Why Business Data Needs a Common Language Before AI Can Use It

An AI system can read an invoice correctly, identify the right supplier, and calculate the exact amount due. It can still make the wrong decision. Perhaps the supplier belongs to a parent company with a special credit arrangement. Perhaps a payment is waiting for an approval that exists outside the system. The data may be accurate, yet the business conclusion may be wrong.

This is becoming one of the central challenges in enterprise artificial intelligence. As companies move beyond chatbots and begin allowing AI agents to approve transactions, coordinate deliveries, and manage financial workflows, the cost of misunderstanding business context rises sharply. A convincing answer is no longer enough. The system must understand what the information means, how it relates to other records, and which rules govern the next action.

The emerging competition is therefore about more than building larger models. It is about connecting intelligence to the institutional knowledge that makes a business function.

What Has Changed in Enterprise AI?

The argument from SAP, presented by Yaad Oren, the company's global head of research and innovation and managing director of SAP Labs U.S., begins with a practical observation. Businesses do not operate through isolated pieces of information. They operate through relationships, procedures, obligations, and exceptions.

A customer record, for example, is not simply a name and a revenue figure. It may connect to outstanding sales orders, customer satisfaction, delayed shipments, contractual discounts, and the manufacturing facilities responsible for delivery. Each connection can change how a company should respond to that customer.

Oren argues that enterprise AI needs three kinds of business knowledge that general-purpose large language models do not automatically possess.

The first is knowledge of relationships between business entities. The second is an understanding of the processes surrounding each task. The third is the ability to recognize exceptions and variations that arise in everyday operations.

SAP Business AI Platform is designed to bring these elements together in a governed environment. Its architecture combines a knowledge graph, shared business definitions, operational context, and tools for building and managing AI agents.

The significance is not that AI has suddenly become infallible. It is that enterprise software is beginning to address a weakness that cannot be solved simply by improving the fluency of a language model.

Why This Matters Beyond Technology

For decades, businesses have invested in enterprise software to make their operations more consistent. Accounting systems track financial obligations. Supply chain applications monitor inventory and shipments. Customer management platforms record sales activity and customer relationships.

The difficulty is that these systems often describe different parts of the same economic reality.

Consider a company preparing to offer a discount to a major customer. The decision may depend on the customer's purchase history, outstanding invoices, delivery performance, and the company's ability to manufacture additional products. Marketing, finance, sales, and operations all hold pieces of the answer.

An AI agent that examines only the marketing records could recommend a discount that creates problems elsewhere. A system that understands the broader relationships could identify the trade-offs before the decision is made.

This is where the economic value of enterprise AI begins to emerge. Better context can help organizations reduce unnecessary delays, coordinate departments, and make decisions with fewer conflicting assumptions.

But there is an important distinction between a technically impressive demonstration and a dependable business system. The former may produce a plausible answer from a limited set of information. The latter must operate consistently within real contractual, financial, and regulatory constraints.

For large organizations, that distinction can determine whether AI remains an experimental tool or becomes part of everyday operations.

How the Technology Works

The underlying architecture combines several components rather than expecting a single model to do everything.

A knowledge graph maps connections among customers, suppliers, products, assets, organizational structures, and business processes. A business ontology defines what important concepts mean across different systems. Operational rules then establish how tasks should proceed and which constraints apply.

These components address different problems.

A graph can establish that two records refer to related entities. An ontology can clarify whether the word "invoice" means a supplier bill, a customer receivable, or an employee expense claim. A policy layer can determine which approvals are required before a transaction proceeds.

SAP's architecture also includes models designed specifically for structured data. The company identifies RPT-1.6, a relational pre-trained transformer, and TabPFN-3.5 Plus from Prior Labs among its efforts to improve AI processing of tabular information.

This matters because enterprise data is often stored in tables containing numerical values, categorical fields, missing entries, and many distinct identifiers. Such information behaves differently from the natural-language documents on which general-purpose language models commonly rely.

SAP's tabular orchestration capabilities in AI Core are intended to handle context selection and data access at runtime, allowing predictions to enter production workflows.

The larger principle is straightforward. Specialized models can analyze structured information, shared platforms can supply business meaning, and agents can use the resulting information to carry out tasks.

Where the Data Lives

The challenge is not simply finding data. It is establishing which records belong together and what they mean.

A sales application might contain an order identifier. A logistics system might record the shipment associated with that order. A financial application might contain the invoice and payment status. Without reliable connections, an AI agent may treat these records as separate events rather than stages in one business process.

The knowledge graph supplies relationships. The ontology supplies consistent definitions. The operational layer supplies rules covering payment terms, service-level agreements, tax requirements, security restrictions, regional policies, and data anonymization.

Together, these elements create a more complete representation of the business.

Yet connected data is not necessarily correct data. If the source systems contain outdated records, contradictory contract terms, or incomplete information, the resulting AI decisions can still be flawed. Companies must maintain data quality and establish procedures for resolving conflicts rather than assuming that a sophisticated architecture will eliminate them.

How Businesses Create and Capture Value

The immediate opportunity is operational efficiency.

In finance, agents could help reconcile invoices, identify unusual transactions, and prepare routine payments for approval. In supply chain management, they could connect production schedules, shipment histories, and order commitments to identify emerging delays. In customer operations, they could combine purchasing patterns and service histories to help employees make better-informed decisions.

These applications could reduce repetitive work and allow employees to focus on cases that require judgment.

For SAP, integrating models, enterprise data, agent development, and governance also creates a commercial opportunity. Customers may find it attractive to deploy more capabilities within an environment that already supports their business processes.

The company positions its Joule Agents across functions such as finance and supply chain as part of this broader strategy.

