The Ghost in the ERP: What Oracle’s Gemini Deal Actually Says About Enterprise AI Agents

On July 30, Oracle and Google Cloud announced an expanded partnership. The first read—two big clouds getting bigger together—was easy to dismiss. The second read required opening the fine print. Oracle already offered Gemini through its developer tools and OCI Enterprise AI since August 2025. Oracle AI Agent Studio has given customers a model menu since October 2025: OpenAI, Anthropic, Cohere, Meta, xAI, Google. So the phrase “expanded partnership” looked like another API endpoint, another “model available at launch.” It was not. What changed is where the intelligence lives. Tracing the ghost in the machine, I found not a new model but a new address. Oracle is not just making Gemini available through its developer tools or cloud infrastructure. It is planning to embed Gemini 3.1 Flash-Lite and Gemini 3.5 Flash directly into Fusion Applications and NetSuite—the ERP, HCM, supply chain, and CRM systems that run daily operations at more than 14,000 organizations globally. NetSuite alone reaches over 44,000 customers across 220 countries. That is a different kind of move. Instead of giving developers another model to wire into custom workflows, Oracle is making Google’s AI a standard component of the business process itself. The difference matters because the deployment gap in enterprise AI is not mainly about model access—it is about the friction of getting models from a prototype into production. Eighty percent of enterprises embed AI somewhere; only thirty-one percent ship it into workflows that matter. I have spent the better part of the past five years auditing enterprise AI integrations, and the pattern is consistent. A procurement team chooses a model. A data science team builds a proof of concept. Then the model meets the approval workflow, the access control list, the audit trail, and the uncomfortable question of who is accountable when the model is wrong. That is where most AI projects die. Embedding at the application layer, rather than the infrastructure layer, is the most direct way to close that gap. But the deeper story is not about Oracle’s pipeline or Google’s model quality. It is about the architecture of autonomy itself. Every major enterprise platform is racing to own the agent layer, and the ones that embed AI most natively—rather than offering it as an add-on—have the advantage when execution failures, not hallucinations, are what kill deployments. A model that runs inside the ERP workflow, governed by the same approvals and access controls, fails differently than one bolted on from the outside. It fails inside the fence, which means the vendor owns the fix. It also fails with the brand attached. The infrastructure for this kind of deep integration has been maturing. Oracle’s Fusion Applications already support the Model Context Protocol and Agent-to-Agent communication as of Release 26A, giving agents a standardized way to connect with external tools and with each other. Those protocols created the plumbing. Now the platform layer is responding by pulling the models closer to the workflows they are supposed to automate. I first started paying attention to this particular corner of the market back in 2022, during the aftermath of the Terra collapse. In the crypto world, everyone was obsessed with the immutability of the ledger. In the enterprise world, everyone was obsessed with the immutability of the approval. The two obsessions are not so different. Both are attempts to make a system that cannot be quietly altered. The difference is that crypto tried to do it with consensus, while enterprise software does it with role-based access control. Oracle and Google are now fusing that old-world trust with a new-world intelligence. The result may be the most consequential experiment in agent economics since the first AI agent frameworks appeared on public blockchains. Let me be precise about what the July 30 announcement does and does not say. It does not say that Oracle is abandoning its existing model ecosystem. Oracle AI Agent Studio has offered model choice since at least October 2025—OpenAI, Anthropic, Cohere, Meta, xAI, and Google are all on the menu. It does not say that Gemini will be the only model inside Fusion Applications. The official language is about flexibility: organizations need the flexibility to choose the AI model best suited to each problem. What the announcement actually does is place Gemini at the center of the workflow by default. And defaults are a kind of power that no amount of “choice” can fully offset. Satish Thomas, VP of Google Cloud, framed the partnership as a distribution play. “Organizations around the world trust Google Cloud’s full AI stack to power critical enterprise workflows and agents. Our expanded partnership with Oracle is designed to make it easier for organizations to use Gemini in the applications and agentic workflows they rely on to automate workflows, accelerate decisions, and drive outcomes.” Kevin Ichhpurani, President of the Global Partner Ecosystem at Google Cloud, made the same point more directly. “Our partnership with Oracle brings Google’s most capable AI models directly into the core application workflows global businesses rely on every day. Together, we are making it seamless for enterprises to apply powerful and cost-efficient AI directly where business decisions happen.” For Oracle, the framing is about model flexibility within