Oracle’s agent strategy arrived on Monday as a catalogue rather than a demonstration. At its AI World customer conference the company introduced 25 applications that act on their own inside Fusion Cloud ERP, human capital management, supply chain and manufacturing, and customer experience. No single product carried the announcement. The volume did.
The applications run on something Oracle calls Fusion Claw, described in its materials as a governed execution runtime. In plain terms, a model reasons over a business process, re-plans when a step fails, and then carries out the task inside the system of record instead of handing a suggestion back to a person.
Oracle paired the launch with an architecture it calls Deep Trust AI. The label is marketing, but the problem behind it is concrete. An agent that can approve a purchase order or close a service ticket needs an identity, a permission boundary and a way to be switched off mid-task.
Enterprise software has spent three years shipping assistants that draft text and answer questions. Moving to agents changes what a vendor is selling. A copilot that proposes a journal entry is a feature. An agent that posts one is a participant in the financial close, and its mistakes land in the ledger.
That distinction is where Oracle’s argument rests. The agents inherit the approval hierarchies, segregation-of-duties rules and audit logs that Fusion customers already maintain, because they write into the same tables. A standalone agent bolted onto the outside of an ERP has to rebuild those controls, and usually rebuilds them badly.
The commercial logic is easy to follow. Fusion applications are the part of Oracle’s cloud business that bills per seat and renews on multiyear contracts, and applications carry higher margins than rented compute. Infrastructure gets the headlines and the capital spending. Applications pay the bills.
Oracle is not alone in making this pitch. SAP has been embedding agents across S/4HANA, Workday across finance and HR, and Salesforce sells Agentforce as an autonomous layer over its customer data. Microsoft attaches agents to the productivity software most employees already open every morning. Each company is promising that its agents know the business context, which is another way of saying the agent sits next to the data.
Analysts said the unresolved question is pricing. Charging per seat assumes a person at the keyboard. When software performs the work, the vendor wants to charge per task, per outcome or per unit of consumption, and finance departments have learned to distrust all three after a decade of cloud bills that grew faster than headcount. Oracle has not published list prices for the 25 applications, and the structure it chooses will say more about the product than any demonstration on stage.
There is a second question, less discussed. Agents that re-plan take actions nobody wrote into the runbook. A procurement agent told to consolidate suppliers may cancel a contract that a factory depends on. The failure is not a wrong answer in a chat window. It is a real commitment, sometimes with a real penalty attached.
Oracle’s answer is procedural. Deep Trust AI bundles the logging, policy enforcement and kill switches that a chief information officer will be asked about during a procurement review, and the company is selling the governance as a feature rather than an afterthought. Whether that satisfies auditors, who have their own tests for who approved what, is a separate matter that will be settled customer by customer.
Oracle has made this kind of claim before. Its Autonomous Database, launched in 2018, promised a system that tuned and patched itself with almost no administrator input, and the phrase “self-driving” did a lot of work in the company’s marketing for years afterward. Customers adopted it, but many kept a database administrator watching, which is the normal fate of automation in systems that carry payroll and revenue recognition.
The current pitch faces a harder version of that test. A study published last year by researchers at MIT found that the great majority of enterprise generative-AI pilots had produced no measurable return, with the failures concentrated in workflows that crossed departmental boundaries. Agents that touch procurement, payroll and logistics cross those boundaries on every task.
Implementation partners will carry much of the delivery. Accenture, Deloitte and the large Indian service firms staff the teams that configure Fusion for multinational customers, and their billable hours determine how quickly an agent can be tuned to a specific company’s approval chain. That work is not in the licence price.
Twenty-five is a large number to announce at once, and the count itself is a marketing metric. What matters is how many of them ship with named reference customers, how many run without a human approving each step, and what happens to the first company whose agent books something it should not have. Oracle’s fiscal year gives it a natural deadline to answer the first of those. The other two will take longer.
