Oracle Fusion AI: Autonomy is earned, one outcome at a time

When a tech executive who has spent his career in AI says the pace of development is so rapid as to be entirely unpredictable, you tend to sit up and take notice. "This is moving at an unprecedented pace. Some of [what we are releasing now] honestly wasn't possible eight months ago, but the models have improved so much and our own tool sets have improved so much," said Kaushal Kurapati, GVP Product Management, Oracle Fusion AI. "I've never seen this pace of change, and I've been working in AI my whole career."

If you feel somewhat blindsided by AI, despite the apparent omnipresence of the technology, don't. Even those deeply embedded in the industry and who are driving the creation and adoption of it, are themselves a little gobsmacked.

Kurapati was discussing AI and trust. Agentic AI is here, and it is being built into one of the company's flasgship lines, Fusion Applications. Speficially, he is referencing Fusion Claw, the governed runtime behind at least 25 of Oracle's new Fusion Agentic Applications. It runs longer, bigger jobs than a single Fusion AI agent could, with continuous replanning and deterministic execution.

And yes, we did ask Kurapati about the Claw in the name, and he confirmed that it has nothing to do with OpenClaw beyond the latter serving as a source of inspiration.

Oracle has added AI in layers, starting with AI Agent Studio, a platform where agents are built and governed across ERP, HCM, SCM and CX. Ready-made agents sit on top, such as Document IO for invoice processing. Just-launched Fusion Claw Agentic Applications include Ledger, Workforce Orchestration and Procurement Spend. Fusion Claw tasks run in short-lived containers with no internet access and no direct access to the system of record, and they work only through the constrained agents. Each action leaves an "outcome receipt" recording who decided what, when, from which sources and with what changes.

Alongside Claw are Fusion Companion and Agentic Home, which work out which workflows, approvals and agents are needed to reach a user's goal (and can be used from tools such as ChatGPT), and the Fusion Setup and Configuration Registry, a public Git repository that turns Fusion setup into skills agents can execute.

For agentic AI to do work, perform tasks, and become useful, it requires trust. And as anyone who has handed the car keys to a newly licensed teenager knows, trust starts small and extends a little at a time. Enterprise software now faces the same question with AI agents, and the stakes are highest in the system of record.

"A lot depends on the kind of workflow you're executing, and the kind of vertical you're operating in," Kurapati said. "Think about it; an agent is doing work on your behalf, so there's liability, there's financial risk. Those risks dictate the degree of autonomy you will be comfortable in giving."

Automated promotions, he quipped, are not on the cards. "A lot of us would like it, perhaps," Kurapati said. "But there are some decisions that are obviously taken by the human because accountability is at stake."

Kurapati leads the platform that builds and governs agents across Oracle's Fusion applications, having rejoined Oracle in 2025 from Salesforce.

Faster, not hands-free

Invoice processing is the obvious test case. Oracle's DocIO agent works through them, but customers aren't ready to let it sign the cheques. "Earlier, it used to take the whole day, and they would process X amount of invoices. Now they can process maybe 10 X in a couple of hours," Kurapati explained. "All the humans are doing now is scanning the output, and giving approval."

Human in the loop, then, but a much quicker loop. Gartner is on the same page: its 2026 Hype Cycle for Agentic AI says fully autonomous agents are not ready for the majority of enterprise use cases, though it still expects agentic AI to make at least 15% of day-to-day work decisions autonomously by 2028.

Notably, in a separate discussion, we asked Alan Petrie, Head of Corporate Data and Analytics at Heathrow Airport (and an Oracle Fusion Applications award-winning reference customer) if there was any frustration between the enterprise AI promises, and the enterprise AI reality. Petrie didn't hesitate. "Absolutely. What happens on the drawing board doesn't translate to what happens in production," he said.

Petrie added that 'personal' AI still works far better than 'enterprise' AI, and agreed that this was at least in part owing to both the complexity and idiosyncrasy of business, and the necessity for governance and control.

Asked by TechDay about the biggest challenge of bringing agents into the system of record (the bit everybody is justifiably reluctant to interfere with), Kurapati pointed to Fusion Claw's 'outcome receipt'. "It will record exactly who made the decision, what point in time, what were the sources used, what were the changes made to the system of record," he said. "It's a confidence-building tool. After a point, maybe you won't look at every single outcome receipt. Maybe you will let go." In a word, added Kurapati, auditability.

Much as you might do with a new hire. "You would first closely monitor them, and once they're doing the job, responsibility and trust builds," Kurapati said. 

Drift happens

Then there is model drift: models get replaced, and agents can behave differently. Oracle's answer is Atlas, its Agentic Testing and Lifecycle Automation Suite. "Every time you roll something into production, you can run it on mock data, simulated data, test data sets, or real data," Kurapati said.

The checking is done in plain English. "You can ask, 'Trace the entire execution path and indicate if the execution is the same as a week before, and the same with a newer model?'," he explained. The rest, he conceded, depends on customer hygiene. "At that point, it's not in our hands as much."

Oracle doesn't publish a single trust score. Its proxy is production. "If they were not trusting it, they wouldn't put it [into production]," Kurapati said, adding that for every agent live in production, roughly 15 more sit in staging and testing.

A Forrester Consulting study for Boomi found 86% of organisations have moved AI agents beyond the pilot stage, yet just 34% trust the actions those agents take. IDC reports that 92% say governing agents is critical, while only 44% have actually implemented a policy.

Move or be left behind

Gartner reckons the average global Fortune 500 enterprise will have more than 150,000 agents in use by 2028, up from fewer than 15 in 2025. Put to him that this may yet prove conservative, Kurapati agreed, and had a chuckle about Neils Bohr's observation that predictions are hard, especially about the future.

His advice to the sceptics is blunt. "You need to adapt. It's evolutionary dynamics in some sense, survival of the fittest, right? So resisting change is probably not the greatest approach."

But there was a caveat, and a good one. "As a human being, I don't want to outsource my thinking ability," Kurapati said.

Which is advice worth keeping a receipt for.