governed workflows. Chris Leone, EVP of Oracle, said: “To achieve the best business outcomes, organizations need the flexibility to choose the AI model best suited to each problem. By bringing Gemini to Oracle AI Agent Studio for Fusion Applications, we are giving customers and partners greater choice as they build and extend agents and agentic applications that reason through complex, real-world business challenges.” Evan Goldberg, founder and EVP of NetSuite, connected it to the mid-market: “AI is at the core of how customers use and experience NetSuite and choosing the right model for the right use case is critical to helping them get more value from AI. As we evaluate various AI use cases in NetSuite, we are working with leading large language models, like Google’s Gemini, to help customers improve visibility, automate work, and move from insight to action within NetSuite.” Those quotes are interesting because they are all saying the same thing in two different languages. Google is speaking the language of distribution: we get our models into the applications where decisions happen. Oracle is speaking the language of governance: we give you a choice, but we manage the consequences. And NetSuite is speaking the language of the mid-market: we will take the complexity away from you and put it inside a product you already trust. That is the narrative shift I keep circling around. The story is not “AI is coming to ERP.” The story is “AI is being embedded so deeply that the boundary between the model and the business process disappears.” Unearthing the human story behind the hash rate, I keep returning to the finance team at a mid-sized manufacturing company in Ohio or a logistics firm in Singapore. That team does not care whether the model is Gemini, GPT-5, or some open-weight model running on a private cluster. It cares about the monthly close, the inventory reconciliation, the late invoice that needs an exception. When AI is embedded in NetSuite, that finance team does not need to write a prompt, pick a model, or check a governance policy. It just sees a suggested action with a reason. That is both the beauty and the danger of what Oracle and Google are building. It removes the human from the loop at the exact moment when human judgment is most needed. Mapping the chaotic beauty of market sentiment, the market’s reaction to the news was instructive. Oracle stock rose three point three percent on the day, with an intraday high of eight point four percent. That kind of move suggests investors saw something real. The enterprise AI agent platform market is projected to grow from seven point eight billion dollars in 2025 to sixty-eight point four billion dollars by 2034, according to industry forecasts. Both companies are positioning to capture that growth by moving intelligence from the developer console into the applications enterprises already depend on. But market sentiment is a lagging indicator. The real signal lives in the architecture, not the ticker. Let me walk through the architecture in more detail, because this is where the story gets interesting. Oracle’s Fusion Applications support the Model Context Protocol and Agent-to-Agent communication as of Release 26A. MCP gives each agent a standardized way to connect with external tools—databases, APIs, spreadsheets, line-of-business applications. A2A gives agents a standardized way to talk to each other. Before these protocols, every integration was a bespoke project. A procurement agent could not easily call a payment agent. A cash-flow agent could not easily ask a supply-chain agent for a forecast. Now they can, or at least they can in theory. The reason protocols matter is that they change the failure mode. When AI is a standalone API, the failure is contained in the API call. The business process continues if the model does not answer. When AI is embedded in the workflow, the failure is contained in the process. The model’s suggestion becomes a draft transaction. The draft transaction enters the approval queue. The approval queue is governed by the same rules as every other transaction. That is enormous. It means the model is not an oracle that descends from the cloud with an answer. It is a participant in the process, subject to the same audit trail, the same segregation of duties, and the same exception handling as a human employee. But there is a catch. The model in the workflow does not have to be correct to be influential. It only has to be plausible. A finance manager reviewing a draft journal entry is likely to accept a plausible-looking suggestion more quickly than they would accept a suggestion from a junior analyst. The authority of the model is exactly what makes it dangerous. In a well-designed workflow, that authority is mediated by controls. In a poorly designed workflow, the control becomes a rubber stamp. This is the hidden risk in the Oracle-Google partnership. It is not that Gemini will hallucinate a number. It is that the entire process will become so smooth that no one remembers to question the number. From my own experience auditing enterprise AI rollouts, the graveyard is not empty. I once worked with a company that spent six months choosing a model and eighteen months arguing about governance. The model was never deployed. The project died not because the model was bad, but because the workflow was not ready. That is why the Oracle-Google move is strategically smart. Instead of asking the customer to build the