Oracle announced 25 new Fusion Claw-powered applications on September 29, extending its enterprise AI software. The central business question is not simply whether an AI system can finish a task, but what it is authorised to do along the way.
From assistance to delegated work
Oracle describes a system that combines AI reasoning with conventional computation. Its controls set permissions, approval requirements and risk boundaries, while an outcome record documents decisions and actions. These are vendor descriptions, not independently verified performance results.
What buyers should test
CapKet analysis: A useful demonstration would start with an ordinary exception: an invoice that does not match its purchase order. Ask whether the system merely flags the mismatch, changes the accounting record or initiates a payment. Those are materially different levels of authority.
A second test should examine failure recovery. If the workflow stops halfway through, can the team identify which steps completed and which still require attention? A log is useful only if the people responsible can interpret it and connect it to the underlying records.
The buying decision also needs a clear owner. Finance, operations and technology teams may each see different risks in the same automated process. Recording who approves a permission change is therefore a practical evaluation step, not proof that a particular product is safe.
What remains to be established
CapKet has not tested Fusion Claw or inspected customer deployments. The announcement does not establish a return on investment for any reader’s organisation. Compare the control question with our NVIDIA agent-guardrails explainer.
Primary source record: Original announcement and Oracle’s explanatory blog. Checked September 30, 2026, around 18:23 IST. The linked issuer material supports the attributed facts; CapKet’s explanatory examples are identified above.