Las Vegas – October 29, 2026 -- Oracle says 25 of its 75 packaged agentic applications for Fusion are now Claw-enabled. Claw is a governed runtime that executes analyst-grade work in finance, supply chain and workforce planning and commits results to Fusion within business-set limits. Oracle announced the expansion at Oracle AI World in Las Vegas on October 26, 2026, and says all of it is available in the current Fusion release.
Oracle extends Claw across ERP, HCM, SCM and CX applications
Oracle first introduced Claw on September 29, 2026. The runtime takes over when an agentic application needs deep analysis across millions of records. Its scope covers reconciliation, analysis, planning, simulation and optimization, and runs can last hours.
Claw-enabled applications include Ledger (financial close), Workforce Orchestration, Supply Planning Allocation, Procurement Spend, Service Health Check and Strategic Field Demand Planning. Customers can build their own extensions with no-code tools in AI Agent Studio, or in VS Code and a command-line interface. Google Gemini and OpenAI models handle reasoning at launch.
Fusion Companion routes a plain-language request to a transaction, an agent, an agentic application or Claw. It runs in Fusion Agentic Home, on iOS and Android, and as a ChatGPT plug-in. Oracle reports 5.7 million agent runs a quarter across its customer base.
Claw plans once and lets deterministic code handle the volume
Oracle says this design keeps results consistent across millions of records and limits token spend to the planning step. Two operating modes matter for risk owners. Research mode performs the analysis and stages actions, which can number in the hundreds, for a person to accept, edit or reject. Full-auto mode re-checks the live state of Fusion before committing anything, so it never acts on a stale plan.
Chris Leone, who leads Fusion development, set a "5pm test": AI that stops pursuing an outcome when staff go home is not owning it.
Oracle cites up to 50% less reasoning time on repeat runs
Claw records the path each run takes, so similar runs reason faster. Oracle's internal testing showed up to 50% less reasoning time on repeat runs, but the company has not published benchmarks. The three Claw scenarios Oracle ran on stage were vendor demonstrations, not audited customer results.
Leone asks customers to pilot one of the 75 applications now
Leone urged customers to take one application off the shelf and run it with a pilot group. The process should have a clear financial measure, such as close duration or freight cost. He suggested trialling Agentic Home now and extending it to a broad user base by January or February 2027.
Governance comes before full-auto mode. Companies must define an operating envelope: which actions Claw may commit alone, which need approval and which must escalate. Finance, risk and audit should sign this off first. Starting in research mode lets Claw stage actions until results match what human planners would have done. The Outcome Receipt should become part of internal audit evidence from day one.
Because Claw acts on the live system of record, gaps in master data, approval hierarchies or policy set-up will show up in its results. They should be fixed before a pilot starts.