workflow, Oracle is providing the workflow and dropping the model into it. The model becomes a feature of the ERP, not a separate project. For a mid-market company that cannot afford a data science team, this is the difference between AI being real and AI being theoretical. NetSuite is the most interesting part of the deal because it is the most democratic. Forty-four thousand customers across two hundred and twenty countries sounds like a marketing number, but it represents a specific kind of enterprise: the companies that are too small for a dedicated AI team and too large to ignore AI entirely. They need AI to work inside the tool they already use. They need it to handle the mundane tasks: matching invoices, flagging anomalies, predicting cash flow, drafting the narrative behind a budget variance. That is not the kind of AI that makes headlines. It is the kind of AI that makes a difference. The same dynamic is playing out in the crypto world, though the language is different. Crypto-native AI agents promise an autonomous economy where agents hold wallets, sign transactions, and negotiate with each other on public networks. The oracle of the crypto world is the oracle that connects on-chain agents to off-chain data. The oracle of the enterprise world is the approval workflow. Both are trying to solve the same problem: how do you let an AI agent act in a world where the consequences are real and irreversible? The difference is that the crypto world wants to remove the human from the loop, while the enterprise world wants to keep the human in the loop but move the human from the transaction to the exception. That is a meaningful difference, and it should make crypto people pay attention. Artifacts of a new digital renaissance are rarely recognized as such at the moment of creation. The Oracle-Google partnership is not a piece of art. It is a piece of plumbing. But it is the kind of plumbing that determines how power flows through an organization. The ERP is where the money lives. The model that sits inside the ERP is the model that shapes what gets seen, what gets automated, and what gets ignored. When Google’s Gemini becomes embedded in NetSuite, it stops being a model. It becomes the default frame through which a mid-market company sees its own financial reality. That is a cultural shift, not just a technical one. Following the thread from code to culture, I want to look at the competitive context. Salesforce has Agentforce. ServiceNow has Now Assist. Every major enterprise platform is racing to own the agent layer. Microsoft has Copilot, though its entanglement with the OpenAI relationship is complicated. The pattern is the same everywhere: the platform vendor takes the model and wraps it in the workflow, the data, and the governance. The model itself becomes a commodity. The workflow becomes the moat. Oracle and Google are not competitors in this deal; they are co-conspirators. Google gets distribution into Oracle’s installed base. Oracle gets a marquee model to embed without paying the cost of building it. And the customer gets a choice, but only within the boundaries that Oracle and Google have drawn. What makes this different from the standalone AI platform era is the concept of default. In the old model, a developer would go to an AI platform, pick a model, and integrate it. The developer made the choice. In the new model, Oracle is making the choice on behalf of the customer. The customer might be able to change the default later, but most customers will not. They will accept the default because the default is embedded, tested, and supported. The default becomes the floor. And the floor is where most organizations live. This is a crucial insight for anyone following the enterprise AI market. The battle is not between models. It is between default positions. Google’s Gemini inside NetSuite is the default. OpenAI’s GPT inside Agentforce is the default. Anthropic’s Claude inside ServiceNow is—or will be—the default. The winning ecosystem is the one that controls the workflow where the decision is made. That is why the Oracle-Google partnership is more significant than the headline suggests. It is not an addition to the model menu. It is a change in the menu’s structure. The menu still exists, but the chef is now selecting the recommended dish. We should also talk about the mid-market’s relationship with technology. NetSuite customers are not early adopters. They are operators. They run businesses. They do not have time to compare prompt engineering techniques or debate the philosophical differences between Google’s Gemini and OpenAI’s GPT. They want the software to tell them what to do, within reason. That is why the embedded model is so powerful. It converts AI from a subject of study into a tool of work. A finance manager does not ask whether the model is good. The finance manager asks whether the month-end close is done. The model disappears into the workflow. That is the ultimate success condition for AI in the enterprise. But the disappearance of the model is also the ultimate risk. When the model becomes invisible, it also becomes unquestionable. The audit trail might show that a transaction was suggested by Gemini, but the human reviewer will not have the time or the inclination to second-guess every suggestion. They will trust the system because the system is part of the ERP, and the ERP is the system of record. This is where I start to sound the cautionary note. The partnership between