SAP, Oracle, and Palantir are facing an uncomfortable problem: business intelligence no longer has to live within the same system that records transactions. For decades, large companies have accepted an almost inevitable logic: if the ERP manages finance, procurement, HR, or logistics, it seemed reasonable that the advanced analytics and automation layers depended on the same provider.
That balance is changing. Databricks, Snowflake, and other data infrastructure players are pushing enterprise AI towards a different layer: governed data, open formats, and agents capable of querying, reasoning, and acting on information that doesn’t necessarily reside solely within a single ERP or a closed decision platform. The movement doesn’t eliminate SAP, Oracle, or Palantir, but it questions where the greatest value is created.
For years, the traditional model worked with tremendous strength. A company deploys SAP or Oracle to record critical operations. Here are invoices, orders, inventories, contracts, payroll, and much of the daily pulse of the business. When it’s time to apply AI, automate decisions, or connect processes, the natural choice of provider is usually the one already controlling the transactional system.
The old power of ERP begins to separate from data
SAP and Oracle are reacting quickly. SAP is strengthening Joule and its Business AI strategy with integrated assistants and agents in business processes, supported by data, security, and corporate governance. Oracle, for its part, has reconfigured its Fusion Cloud suite toward what it calls “agentic apps,” designed so users can request business results rather than just execute isolated tasks within the software.
This response makes sense. If enterprise software is reduced to a transactional database with interfaces, the value margin shifts elsewhere. ERPs will remain necessary because a company needs reliable systems to record purchases, sales, payroll, taxes, manufacturing, and supply chain. But the layer that interprets this data, cross-references with other systems, and proposes decisions is beginning to become independent.
Palantir represents another model. Its Ontology isn’t limited to querying tables: it seeks to represent real business objects such as factories, vehicles, orders, products, or assets, and connect them with actions, permissions, and operational flows. Palantir’s documentation defines Ontology as an operational layer over digital assets integrated into Foundry, capable of linking data with real-world equivalents.
The potential is very powerful, especially in complex organizations. The strategic cost is clear: the more value that semantic layer generates within a specific platform, the harder it becomes to detach from it. The company isn’t just buying software. It’s essentially transferring part of its operational knowledge into a structure dependent on a provider.
Databricks and Snowflake attack from below
Databricks and Snowflake’s offensive doesn’t come from the ERP but from data infrastructure. Their thesis is different: enterprise intelligence should be built on top of the organization’s data, not necessarily within the application that originated it. This includes ERP data, but also data from POS systems, e-commerce, CRM, sensors, internal spreadsheets, documentation, campaigns, inventory, and logistics systems.
Databricks has taken a further step with Genie One, a new line of AI agents for business teams. According to The Wall Street Journal, the product is designed so finance, marketing, or sales professionals can get answers and make decisions based on corporate data. The centerpiece is Genie Ontology, a contextual layer that organizes data, documents, applications, and people so agents can respond more accurately.
The case of Albertsons Companies helps illustrate the shift. The American supermarket chain works with Databricks on merchandising and pricing intelligence. During the Data + AI Summit, Databricks explained that Albertsons processes 70 billion transaction rows in its Lakehouse and models relationships between items and promotions using reusable structures capable of showing substitution effects, induced sales, and cannibalization among products.
The question is no longer just “how much did this cheese sell last week?”. The more interesting question is: if a promotional blitz is launched for a specific brand, how much can it steal from private label, what effect will it have on related products, and how much shelf space should be adjusted? Previously, such a query could require weeks of extraction, reconciliation, and analysis across ERP, logistics, POS, and BI tools. Now, the goal is for an agent to operate on a data layer prepared to respond to these kinds of decisions.
Accenture also highlights Albertsons as an example of adopting agentic solutions on Databricks for pricing intelligence, with historical analysis, forecasting, and explainability tailored for category managers.
Snowflake follows a similar approach. The company emphasizes open architectures, governed data, and interoperability so that agents work on reliable definitions rather than improvisations. Recent announcements about open formats and Iceberg aim to ensure data can be accessible across different tools without being trapped in a single proprietary layer.
The real battleground is the semantic layer
At its core, the debate isn’t about replacing ERP with AI agents. It’s about who controls the company’s semantic layer. That is, who defines that a product code belongs to a category, that a promotion affects a segment, that a store behaves comparably to another, that a logistical delay impacts margin forecasts, or that a business decision influences inventory, procurement, and sales.
For a long time, this intelligence was spread across consultants, reports, internal processes, and proprietary modules. Palantir made it into a high-value operational layer. SAP and Oracle are trying to embed it into their suites. Databricks and Snowflake want it to be built on top of the data platform.
For the board, this becomes an even more uncomfortable question: it’s not enough to choose which AI tool to buy. They must decide where to build the company’s operational brain and under what rules. Building inside the ERP can offer tighter integration but also greater dependence. Building within a closed operational intelligence platform may provide quick returns, but exiting can be difficult. Relying on open, governed, and reusable data demands more technical discipline but offers more flexibility.
The ERP isn’t going away. SAP and Oracle will continue to hold strong because they record processes that can’t be improvised. Palantir remains relevant, especially for organizations needing to connect decisions, operations, and security in very complex environments. But the center of gravity is shifting.
A company that treats its data as a byproduct of ERP might end up purchasing intelligence multiple times: once from the transactional provider, another from analytics, another from agents, and another from consulting firms connecting it all. Conversely, a company that organizes its data as a strategic asset will be better positioned to choose models, agents, and applications without constantly rebuilding its architecture.
Enterprise AI isn’t just about automating human tasks versus manual ones. It’s about separating the system that records business from the system that understands it. This divide could be the defining tech battle of the next decade.
Frequently Asked Questions
Does this mean SAP and Oracle will lose importance?
Not exactly. Their systems will remain essential for recording critical operations. What changes is that intelligence and automation can be built outside the ERP, on a broader data layer.
What distinguishes Palantir from Databricks or Snowflake?
Palantir offers a closed, highly integrated operational layer representing objects, decisions, and actions. Databricks and Snowflake focus more on data infrastructure, enabling agents to work on governed and reusable data.
Why do open formats matter in enterprise AI?
Because they reduce dependency on a single provider and allow different tools to access the same data with common rules. In AI, this can help agents work on consistent information.
What’s the main risk for businesses?
The key risk is creating a new technological dependency just as they are trying to modernize with AI. The important decision isn’t just which agent to buy; it’s where the operational knowledge is stored and governed.
