running a language model over every ledger line would be slow and expensive. Using it to decide what to do and conventional code to do it keeps token consumption proportional to the judgment required, not the data volume. Note that this economic advantage is Oracle’s own claim, and the company has not published measured savings.
Reliability. Arithmetic, posting and reconciliation are exactly the places where “probably right” is not good enough. Handing them to deterministic code removes a class of hallucination risk.

The model never writes directly to your records

SiliconANGLE reports that Claw runs in an isolated environment that keeps the language model from directly changing Fusion business objects. In effect, the model proposes; deterministic software validates and executes. This is the same principle we described in why action-layer governance is the new standard for enterprise AI agents: control what an agent can do, not just what it can say.

Model choice is limited for now

The initial release supports frontier models from Google Gemini and OpenAI, and, per SiliconANGLE, customers cannot yet bring their own lower-tier open models. Oracle has said it plans to add more models. If model flexibility or data-residency preferences drive your AI strategy, that is a constraint to note.

How the Governance Model Works

Oracle’s governance story has three named parts, according to ERP Today’s summary:

Enterprise Operating Envelope. Defined before a run, it captures objectives, procedures, policies, permissions, risk thresholds, decision rights, approval requirements and escalation boundaries.
Outcome Trust Harness. Applies the envelope to each run and controls the agent’s identity, capabilities, data and actions.
Outcome Receipt. Recorded after a run, it logs the authority applied, evidence used, decisions made, transactions executed and results. Oracle says it holds more context than a conventional audit log, including the full text of the policies used.

Graduated autonomy

Automation is not all-or-nothing. Customers can require employee review of plans, or let the application execute within delegated authority. Rachelson said that “any task can be in full auto mode,” while expecting organizations to start with more human review and expand as they gain confidence.

Claw can also work with third-party systems through APIs, the Model Context Protocol and Agent2Agent, according to SiliconANGLE, which matters because few finance or supply-chain processes live in a single application.

Pricing: Included, But Not Free

Claw is included at no additional charge with Oracle’s agentic applications product. However, per SiliconANGLE, that product requires a separate purchase beyond a standard Fusion subscription, and customers pay for AI units consumed as work is completed. This is a consumption model, which means budgeting depends on how much work your agents actually complete. Our guide to AI agent pricing models and the “agentic enterprise license” covers how to forecast that kind of spend.

What to Be Skeptical About

Good governance language is not the same as proven results. A few honest caveats:

Vendor-reported benefits. SiliconANGLE’s coverage contained no independent analyst comment, and Oracle did not provide measured cost savings or evidence of production performance.
Agent hype has a track record. Gartner has predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. Governed runtimes like Claw directly address the third reason, but not necessarily the first two.
Lock-in. A runtime that lives inside one vendor’s ERP is powerful because it is close to the data, and sticky for the same reason.
Receipts are only as good as your policies. An Outcome Receipt can show which policy was applied, but it cannot tell you the policy was a good one.

A Practical Playbook for ERP and Finance Leaders

ERP Today’s own practical advice for ERP teams is worth echoing: keep process documentation current, involve controllers and internal audit early, and expand automation in stages with measured ROI.

Before you switch anything on

Document the process first. An operating envelope can only encode decision rights that your organization has actually written down.
Bring audit in on day one. Ask internal audit whether Outcome Receipts satisfy your evidence requirements.
Pick a bounded pilot. Exception investigation in the ledger is a natural start because the output is a recommendation a human can verify.

While you scale

Start in review mode. Require human approval, measure how often reviewers change the plan, and only widen delegated authority when that rate is low and stable.
Track cost per outcome. With consumption pricing, compare AI units spent against the hours saved.
Inventory every agent. As vendors embed agents in more applications, visibility becomes the issue. See our analysis of AI agent sprawl.

Conclusion

Fusion Claw is notable less for what its agents can do than for how Oracle has chosen to constrain them: models reason, deterministic code executes, policies are defined in advance, and every run leaves a receipt. That design is a credible template for putting agents near systems of record. The open question is whether the economics and reliability hold up in production, and so far the evidence is Oracle’s own.

Key takeaways:

Separate judgment from execution in any agent design you build or buy.
Treat “full auto” as something a process earns, not a default setting.
Demand measured results, not just governance vocabulary, before expanding scope.
Plan for consumption-based costs and model-choice limits from the start.

Frequently Asked Questions

What is Oracle Fusion Claw?

It is a governed agentic execution runtime for Oracle Fusion Agentic Applications, announced on September 29, 2026. It lets AI agents plan work with a frontier model and then execute it through deterministic enterprise software.

Which AI models does Fusion Claw use?

At launch it supports frontier models from Google Gemini and OpenAI. Oracle has said it plans to add more models, but customers cannot yet choose their own lower-tier open models.

Can the AI change my ERP data directly?

According to SiliconANGLE, Claw runs in an isolated environment that keeps the language model from directly changing Fusion business objects. Changes go through deterministic execution under the policies you define.

What is an Outcome Receipt?

It is a record created after each run showing the authority applied, evidence used, decisions made, transactions executed and results. Oracle says it includes more context than a typical audit log, including the full text of the policies used.

How much does Fusion Claw cost?

Claw is included at no extra charge with Oracle’s agentic applications product, but that product requires a separate purchase beyond a standard Fusion subscription. Customers pay for AI units consumed as work is completed.