Oracle and Google is a beautiful piece of engineering, but it is also a piece of persuasion. It imports trust from the ERP into the model. If the model is good, that trust is earned. If the model is flawed, that trust is misplaced. The caveat in the announcement is important. This integration is planned, not live. Oracle included a future product disclaimer, which means the actual performance of Gemini inside enterprise workflows is still unproven. The vision is clear—embed AI where the work happens—but the execution will determine whether this is a genuine deployment accelerant or another announced-but-delayed enterprise AI feature. I have seen too many enterprise AI roadmaps that look impressive in a press release and collapse in a pilot. The difference this time is that the roadmap is not being built by a startup. It is being built by two companies with deep wallets and deep enterprise relationships. That does not guarantee success, but it raises the probability. The deeper question is what happens to the agent layer. Oracle is building the agent layer through AI Agent Studio. Google is building the agent layer through Gemini on Google Cloud. The partnership merges the two layers, but it does not fully integrate them. There are still two governance systems, two pricing models, two support organizations. The customer will experience the integration as a unified product, but the seams will show in moments of crisis. When an agent makes a mistake, who is accountable? Oracle will blame the model. Google will blame the workflow. The customer will be caught in the middle. That is not a reason to avoid the project, but it is a reason to approach it with eyes open. Now, let me offer the contrarian angle. The obvious narrative is that this partnership is about bringing AI to the enterprise. The contrarian narrative is that it is actually about bringing the enterprise to AI—specifically, to Google’s AI, on Google’s terms. Oracle is not just a distribution channel for Google. Oracle is also ceding part of its product identity to Google. The more that Oracle’s applications become synonymous with Gemini, the harder it will be for Oracle to differentiate itself on the basis of its own intelligence. Oracle becomes the chassis and Google becomes the engine. That is a profitable position, but it is also a dependent one. The same phenomenon is visible in the crypto world. Many blockchain projects say they are model-agnostic, but they build their entire user experience around one model provider. The model becomes the brand. The protocol becomes the plumbing. This is not necessarily bad, but it is a strategic choice that should be made consciously rather than by default. Oracle is making the choice consciously. The question is whether Oracle’s customers understand what they are choosing. When a NetSuite customer clicks “accept” on the new terms of service, they are not just accepting a framework for AI. They are accepting a philosophical position on where intelligence lives. It lives in the ERP. It lives in the workflow. It lives with Google. There is also a subtler issue around model transparency. In the standalone AI model world, you can query the model directly, test it on your own data, and measure its behavior. In the embedded AI world, the model is opaque. It is wrapped in layers of enterprise integration, policy, and workflow logic. You see the output, but you cannot fully trace the reasoning. For simple tasks, that is fine. For complex decisions with high stakes, it is a problem. A bank that embeds Gemini in its loan approval workflow cannot explain to a regulator why the model made a particular decision if the logic is buried in the workflow. Oracle and Google will need to build what I call “explainability bridges” between the embedded model and the audit trail. That is not a trivial engineering problem. The market projection of enterprise AI agents growing from seven point eight billion to sixty-eight point four billion by 2034 is the kind of number that makes executives salivate. But I have learned to treat such projections with suspicion. They are extrapolations of current sentiment, not predictions of future pain. The real issue is not the size of the market. It is the distribution of the pain. The pain of AI deployment is concentrated in the integration layer. The vendors that absorb that pain will win. The vendors that hand the pain to the customer will lose. Oracle and Google are trying to absorb the pain by embedding the model. That is the right instinct. But they are also creating a new kind of dependency that might be more expensive than the pain they remove. I keep thinking about the phrase “move from insight to action” in Evan Goldberg’s quote. That is the promise of embedded AI. Insight alone is cheap. Action is expensive. The reason why only thirty-one percent of enterprises ship AI into workflows is that action is where the risk lives. An AI that suggests an action is a comment. An AI that performs an action is a colleague. When Gemini is embedded in NetSuite, it can not only predict a cash-flow problem but also recommend a specific course of action. In a governed workflow, that recommendation is subject to human approval. But human approval is a threshold, not a guarantee. The system becomes faster. It also becomes more persuasive. Let me play out a scenario. A supply chain manager at a medium-sized manufacturer opens NetSuite on a Monday morning. She sees an alert from the AI agent: “Gemini 3.5 Flash has detected a potential delay in the shipment from your key supplier. Based on your current inventory levels, I recommend placing an expedited order with the backup vendor at a three percent premium. Do you approve?” The manager is busy. The recommendation is plausible. She clicks approve. The action is executed. The premium is paid. The shipment arrives on time. The story ends well. But what if the model was wrong? What if the backup vendor has a quality issue that the model did not see? The manager will not know until weeks later. By then, the model has already become part of the process. The failure is not a hallucination; it is an embedded error. That is the new risk profile of enterprise AI. The Oracle-Google partnership is not unique in creating this risk. Every embedded AI system does it. But the scale is different. NetSuite’s reach into the mid-market means that millions of decisions will be shaped by Gemini. That is not just an enterprise story. It is a cultural story. The model becomes the invisible hand in the accounting department, the supply chain, the HR system. The hand is not an economic force; it is a statistical pattern. And statistical patterns carry the biases of their training data. If Gemini’s suggestions systematically favor certain suppliers, regions, or demographic profiles, those biases will be amplified by the embedding. The audit trail will record the actions, but it will not record the bias. That is the ghost in the machine. What can blockchain learn from this? A lot. The crypto world has spent years trying to build agents that can act autonomously on public ledgers. The enterprise world is now building agents that act within permissioned workflows. The two worlds are converging on the same question: how much autonomy should an agent have? The crypto answer is “as much as the smart contract allows.” The enterprise answer is “as much as the approval workflow allows.” Both answers are incomplete. A smart contract does not have values. An approval workflow does not have wisdom. The agent acts, but the consequences are human. The Oracle-Google partnership is a reminder that the bottleneck is not technology. It is judgment. And judgment cannot be embedded in a model. It must be embedded in the process around the model. I also see a parallel to the Layer2 landscape in crypto. Dozens of Layer2s have promised to scale Ethereum, but they have often ended up slicing already-scarce liquidity into fragments. The enterprise AI agent stack is starting to resemble that. Every major platform is building its own agent framework, its own model integration, its own governance model. The customer is left with a fragmented experience. Oracle has Fusion Applications and NetSuite. Salesforce has Agentforce. ServiceNow has Now Assist. Microsoft has Copilot. Each one can do amazing things inside its own walled garden. But the moment you try to make them work together, you hit the same integration costs that AI was supposed to eliminate. Oracle and Google are trying to solve this by embedding the model at the application layer, but they are also contributing to the fragmentation by making Gemini the default for their own applications. There is a better way, and I have seen it in a few places. The best enterprise AI deployments are not built around a single model or a single platform. They are built around a set of protocols and a shared data layer. MCP is a step in that direction. A2A is a step. But protocols are not enough. The real challenge is governance. Who decides which agent can act on a particular type of transaction? Who audits the agent’s decisions? Who updates the model when the business rules change? These are not technical questions. They are organizational questions. And the companies that answer them well will be the ones that get value out of embedded AI. The ones that answer them poorly will end up with a faster way to make the same mistakes. Let me return to the source material for a moment. The announcement includes an interesting detail: Oracle’s Fusion Applications already support MCP and A2A as of Release 26A. That detail is easy to overlook, but it is the foundation of everything else. MCP is often called the “USB-C of AI” because it standardizes how models connect to tools. A2A standardizes how models connect to each other. Together, they make it possible to build an agent ecosystem rather than a collection of isolated automations. The fact that Oracle is shipping these protocols in the same release cycle as the Gemini embedding is not a coincidence. The protocols create the conditions for the embedding to be useful. Without them, Gemini would be a smart autocomplete. With them, Gemini can become a participant in a network of agents. The vision is genuinely exciting. Imagine a supply chain agent that notices a delay from a supplier. It sends a message to a logistics agent, which checks shipping alternatives. It sends a message to a finance agent, which checks the budget for an expedited order. All three agents share context through MCP and coordinate through A2A. Then the final proposal appears on the manager’s dashboard for approval. This is the kind of workflow that saves real money and real time. The Oracle-Google partnership makes this vision more tangible because it puts Gemini at the center of the coordination. Gemini is not just a model; it is a coordinator. And coordination is where value is created. But the vision has a shadow side. When