Why Structured Data May Be AI’s Next Enterprise Frontier

Enterprise-Focused AI Models

One side of the AI market is dominated by models built for language and other unstructured content. Those systems are strong at writing, summarizing, searching, coding and conversational tasks. They have changed how many knowledge workers interact with software and have made AI visible.

The other side is less visible and arguably more consequential. It is focused on prediction inside the systems that actually run a company. This is the more “traditional” world of AI that used to be the domain of data scientists before the term vibe coder or prompt engineer ever emerged. That traditional world has focused on high value data such as spotting likely churn, fraud, delays, defects, defaults, missed upsell opportunities and supply chain disruptions from the raw data sitting inside enterprise infrastructure. That work has enormous economic value, and it has never been easy. It has usually required specialized teams, long setup cycles and a good deal of manual labor.

In the case of the recent announcements, KumoRFM-2 is built for structured and relational enterprise data, not as an adaptation of a text model, but as a model meant to work closer to the shape of the underlying data itself. Snowflake’s Project SnowWork pitches “outcome-driven” AI grounded in governed enterprise data. Oracle says its AI Database architects AI and data together across operational databases and analytic lakehouses, with agents that can securely access real-time enterprise data. SAP’s planned Reltio acquisition rests on a similar premise: AI performs better when enterprise data is unified, cleansed, harmonized, and context-rich.

"Every enterprise runs on structured relational data, but until now, every AI tool has required companies to destroy the most valuable part of that data before they can use it,” said Vanja Josifovski, CEO and co-founder of Kumo, which was also on a recent Forbes AI 50 list.

Much of the public conversation around AI still assumes one model class such as OpenAI’s GPT, Anthropic’s Claude, Google’s Gemini, or the increasing set of models emerging from other vendors can do almost everything. While that might be the desired goal, it does not reflect how large organizations actually operate.

Enterprises do not have one kind of data. They have many. Contracts, emails, support transcripts, images, tickets, clickstreams, transaction records, account hierarchies, ledgers, supplier databases, product catalogs, claims histories and machine logs all exist side by side. Each has its own structure, its own rhythm and its own business value. A model built to predict the next word in a sentence is not automatically the right instrument for every other job. Instead of pushing all enterprise data through a language-first lens, structured data deserves its own model architecture and its own place in the enterprise AI stack.

Predictive modeling inside the enterprise is not new. Companies have been using machine learning for years to score leads, rank risks, estimate demand, flag fraud and forecast outcomes. The trouble has been the cost and complexity of getting models into production.

The standard workflow is one that data scientists know well. Teams pull data from several systems. They clean it, prepare it, handle missing information and define features by hand. They can train one model for one task. Tune it. Monitor drift. Rebuild when the business changes. Then do the same thing again for the next use case. This was the usual process for AI before generative AI flipped the script.

That process can produce strong results, though it behaves more like custom craftsmanship than like scalable software. It depends on scarce talent. It moves slowly. It makes experimentation expensive. It turns each predictive problem into a fresh project.

These enterprise-focused vendors make the pitch that a relational foundation model can reduce some of that overhead. Rather than asking teams to build every predictive system from scratch, the model starts with a broader learned understanding of how structured business data behaves across linked tables. The value proposition is efficiency, less feature engineering, less custom prep work and faster movement from raw data to usable prediction.

What Structured Models Offer That The Big AI Models Do Not

Trying to use one of the super popular LLMs for enterprise structured data is often like trying to put a square peg in a round hole. Teams that want to make a general-purpose language model work on a structured prediction task often have to flatten records into text, serialize tables, wrap the model in orchestration logic or chain multiple steps together. That may be workable for some cases but it causes as many problems and complexity as it solves. Hallucination of enterprise data is a major problem. Important structure can get diluted and precision can suffer. The system may sound smart long before it proves useful.

A model designed for structured enterprise data starts from a different place. It is built to work with rows, keys, tables and relationships. It is aimed at finding patterns in how business entities connect and change over time. That makes it more relevant for tasks where the real signal is not in one field but in the interaction between many fields spread across many records.

This is important because most enterprise outcomes are relational. They do not emerge from a single event or a single line in a spreadsheet. Customer churn may show up through a sequence of purchases, returns, discounts, service contacts, timing gaps and account behavior. Fraud may look harmless in isolation but suspicious when viewed across related accounts, transactions, devices or counterparties. A delayed shipment may be tied not to one supplier or one port, but to a chain of interactions across routes, weather, inventory position and warehouse constraints.

This is why structured data has such strategic weight. It contains the logic of the business in motion. It captures how entities connect, how events unfold and how small signals compound into expensive problems or profitable opportunities. A model built for that environment has a better chance of surfacing what conventional dashboards miss and what language-first systems are not naturally designed to see.

A Shift To More Pragmatic, Value-Focused AI

The next wave of AI may be less sensational and more valuable. It will likely focus on matching the model to the data type and the business decision. Language models will continue to be important, of course, especially as they see increasingly broader adoption and capabilities. They will sit at the interface layer, helping people interact with systems, search information, summarize outputs and navigate complexity.

Structured models on the other hand will be just as useful but less visible to the typical user. They may increasingly sit closer to the operational core, where prediction and decision support matter most. Enterprise-focused structured model releases into a crowded AI market make a case for segmentation. One model class for language. Another for structured prediction. Different tools for different layers of the enterprise.