Three separate announcements this month show a pattern: the major enterprise platform vendors are pushing AI deeper into their existing systems rather than asking customers to bolt on external agents. Oracle added an AI-native builder to Fusion Applications. Google renamed NotebookLM to Gemini Notebook and gave it code execution inside a secure cloud environment. Microsoft shipped Copilot features in SharePoint that generate Office documents, discover reusable skills, and scaffold automations.
Taken together, these releases suggest that 2026 is the year when "agentic" stopped being a buzzword in vendor slide decks and started appearing in actual production tools. The approaches differ in architecture and audience, but they share one assumption: enterprise AI works best when it lives inside the systems where data and governance already exist.
Oracle's Fusion Agentic Applications: agents that inherit security by default
Oracle's update to AI Agent Studio for Fusion Applications lets customers build and run AI agents directly inside Fusion Cloud Applications. The company is calling these "Fusion Agentic Applications" and describes them as a new class of enterprise software where multiple specialized agents collaborate to accomplish business objectives. The agents perform reasoning and decision-making before executing business processes using Fusion's existing business objects, workflows, approval frameworks, and policies.
The key architectural claim is that these applications inherit Fusion's security and governance policies automatically. Audit trails are built in. Oracle argues that when AI systems are built outside enterprise applications, organizations have to separately address identity management, data access, approvals, audit trails, observability, and lifecycle management. Their approach puts all of that inside the runtime environment from the start.
The builder environment supports no-code, low-code, and pro-code development within a single Fusion-native framework. Business users can create agentic applications by describing what they want in natural language. Developers can use Visual Studio Code, CLI tools, Git-based environments, OpenAI Codex, or Anthropic's Claude Code through the new AI Studio Skill.
Oracle says AI Agent Studio for Fusion Applications is available at no additional cost to existing Fusion Applications customers and partners. That pricing model matters: it lowers the barrier to experimenting with agentic workflows because there is no separate per-seat or per-agent license to evaluate.
I am skeptical about how much natural-language-to-agent actually works for complex business processes. Oracle's demos show simple approval chains and data lookups. The real test will be whether a procurement manager can describe a multi-step approval with conditional routing and get something that does not break on edge cases. The architecture is sound, but the interface abstraction is hard to get right.
Google's Gemini Notebook: code execution inside a research tool
Google renamed NotebookLM to Gemini Notebook and added native code execution through a secure cloud computer. The tool can now generate and run code using information from a user's uploaded sources. Google says the feature supports complex data analysis grounded in source documents, as well as new output formats.
The secure cloud computer is introduced in stages. It is available first to Google AI Ultra users and Workspace business customers with AI Ultra or AI Expanded Access. Google says it will reach all Pro users on the web over the coming weeks.
Gemini Notebook remains a standalone product despite the name change and integration with the Gemini app. Users can create and access notebooks inside the Gemini app, with changes synchronized between the app and the standalone product. Business use cases identified by Google include analyzing financial documents, creating onboarding resources from training manuals, and preparing meeting plans using product specifications.
The code execution capability is the most interesting part. It allows the tool to move from passive summarization to active computation. A user could upload a CSV of sales data and ask Gemini Notebook to calculate monthly averages, build a chart, or run a regression. The code runs in a secure environment and the results stay grounded in the uploaded sources.
This is a different model from Oracle's. Oracle embeds agents inside the business application. Google gives you a standalone tool that can pull from your uploaded sources. Neither is inherently better, but they serve different workflows. Oracle's approach makes sense for operational processes like purchase order approvals. Gemini Notebook works for research and analysis tasks where the user controls the inputs.
Microsoft's SharePoint Copilot: generating Office files from existing content
Microsoft's July 2026 update to Copilot in SharePoint focuses on reducing repetitive work. Users can ask Copilot to create Word documents, Excel workbooks, and PowerPoint presentations directly from SharePoint content. Instead of starting with an empty document, the system generates complete Office files based on existing SharePoint pages, lists, or documents.
The update also introduces a /Skills command that lets users discover reusable automations already created by their organization. Built-in Skills provide scaffolding and starter templates for common business processes like site setup. Microsoft is trying to solve the problem of people rebuilding the same workflows across departments.
These capabilities follow earlier Copilot features in SharePoint for site creation, interactive reporting, and workflow automation. The July update feels like a consolidation play: take the individual AI features and connect them into a coherent content generation pipeline.
The practical impact depends on how well the generated documents match what a human would produce. Generating an Excel workbook from a SharePoint list is straightforward if the schema maps cleanly. Generating a PowerPoint presentation with proper narrative flow is harder. Microsoft will need to iterate on the quality of the output.
Comparing the three approaches
Oracle, Google, and Microsoft are all building AI into platforms, but they start from different foundations.
Oracle starts from the business process layer. Fusion Agentic Applications execute inside the same system where ERP, HCM, SCM, CRM, and EPM run. The agents have direct access to business objects, approval workflows, and audit trails. The tradeoff is that you must be a Fusion Applications customer to use any of it.
Google starts from the research and analysis layer. Gemini Notebook is a standalone tool that connects to user-uploaded sources. It has more flexibility in what data it can consume (PDFs, videos, audio, web pages) but it does not integrate with enterprise business processes natively. The code execution feature is a capability that Oracle and Microsoft do not directly offer in their announcements.
Microsoft starts from the content layer. SharePoint is a document management and collaboration platform, not a transaction processing system. Copilot generates Office files and discovers skills, but it does not execute business workflows the way Oracle's agents do. Microsoft's strength is the integration with the rest of Microsoft 365.
For a technical audience, the interesting question is how these platforms handle the hard problems: state management, error recovery, and auditing. Oracle is explicit about audit trails and governance inheritance. Google does not publish details about the secure cloud computer's execution environment. Microsoft relies on existing Microsoft 365 compliance and retention policies.
Practical takeaways
If you run Oracle Fusion Applications, the new AI Agent Studio is worth testing because it comes free. Start with a simple approval workflow where an agent checks an invoice against contract terms and either approves or escalates. You get governance and audit trails for free, so the risk of deploying a rogue agent is lower than with a custom-built solution.
If your team does research or data analysis, Gemini Notebook with code execution is worth trying for exploratory work. Upload a dataset, ask questions, and see whether the generated code is correct. The main limitation is that the code execution environment is not transparent about what packages or libraries are available. Expect to validate the output manually for now.
If you live inside Microsoft 365, the SharePoint Copilot features remove some friction from content creation. The /Skills discovery feature is the most valuable piece if your organization has invested in Power Automate or other automation tools. The document generation feature saves time on routine reports and onboarding materials, but do not trust it for anything that requires precise formatting.
The broader pattern is that enterprise AI is moving from standalone copilots to embedded agents. The vendors who own the business logic and the data are in a strong position. Independent AI tooling will need to find its niche in areas the platforms do not cover, such as multi-platform orchestration or highly specialized domain models.