However, the original information provides no quantified evidence of additional revenue, cost savings, or investment returns. The commercial case remains dependent on what organizations achieve in practice, including the accuracy of decisions, the cost of implementation, and the amount of manual work eliminated.

A credible evaluation should measure these outcomes rather than count the number of agents deployed.

Platform or Product?

A standalone AI product can perform a specific task extremely well. A platform becomes more valuable when multiple tasks need to share information, definitions, and controls.

Imagine a finance agent and a supply chain agent working with the same supplier. If each maintains its own interpretation of the supplier's identity, contractual terms, and transaction history, inconsistencies can spread across the organization.

A shared platform can give both agents access to the same underlying business representation. It can also centralize important permissions and audit records.

This creates the possibility of a compounding advantage: each new agent can benefit from relationships and definitions that already exist.

Nevertheless, not every company needs a comprehensive enterprise platform. A small organization automating email classification may require only a limited set of integrations and rules. A multinational corporation operating across numerous departments and jurisdictions faces a substantially different problem.

The appropriate architecture depends on the complexity of the business, the sensitivity of its data, and the consequences of an incorrect decision.

Competition and Competitive Advantage

Enterprise AI providers approach the market from different starting points. Model developers focus on reasoning and language capabilities. Data infrastructure companies concentrate on organizing and accessing information. Enterprise software providers bring established business applications, workflows, and institutional data.

SAP's potential advantage lies in connecting AI to the business processes its software already supports. But that position is not proof that its approach will outperform every alternative.

Competitive outcomes will depend on integration with external systems, implementation costs, model performance, governance quality, and the ease with which customers can adapt or replace individual components.

The strongest position may belong to providers that combine reliable models with deep business context and practical controls. Intelligence matters, but so does the ability to use it within an organization without disrupting operations.

The Risks of Giving AI the Power to Act

When an agent moves from answering questions to executing transactions, governance becomes an operational necessity.

A rejected order can damage a customer relationship. A blocked invoice can disrupt a supplier. An incorrect credit-risk classification can influence decisions across an entire account.

SAP describes a framework in which agents on its platform receive individually provisioned identities, scoped permissions, and audit trails documenting their actions and authorization boundaries. Its AI capabilities also undergo ethics and risk reviews, with feature-level assessments intended to address relevant regional policies, privacy laws, and regulatory requirements.

One example illustrates why exceptions matter. SAP describes a parked document, such as a catering invoice that a financial controller must allocate across several budgets before it can proceed. If that employee is on vacation or parental leave, an agent monitoring the workflow might interpret the delay as a process failure.

A system with broader context could recognize the reason for the delay and respond accordingly.

Still, recognizing an exception must not mean ignoring controls. Some decisions should remain subject to human approval, particularly when financial losses, legal obligations, or customer rights are involved.

Trustworthy automation requires both flexibility and firm boundaries.

How the World Could Be Affected

The wider economic implications extend beyond software companies.

As businesses automate routine decisions, productivity gains may come from better coordination rather than simply faster processing. Manufacturers could improve the connection between production and delivery. Financial teams could spend less time reconciling records. Service organizations could handle routine requests while directing complicated cases to experienced employees.

These changes could also reshape competition. Organizations with well-structured data and clearly documented processes may find it easier to deploy AI than companies whose operations depend heavily on disconnected systems and informal knowledge.

The effects on employment will vary by task and industry. Some repetitive administrative responsibilities may decline, while demand could increase for people who can design workflows, maintain data quality, supervise automated decisions, and resolve unusual cases.

There are also geopolitical and regulatory considerations. Enterprises operating internationally must navigate different privacy requirements, security expectations, and regional restrictions. A platform that supports automation across borders must account for those differences rather than treating every market as identical.

The opportunity is substantial, but so is the need for responsible implementation.

What Businesses Should Do Next

Companies do not need to begin by automating every department. They should start with a clearly defined process whose performance can be measured.

First, identify the relevant records and determine whether customer, supplier, transaction, and operational data can be connected reliably.

Second, document the rules that govern the process, including approval requirements and common exceptions.

Third, begin with an agent that recommends actions or prepares transactions for human review. Expand its authority only after its performance has been tested under realistic conditions.

Finally, measure the results. Processing time, error rates, operating costs, customer outcomes, and the frequency of human intervention provide more useful evidence than the number of AI features installed.

This approach is relevant to small businesses as well as multinational corporations. A smaller company may not need an elaborate knowledge graph, but it still benefits from consistent customer records, clear operating procedures, and defined limits on what its AI tools can do.

The objective is not maximum automation. It is reliable automation that creates measurable value.

Enterprise AI is entering a phase in which the central question is no longer whether a model can generate an intelligent response. It is whether that intelligence can be translated into a correct, authorized, and useful business action.

SAP's approach highlights the importance of combining structured-data models, knowledge graphs, shared business definitions, operational context, and governance. None of these elements guarantees perfect decisions. Together, however, they can address important weaknesses that remain when AI operates without a reliable understanding of the organization around it.

The competitive advantage of the next generation of enterprise software may therefore come from more than model size or conversational ability. It may come from the quality of the connections between data, decisions, and institutional knowledge.

For businesses, the practical lesson is clear. Before handing more authority to AI, make sure the system understands what the data represents, how the organization works, and when it should stop and ask a person.

That is the difference between an AI system that merely sounds intelligent and one that can become a dependable part of the business.

Why Business Data Needs a Common Language Before AI Can Use It

enterprise AI, AI agents, SAP Business AI, knowledge graphs, business context, AI governance, enterprise automation, structured data AI, business ontology, AI decision making, supply chain AI, financial automation, enterprise data, AI business strategy