Oracle Built the System of Execution. Nobody Asked What It's Executing On.
Oracle's 22+ Fusion Agentic Applications execute business decisions autonomously. The transactional layer is covered. The document knowledge layer isn't.
Oracle announced Fusion Agentic Applications yesterday at a London event — 22+ AI agents embedded across Finance, HCM, Supply Chain, and Customer Experience. These don't advise. They execute. Cash collection actions. Workforce scheduling decisions. Vendor sourcing. Cross-sell program management. Oracle EVP Chris Leone put it plainly: "AI is moving from advisers and copilots to being able to execute work." Workday ran the same play six days earlier with 300+ agentic skills. The coverage was wall-to-wall. One question nobody asked.
The transactional layer is covered. The document knowledge layer is not.
Oracle's claim that these agents are "grounded in systems of record" is accurate. The transactional layer — invoices, schedules, approval hierarchies, purchase orders — is structured, maintained, and current by design. That's what "system of record" means, and Oracle's native integration with that layer is a real engineering achievement. EVP Steve Miranda described the goal: moving "beyond passive systems of record" toward applications that "reason, decide and act in pursuit of defined business objectives." Correct direction.
The gap is that enterprise operations don't run on transactional data alone. They run on policies. Credit terms. Labor agreements. Vendor qualification frameworks. Compliance procedures. Regulatory interpretations. These documents tell the transactional system how to execute — and they live in SharePoint folders, shared drives, and PDF repositories that nobody actively governs. Oracle's tight integration with the transactional layer doesn't extend to that layer. Nothing does, by default.
The failure modes are domain-specific
Finance: cash collections on superseded credit policy
Oracle's finance agent handles "automated cash collections risk analysis." Risk analysis against what, exactly? If the credit terms document was last updated 18 months ago and the board revised credit limits in Q4, the agent executes collection actions based on superseded policy. This isn't a recommendation a human reviews before acting on it. It's an execution.
HCM: workforce scheduling against outdated labor compliance
The HCM agent does "workforce scheduling with real-time gap detection." Gap detection relative to what compliance baseline? State labor regulations change. Union agreements get renegotiated. If those updates haven't reached the documents the agent reads, the gaps it detects are gaps against an outdated framework. Legal exposure accumulates quietly until the next audit traces it back.
Supply chain: sourcing against an expired approved-vendor list
Oracle's supply chain agent handles "AI-driven sourcing from design data." Which vendors are currently approved? If the approved-vendor list hasn't been refreshed since last quarter's review, the agent routes sourcing to vendors whose status has since expired, been suspended, or flagged for quality issues. The transactional system knows the vendor ID. It doesn't know what happened to that vendor's approval last month.
The compounding problem
A human working from a stale policy document makes one mistake. An AI agent working from the same document makes that mistake at the volume and velocity of every transaction it touches — until an auditor pulls a sample, traces the decision chain, and finds the source. At enterprise scale, one outdated procedure document doesn't produce one error. It produces a systematic error embedded across every execution that referenced it.
Workday's Aashna Kircher cited "deterministic rails" as the safety mechanism in their agentic deployment. Deterministic rails means the transactional data structure. The same gap applies there too — as agentic deployment scales, document knowledge becomes the failure surface nobody was watching.
What the document knowledge layer needs before agents touch it
Oracle's framing is right in principle: agents need to be grounded in enterprise knowledge to operate safely. The "system of record" concept just needs to extend from the transactional layer to the document-knowledge layer that governs how transactions should actually be executed.
That layer has to meet a specific standard before it's ready for autonomous execution. It has to be current — reflecting what's actually true today, not what was accurate when someone last saved the file. It has to be contradiction-free, meaning two policies on the same topic agree, or the conflict is surfaced and resolved before any agent can act on either. Every agent action has to be traceable back to the specific document version that informed it. And policy updates need to propagate immediately, with a timestamped audit trail so changes are accountable.
This is the infrastructure problem the announcement didn't address. Transactional grounding: Oracle's responsibility, and they've handled it. Document knowledge grounding: the enterprise's responsibility, and most haven't treated it that way. Mojar AI is built specifically for this layer — RAG infrastructure that actively manages documents, detects contradictions across a knowledge base, attributes sources to every answer, and ensures changes are reflected immediately.
The transactional grounding problem belongs to Oracle. The document knowledge grounding problem is still open.
The window to prepare is already narrowing
Oracle, Workday, SAP, Microsoft — every major enterprise software vendor is moving the same direction, and they're moving fast. The agentic enterprise era isn't approaching; it's deploying. By the time agents are live at scale, the window to prepare the knowledge layer that governs their decisions has closed.
The question enterprise operations teams need to answer now — before the Oracle rollout, before the Workday deployment, before any agentic application touches a consequential workflow — is whether the policies, procedures, and compliance documents these agents will execute on are current, internally consistent, and traceable. Because once execution starts, there's no human in the loop catching the difference between a 2024 credit policy and what the board actually decided in Q4 2025.