Oracle Wants to Run Your Operations.
Write the Rules First.
A conversation with Rondy Ng
EVP, Applications Development, Oracle

On day 2 of Oracle AI World 2026, Rondy Ng, Oracle’s EVP of Applications Development, sat down with The Enterprise Edge and explained why Oracle plans to run its customers functions and operations themselves, within rules the customer writes, instead of handing functional teams more AI tools. The agentic apps and governance layers are shipping now. The customer evidence is due within two quarters. Agentic AI will reward companies that define their outcomes and controls before turning on autonomy. Each leader has a different first move.

​

• What Oracle is building: Fusion Agentic Applications that model a whole department, such as the ledger, collections or payables, and drive it toward outcomes a manager sets. Oracle released 25 apps powered by its new Fusion Claw runtime this fall, bringing the total to 75.

• How it works: A frontier model plans the work. Fusion Claw runs the heavy computation deterministically on Oracle Cloud Infrastructure. An Enterprise Operating Envelope turns your policies and desk manuals into limits the system enforces, and an Outcome Receipt records every decision for auditors.

• Where it pays first: High-volume, rules-heavy work. Ng would make accounts payable “100% machine owned.”

• What’s still missing: Proof of customer ROI. Asked whether customers can generate it yet, Ng said: “No, not yet.” He expects proof within three to six months.

• What to do now: Pick two or three finance outcomes. Write the envelope before you raise the autonomy setting. Fund at least one growth use case alongside the cost cuts.

​

What does it mean for software to "own the operating model"?

​

For twenty years, ERP vendors have sold finance teams better instruments: more reports, faster screens, a cleaner journal-entry form. The accountants still did the flying.

​

Rondy Ng wants the software to take the controls. “We’re not providing you a bunch of tools to let you actually do your job,” he told me. “We want to own your core way of driving things.”

​

That one sentence matters more than might be immediately evident. Oracle calls the idea a “system of outcomes” running alongside the system of record. People set the objectives, the authority and the accountability, and the software carries the work through. Ng’s fellow EVP Chris Leone summed it up: “The next era of enterprise AI is all about delivering software that increasingly does the work.”

​

Ng admits the thinking settled recently. Six to nine months ago, he said, Oracle was selling “hundreds of agents” that automated tasks and “answer[ed] questions here and there.” That helped, “but that was not a strategy about how you should run accounting totally differently.”

​

Buyers have felt the same gap. Gartner’s November 2025 survey of 183 finance leaders found 59% of finance functions using AI, while 91% rated its impact low or moderate. Scattering assistants across a department rarely changes how the department performs. Ng is betting that redesigning the department’s work end to end will.

​

What is an agentic app, in plain finance terms?

​

When Ng explains an agentic app, he describes a department, not a feature. The Ledger app is modeled on a working day in accounting. A manager sets the outcomes: bring these accounts to this level of accuracy and completeness, clear these exceptions, finish these intercompany reconciliations.

​

His example is ordinary on purpose. Compare a liability account’s balance today with its balance on the same day last quarter. Some of the difference is legitimate because spending changed. Some of it points to trouble, such as a manual journal entry nobody can explain. Today a specialist pulls reports, dumps them into Excel, runs macros, chases evidence and gets sign-off. With the agentic app, you give it a standard operating procedure and a single instruction, and it returns the substantiation.

​

“That’s complex work handled by specialists today with reports and Excel. Now it’s done with one single command. And that’s just one example of millions.”

​

The human role moves up a level. “You are the accounting manager,” Ng said. “You define what people do, the goals.” Some work runs on its own and some goes to people. The app walks the close from day minus two to day minus one to the close itself. Each subsidiary then gathers its evidence, the local CFO signs off, and the package goes to corporate consolidation.

​

Collections follows the same pattern. Oracle’s Collectors Workspace app, launched in March, targets days sales outstanding and promise-to-pay conversion. Picture a collections manager with 12 collectors and a queue of disputes, returns, quality complaints and late payers. Deciding who calls whom first means weighing perhaps 20 factors, including terms, payment history, delinquency behavior and broken promises to pay. The app reads that data in real time and builds a ranked worklist. It can send reminder and warning letters automatically and draft a call script so the collector arrives prepared. Ng sees that as the difference from “giving you 15 reporting tools to get the data out.”

​

For buyers, this changes how to evaluate the product. Score an agentic app against how your department’s manager actually decides things each morning. A feature checklist won’t tell you much.

​

Why does Fusion Claw separate the thinking from the doing?

​

Fusion Claw, announced September 29, is the runtime under the newest apps. It borrows its name and approach from the open-source OpenClaw agent and adapts both for enterprise controls. The core design choice is a division of labor. Ng explained why.