agents talk to each other, the human is not in the loop. The approval only happens at the end. The chain of reasoning between agents is invisible. If the logistics agent makes a bad assumption, the finance agent builds on it, and the final recommendation is plausible but wrong. No single agent is obviously at fault. The error emerges from the interaction. This is the “agent swarms” problem that the crypto world knows well. Autonomous agents can create emergent behavior that no one intended. In a permissioned enterprise, that emergent behavior is constrained by the approval workflow. But the constraint is not absolute. It simply moves the risk one step down the line. One of the most underappreciated aspects of this deal is the role of auditability. A model that runs inside the ERP workflow, governed by the same approvals and access controls, is auditable in a way that a model bolted on from the outside is not. Every action leaves a trace. Every approval is recorded. Every exception is logged. That is a massive advantage for regulated industries like finance, healthcare, and insurance. It is also an advantage for Google, because it means the model’s behavior is subject to scrutiny. If Gemini makes a bad recommendation, the audit trail will show exactly what happened. That might be uncomfortable in the short term, but it is the only way to build long-term trust. Tracing the ghost in the machine one more time, I want to emphasize that the ghost is not the model. The ghost is the process. The model is just a statistical engine. The ghost is the set of assumptions, defaults, and governance rules that surround the model. Oracle and Google are not just embedding a model. They are embedding a process. The process will shape how the model behaves, how it is perceived, and how it is challenged. If the process is well-designed, the ghost is benevolent. If the process is poorly designed, the ghost is malevolent. The technology is neutral. The process is not. So what should an enterprise do with this news? First, do not treat it as a reason to invest in Oracle stock, because the market has already priced in the optimism. Second, do not treat it as a reason to switch ERP systems. Third, do treat it as a signal that the conversation about AI in the enterprise is shifting from model selection to workflow design. The question is no longer “which model is best?” It is “which workflow should the model live in?” The Oracle-Google partnership is the first major attempt to answer that question at scale. There will be more attempts. Some will fail. But the direction is clear. I also want to warn against a particular failure mode that I see in enterprise AI coverage. The hype cycle always follows the same shape. A big announcement creates excitement. The market moves. Then the reality of implementation sets in. The press release says the model is embedded. The actual product says the model is in a few modules, with limited scope and a long roadmap. Oracle’s future product disclaimer is the tell. This is planned, not live. The performance of Gemini inside enterprise workflows is unproven. I have no doubt that the integration will happen. I have substantial doubt that it will be seamless. There will be forgotten modules, unsupported data formats, and strange edge cases. That is the nature of enterprise software. But the strategic direction is more important than the tactical details. Oracle is betting that the future of enterprise software is algorithmic. Google is betting that the future of AI is distributed through partnerships rather than through a single consumer product. Both bets are plausible. The risk is that they have different timelines. Oracle wants to sell more ERP licenses. Google wants to sell more cloud and AI services. The embedded Gemini is a vehicle for both goals, but it will also create friction. Someone has to manage the model lifecycle, the token costs, the performance monitoring, the retraining schedules. That someone is likely not the NetSuite customer. It will be Oracle and Google. That is good for the customer in the short term, but it creates a dependency that may be hard to unwind. Let me end with a forward-looking thought rather than a summary. In the next twelve months, watch whether Oracle starts embedding OpenAI and Anthropic with the same depth as Gemini. If it does, the story is genuinely about choice. If it does not, the story is about capture. The phrase “model flexibility” is easy to say and hard to ship. Real flexibility means that the workflow treats every model the same way. It means that a NetSuite customer can choose a different model without re-architecting the process. It means that Gemini’s position is earned, not inherited. The market should reward true flexibility. But it should also reward those who build it. The enterprise AI agent market is about to become the greatest distribution fight of the decade. Oracle and Google have fired an important shot. Salesforce and ServiceNow will answer. Microsoft will answer. And somewhere in that chaos, the crypto world will find its mirror image. The autonomous agents of the public ledger are the cousins of the governed agents of the ERP. They share the same parent: code that acts in the world. The question is not whether the agents will act. It is who will be accountable when they do. The Oracle-Google partnership does not answer that question. It just makes it impossible to ignore. The ghost is in the machine. The machine is everywhere.