Oracle: a business agent platform designed to make AI profitable

“Up until now, AI has been used to assist with work,” said Leone. “Now it will be used to do the work.” And without set hours: “If your AI stops working towards a goal at five o’clock in the afternoon, when the employee goes home, you’re not taking responsibility for that outcome.”

“AI should mean you hardly have to think about software at all,” said co-CEO Mike Sicilia in the opening keynote. “Why should anyone have to navigate through six systems just to do one thing?” The architecture is layered. Users no longer open menus and forms: they simply describe what they want to achieve. They do this via Agentic Home, Fusion’s new home screen. It is also a smartphone app with voice commands for the ‘tens of millions’ of users who do not spend their days in front of a computer: technicians, shop assistants, nurses and warehouse staff.

Fusion Companion handles these requests; it is a sort of intelligent switchboard that knows who the user is and what they are authorised to do, and routes the work: to a simple transaction, such as a holiday request; an agent-based application – that is, a programme built around a specific objective (‘deliver on time at the lowest cost’) that runs continuously; or, for longer and more complex tasks, to Fusion Claw, the specialist. Everything is based on data from the management system.

An agent that queries the language model record by record is very costly and makes more errors when dealing with large volumes of data. Oracle’s solution separates reasoning from execution to save tokens, the currency of AI. The language model plans and writes the logic, then the computation is handed over to traditional systems. Tokens are only consumed during the reasoning phase, and the system ‘takes notes’ on the paths it has already followed. In internal tests, Oracle says, reasoning time in subsequent executions is reduced by up to 50 per cent.

Greater autonomy requires greater safeguards. “More artificial intelligence does not mean fewer checks,” said Sicilia. “It is trust that enables AI to work.” Every time Claw is run, it operates within a set of rules defined by the company (policies, approval thresholds, procedures) and generates a ‘receipt’ detailing who decided what, regarding which data and with what alternatives.

But if an agent makes a mistake, who is held accountable? “The person who defines the level of governance, the policies and the authority: that person is responsible,” replied Chris Leone to a question from *Il Sole 24 Ore*. “Let’s put a control framework (harness) around the agent. It cannot act outside the delegated authority.” If the harness breaks, “the fault lies with development”; if the agent makes a mistake whilst remaining within the rules, the responsibility lies with whoever wrote them. “It’s an engineering problem, not a people problem.”

On Wednesday, also at AI World, Oracle will hold its Investor Day. In the first financial quarter, which ended in August, revenue rose by 30 per cent to $19.3 billion and net profit by 60 per cent to $4.7 billion. However, the share price has lost more than half its value over the past year. The market is concerned about the cost of the race to build data centres. These initiatives also serve to demonstrate that, for Oracle, AI generates revenue and is not merely an expense. As Sicilia said, ‘the goal is no longer AI in businesses, but businesses capable of achieving more’.