Oracle AI World in Las Vegas made one thing clear: AI in Fusion has moved past the demo stage. Oracle reported 3,648 agents in production and 7,000+ customers using AI and agents. I spent the week hearing from 12 Oracle customers about what’s actually working inside their organizations, and their lessons are a practical playbook for anyone ready to scale.

THE NEWS / NUMBERS
Oracle came in with numbers: 5.7M agent runs, 3,648 agents in production, 7,000+ customers using AI and agents, and 100K+ certified builders. Oracle also announced a native mobile app for frontline workers and a new way to implement Fusion (more on that below.) On day 2, Oracle co-CEO Clay Magouyrk sat down with OpenAI CEO Sam Altman. Clay shared that every Oracle employee has a ChatGPT license, and 70% of Oracle’s software is now developed by AI.

FUSION CLAW
Fusion Claw is the engine behind Oracle’s Fusion Agentic Applications. It moves agents from answering questions to actually finishing complex, multi-step work. A frontier model (Gemini or OpenAI) does the reasoning and planning, and Fusion’s own deterministic computation handles the high-volume execution. Every completed job produces an Outcome Receipt that shows the authority the agent had, the evidence it used, and the decisions and transactions it made. For HR, Workforce Staffing is the one to watch. Scheduling is a beast.

SECURITY & GOVERNANCE
Sam’s #1 recommendation for what to do with AI right now: “find security vulnerabilities and mitigate them.” That theme came up in almost every customer conversation too. Because Oracle’s AI lives inside Fusion, the agents inherit Fusion’s security. Spectrum’s agents use the same security profile as the UI, so if you can’t see it in Fusion, the agent can’t show it to you. Heathrow’s analytics inherit the same model. With Fusion Claw, each organization sets its own policies, permissions, approval limits and how much autonomy each process gets.

Customers are pairing that with their own guardrails. Spectrum’s agents can’t hire, promote or terminate on their own, and once those rules were set, the team could move fast. Nationwide’s mantra is start small and build trust. Heathrow is going “slowly and measured,” but fast enough to keep up.

AI LESSONS LEARNED: ADVICE FROM CUSTOMERS

Keep AI inside the system. Baylor’s advice is to use the AI built into Fusion. For performance reviews, it draws on prior documents and comments. Paste that into an outside chatbot and the history (and the security) stays behind.

Start where the volume is. Spectrum’s lesson: aim the first agent at the biggest pile of work. For them, that was 80-90K HR tickets a year across 45 states. Northwell did the same and took a 20-person, weekday-only HR service center to 24/7 support for 100K+ employees.

Put agents on payroll checks. Spectrum built a payroll anomaly dashboard in 2 days. Northwell flags over- and underpayments during close. Their advice is to keep a human making every call.

Predict future issues and act upon that knowledge. Heathrow uses Fusion Data Intelligence to predict attrition from a decade of history, with salary, leave and transfer patterns as the key signals. Their lesson is that a chart showing what changed isn’t enough. Leaders need dashboards that tell them what to act on.

Reskill, don’t reduce. Bank of Hawaii’s advice is to shift people toward human-to-human customer work. The UK MoD learned through skills assessments that some of its strongest future cyber candidates are chefs. Yes, chefs.

THE FUSION SETUP & CONFIGURATION AGENT
Oracle’s new Fusion Setup & Configuration agent puts implementation know-how into a public Git repo that AI agents can execute. They plan, configure, test and validate, and people approve before anything goes live.

Every vendor is talking about AI that implements software right now. What I haven’t seen yet is customer proof of faster go-lives or lower cost. Conversely, faster configuration won’t fix change management or adoption. That’s where good Consulting partners still earn their keep.

If you’ve gone live on Fusion recently, I want to hear how it went. Share your review here: http://ravenintel.com/review