The five parts, one at a time
1. The business goal
Every run starts with an outcome, not a prompt for a chat. Oracle frames Claw work as "Claw Outcomes": reconcile the ledger, build the staffing plan, consolidate shipments, set territories. The goal is what the run is judged against.
2. Enterprise Operating Envelope
The envelope is the company's rulebook for the run, written before anything executes. Oracle lists what it holds: objectives, standard operating procedures, policies, constraints, permissions, risk thresholds, decision rights, approval requirements and escalation boundaries. Think of it as a job description plus a delegation-of-authority matrix. Launch coverage reports that AI can help turn company documents into the executable rules the envelope enforces.
3. Outcome Trust Harness
The harness applies the envelope to one specific run. Oracle says it governs identity, capabilities, data and actions for that run. Trade press adds that each job runs in a dedicated, isolated workspace that only the requesting application can open, has no independent internet access, and is removed when the job finishes. SiliconANGLE reports that the model has no direct access to business objects: it works through the runtime.
Inside the run: reason and execute
This split is the core design decision. The frontier model (Gemini or OpenAI at launch) handles what needs judgement: understanding the goal, planning the steps, interpreting surprising results, and re-planning. Deterministic computation handles what needs precision and volume: calculations, business rules, optimisation, simulation and transaction processing. Oracle's argument is that this makes complex work economical, because expensive AI inference is used only where intelligence is needed, and exact results come from code rather than from a model's guess.
Oracle also describes continuous re-planning: when an input changes mid-run, the model revises the plan and the runtime executes the new one. SiliconANGLE gives the staffing example of a plan being reopened automatically when an employee calls in sick.
4. Automation level
Customers choose, per process, how much authority to delegate. Oracle describes the range as "quick assistance to governed full auto execution within explicitly delegated authority". Coverage describes two practical modes: a research mode where Claw models, simulates and optimises and a person reviews before anything is implemented, and an autonomous mode within granted permission.
5. Outcome Receipt
Each completed run produces an auditable receipt: the authority applied, the evidence used, the decisions made, the actions and transactions executed, and the result. This goes further than a technical log. It is written for a process owner, governance committee or auditor to review.
Where Claw sits in the Fusion AI stack
| Layer | What it is | Who mainly works with it |
|---|---|---|
| Fusion Agentic Applications | Ready-made AI applications and workspaces by business area. 75 in total after the Claw launch. | Business users and managers |
| Fusion Claw | The governed execution runtime that powers 25 of those applications, the ones built for complex, long-running work | Process owners, consultants, security and audit |
| AI Agent Studio | The toolkit to build and extend agents. Futurum Group reports Claw ties into Agent Studio for no-code and pro-code extension. | Consultants and developers |
| Models and infrastructure | Gemini and OpenAI frontier models, running on Oracle Cloud Infrastructure | Oracle-managed |
Earlier Agentic Applications, such as those on our Agentic Applications page, mostly watch an area, prioritise exceptions and recommend actions. Claw-powered applications are aimed at the next step: completing large jobs within delegated authority.
The 25 launch applications
Oracle's announcement highlights four Claw-powered applications, one per pillar:
- Ledger (finance): reconciles accounting entries, investigates exceptions, performs financial analysis.
- Workforce Staffing (HR): builds executable staffing plans that balance skills, accreditations, availability, schedules, labour rules, cost and service needs.
- Shipping Consolidation (supply chain): models consolidation alternatives across timing, capacity, service commitments and cost.
- Account Territory Growth Plan (sales): models and optimises territory configurations and opportunity allocation against capacity and constraints.
The other 21 were not named in the announcement or the coverage we checked. We will add them when Oracle publishes the list. Check Oracle's documentation for what is enabled in your environment.
What a consultant designs for a Claw rollout
Claw executes the process you define, so the design work moves from screens to authority. For each process, agree and document:
- The outcome and its success measure. For example, "all reconciling items explained, with fewer than X open exceptions above Y value".
- The envelope. Which policies and SOPs apply, which constraints are hard, and what the risk thresholds are in money, volume or count.
- Decision rights and approvals. What Claw may post on its own, what needs a human approval, and who that human is.
- Escalation. When Claw must stop and hand back, and to whom.
- Data readiness. Master data quality, open interfaces and known data gaps. Analysts flag fragmented data as the main brake on value.
- Receipt review. Who reads Outcome Receipts, how often, and what triggers a change to the envelope.
- Cost tracking. AI unit consumption per run against the time saved. Info-Tech Research Group notes that costs vary with actions, model choice, testing and human review, and that unused pooled units can expire.
Begin with a person reviewing every plan, and widen the delegated authority only when the receipts show consistent, correct decisions. Our Build Readiness Checklist covers the security, data and testing blocks.
Cost model, as reported
SiliconANGLE and TechTarget report that Claw is included at no additional charge with the Fusion Agentic Applications offering, which is sold separately from standard Fusion subscriptions, and that usage is paid in Oracle AI units as work completes. The Letter Two reports that AI units are consumed by the model's planning work, while the deterministic execution consumes none, and that the runtime reuses successful solution paths from earlier jobs. These points come from reporting, not from the announcement, so confirm them against your contract.
Checked Component names, the four named applications, model providers, the 25 and 75 counts, and OCI hosting come from Oracle's announcement, Oracle Extends Fusion Agentic Applications with Introduction of Fusion Claw (29 Sep 2026). Isolation, modes, pricing and Agent Studio details come from The Letter Two, SiliconANGLE, TechTarget, CIO and Futurum Group. Verified 1 Oct 2026. The runtime is new; behaviour, availability and licensing can change, so verify in your own environment.