​

“Clearly, the reasoning and the planning is where the LLM does a great job,” he said. Deterministic execution handles everything else: the computation and the reconciliations across millions of transactions, run in containers on OCI. “If you actually put [that] into an LLM, good luck getting any meaningful answers for that kind of volume.”

​

This split addresses two of the three reasons Gartner gives for its prediction that more than 40% of agentic AI projects will be canceled by the end of 2027: rising costs and inadequate risk controls. Paying model rates only for planning keeps token bills in check. Running the arithmetic in ordinary code removes the risk of a model inventing an answer at the step where accuracy matters most. The third reason, unclear business value, is the one Oracle still has to prove.

​

Some practical details:

• Models: Claw currently uses Google Gemini and OpenAI models, and Oracle says more are coming.

• Price: According to TechTarget, Claw is included with Agentic Applications subscriptions, which are priced separately from Fusion. Oracle SVP Natalia Rachelson described the goal as tying “outcomes to consumption.”

• The limit: HyperFRAME Research notes that deterministic execution guarantees the steps run correctly. It doesn’t guarantee the plan was right. Testing still matters.

​

What is the Enterprise Operating Envelope, and why does it matter most?

​

Earlier agentic tools left a gap that cost savings and autonomy settings can’t close: something has to tell the agent what it may not do. “To me, the governance in particular in my space is huge,” Ng said.

​

Oracle calls its answer the Enterprise Operating Envelope. It contains the organization’s SOPs, controls, policies, risk thresholds and delegated authority, the same rules that already govern human accountants. Ng’s payables example: one company may have 75 different hold conditions that require a person’s attention. Those conditions already exist in the desk manuals AP clerks follow every day.

Traditional automation required someone to rebuild those manuals as workflow rules. “It’s impossible for them to codify [it] themselves over time,” Ng said. “That’s too IT.” In the new model, the written policy is the input. An Outcome Trust Harness turns the text into enforced controls covering identity, permissions, data and actions. Each completed job produces an Outcome Receipt that records the authority applied, the evidence used, the decisions made and the transactions posted.

​

This matters for two reasons.

1. The controller writes the rules, not IT. Whoever owns the policy document now configures the automation. Finance leaders gain direct control and take on a new responsibility.

2. The audit trail is built in. In regulated industries, the receipt may be what lets the system go live at all. KPMG’s Rob Fisher said at launch that “greater autonomy is often only possible when built upon a solid foundation of trust, accountability and proper controls.”

 

The autonomy setting sits on top of all this. A team can start with simple assistance and move toward governed full-auto execution, one outcome at a time, within the authority it has delegated.

​

How fast should a finance team move, and when will ROI show up?

​

Ng’s advice sounds familiar: crawl, walk, run. What’s useful is where he places each finance function on that scale. Some work is low risk to turn on. Liability substantiation needs little change management, perhaps one short SOP to set the output format. That kind of win builds trust. Running day minus two of the close at 80% autonomy is a much bigger change management project.

Payables is where he’d go furthest:

 

“Why do you actually have 200 people doing data entry? It should be completely touchless. If I were running my own company, I would make it 100% machine owned, from documents coming in to getting the payments almost ready to be paid.”

He also points out that the offshore AP team’s cost goes beyond wages. It includes management overhead and operational complexity. For accounting judgment calls, he expects slower adoption. Every function will move at its own pace, depending on comfort, change readiness and company culture.

​

On ROI, he was direct. Do customers have a lot of proven cases for Claw-powered agentic apps? “Not quite.” Can they generate ROI yet? “No, not yet.” Will they in the next three to six months? “Absolutely.” Analysts are waiting for the same evidence. At Oracle’s April analyst summit, Forrester noted that Oracle cited more than 1,000 prebuilt agents but no production deployment counts or outcome metrics.

​

Ng also repeated a remark he heard all week: AI tools mostly mean people get more done and are expected to do more. Survey data backs that up. In Hubstaff’s October 2026 survey of 1,068 U.S. workers, 46% said expected workload had risen since AI arrived, and 5% said it had fallen. Hubstaff CEO Jared Brown said: “AI is giving time back to individual employees, but most organizations aren’t capturing it as capacity.”

​

That’s the strongest argument for Oracle’s approach. When assistants save individuals time, the savings scatter across the organization. When the work itself moves from people to a system, the capacity becomes something a CFO can count.

​

Is the payoff lower costs or faster growth?

​

Back-office automation usually gets justified as cost savings. Ng would rather measure how quickly a company can act on a new opportunity. Suppose you could onboard 15 new suppliers at a tenth of the old cost. Or suppose you could open a business unit in a country where you’ve never operated and support it with your current finance team.

​

That second scenario has a product behind it. At AI World, Oracle announced AI Success Navigator, built on the Fusion Setup and Configuration Registry. The registry packages Oracle’s setup knowledge as skills that agents use to design, configure, test and validate a deployment. Early adopter Lexington Health reports cutting certain design and configuration work by 80%. That figure comes from the customer, and no one has independently measured it.

​

Boards want the growth story. A Gartner survey of 204 finance leaders, published in July, found 45% of finance AI investment leaning toward productivity and only 20% toward decision quality, while boards weight growth and better decisions more heavily. Functions that funded AI to create new offerings or enter new markets were more than twice as likely to report high value. McKinsey’s 2025 State of AI shows the same pattern from a different angle. Only 39% of companies report any EBIT impact from AI, and most of those put it under 5%. The roughly 6% of high performers are nearly three times as likely to have fundamentally redesigned their workflows. That’s the same thesis Ng describes: change how the work is done, rather than adding AI to the old process.

​

Who knows how to fly this plane? The 12-24 month outlook

​

Ng isn’t worried about the technology. He’s worried about expertise, and he described the problem with an image he uses with his own engineers:

​

“You guys are the best engineering teams to build a plane, but none of you have been pilots. Now I’m telling you to go out there and tell people how to be the best pilot.”

​

No one has years of experience running a finance department with autonomous systems. To build that knowledge, Ng is placing development managers and product managers inside about 100 of Oracle’s roughly 12,000 Fusion customers. The goal is to create reference cases. He says his best customers keep Oracle expertise in-house and keep pace with each release.

​

From this conversation and the AI World announcements, I expect five things.

1. Published evidence will decide deals by mid-2027. Those 100 embedded customers will produce Oracle’s first public results. Buyers will start asking vendors for exception rates and human minutes per outcome instead of agent counts. HyperFRAME has already requested those numbers.

2. Touchless AP will become the expected standard. AP automation is already the second most common finance AI use case, at 37% of adopters in Gartner’s 2025 survey. The benchmark that matters will move from cost per invoice to the share of invoices that need any human handling.

3. Pricing disputes will move to renewals. Oracle’s AI units, and its stated aim to tie outcomes to consumption, haven’t been tested at contract renewal. Forrester noted that Oracle hasn’t committed to a consumption model for agent-heavy workloads at renewal.

4. Writing the envelope will become a core finance skill. Turning policy into enforceable instructions for the system will become as important to the controller’s team as the month-end close checklist.

5. The gap between Fusion customers and legacy-ERP customers will widen. Ng told Forrester in April, “These agents are not coming to [EBS].” Companies still on E-Business Suite, JD Edwards or PeopleSoft will fall further behind on automation every quarter.

​

The competition is real. SAP, Salesforce and Workday are investing heavily in the same direction, and Futurum puts Oracle second in ERP with an 11.0% share and $6.6 billion in 2025 revenue. Oracle’s advantage is in how the pieces work together: planning and execution are separated, governance is written in plain language, and every action produces an audit receipt, all inside the system that already holds the ledger.

​

What's your edge?

​

For CEOs: 

• Ask for a growth case alongside the cost case. Gartner’s July 2026 data shows boards value growth and better decisions, and that AI aimed at new offerings or markets is more than twice as likely to deliver high value. Ask your team what entering a new country or acquiring a company would cost if Oracle’s setup and configuration skills handled the build and your current finance team handled the run.

• Name one function to make machine-owned, with a date. AP is the obvious choice. Set a target for the share of invoices that need no human touch and assign one executive to own it.

• Decide where freed capacity goes before you free it. In Hubstaff’s survey, 46% of workers said AI had increased their workload. If you don’t plan to redeploy, reskill or reduce, the savings disappear into longer to-do lists

​

For CFOs: 

• Write the envelope first. Gather your desk manuals, approval matrices, AP hold codes and delegation-of-authority limits. They become the Enterprise Operating Envelope. Bring your external auditor in early so the Outcome Receipt is accepted as control evidence before go-live, not reviewed afterward.

• Start with substantiation, then move to the close. Liability and accrual variance analysis needs little change management and shows results within a quarter. After that, aim for a measurable share of the day-minus-two close running autonomously.

• Limit pilots to three to five use cases. Gartner’s Ash Mehta: “If everything is a priority, nothing gets funded.” Only 7% of finance organizations in Gartner’s research report high impact from AI. Spreading effort too thin is a big part of the reason.

​

For CIOs:​

• Fix master data before raising autonomy. Finance leaders cite data quality and data literacy as their biggest AI obstacles (Gartner). A collections agent working from bad customer records will chase the wrong invoices quickly.

• Plan a path off legacy ERP. Oracle’s agents run only in Fusion SaaS. If you’re on EBS, JD Edwards or PeopleSoft, Forrester recommends negotiating a commercial bridge to Fusion AI within 12 to 18 months.

• Set an AI front-end policy. Fusion Companion lets employees act on Fusion data from tools like ChatGPT, within existing permissions. Decide which front ends you’ll allow before employees choose for you.