Structured Data Is AI's Overlooked Engine

The most valuable data in most companies isn't in their chatbot logs or their marketing copy. It's in the tables — the spreadsheets, the ERP systems, the transaction records, the demand forecasts that get run every Monday morning and quietly drive billions of dollars in decisions. Nobody made a TED Talk about it. Nobody put it on a billboard. It just sits there, doing the actual work.

That gap between where AI's spotlight has fallen and where enterprise value actually lives is the entire premise behind Fundamental, a company built by Jeremy Fraenkel — a CEO and co-founder who came out of finance and machine learning and found, when the LLM wave arrived, that none of the new tools solved his actual problem.

"Businesses don't actually run on unstructured data," Fraenkel told Tanuja Randery, VP and Managing Director of EMEA for AWS, in a recent episode of the AWS Executive Insights podcast. "They really run on tables. Spreadsheets, databases, ERPs, transaction logs. It's all structured data. And that part of the enterprise world never had its ChatGPT moment."

That framing is worth sitting with. The ChatGPT moment was visible. It was theatrical. You could show your grandmother what it did. Tabular data doesn't demo like that. A fraud model's output is a number, possibly followed by a declined transaction. There's no prose to admire, no image to share. And yet that number — and the prediction engine behind it — is what keeps a financial institution from hemorrhaging money it can't account for.

The Hard Problem Nobody Wanted

Fraenkel is candid about why the tabular space stayed underserved. Part of it is attention economics: when GPT-4 arrives and does things that feel like magic, the whole industry chases the trick. But the other part is technical, and it's more stubborn.

"Tables don't behave like language," he said. "Enterprise data is messy, relational components, they're incomplete, they're constantly changing. And numbers also have different structures than words."

This matters architecturally. You can't simply repurpose a language model's training approach for structured data. The internet is full of text — that's where LLMs get their raw material. But the most valuable tabular datasets are behind corporate firewalls, proprietary by nature, and unavailable for training. Fraenkel's team sourced from open data and synthetic data, assembling what Fundamental's own website describes as training on billions of tables to build NEXUS, their foundation model for structured prediction.

The older answer to this problem was gradient-boosted tree algorithms — XGBoost being the most recognizable name — and for a long time, those were considered sufficient. Fraenkel's argument is that "sufficient" was always a compromise, and that the foundation model paradigm, applied to tabular data, can do substantially better. That's a claim that requires scrutiny, and he seems to know it.

"What I tell every single enterprise is don't trust anything I showed you," he said. "Don't trust any of the numbers I mentioned. Don't trust any of the benchmarks. Try it for yourself and see the results."

That's an unusual sales pitch. It's also, notably, the only defensible one for a product category where the proof has to live in production, not in a demo.

The Enterprise Reality Check

Fraenkel's account of what actually happens when you bring AI into large enterprises is one of the more grounding parts of this conversation — and it cuts against a common assumption. People outside these organizations tend to imagine that Fortune 500 companies have their data organized, catalogued, and ready to be analyzed. They don't.

"Those enterprises were built before the age of AI," he said. "Their internal tools were never really designed to work together. And so that creates an immense amount of challenge."

What this means in practice: enterprises with enormous data reserves still deploy armies of data scientists just to clean, reconcile, and prepare that data before any actual modeling begins. The complexity compounds with every additional use case. This is the mundane reality that vendor pitches tend to skip over — and it explains why Fundamental had to build connectors to platforms like Databricks and Snowflake. The best model in the world doesn't help if the customer can't point their data at it.

The friction extends beyond infrastructure. Fraenkel described working with a major company where the business leaders, when shown a meaningful accuracy improvement on a prediction problem, essentially shrugged. It took bringing in their own data scientists to shift the reaction from indifference to genuine interest. The cultural translation problem — making executives understand why a small improvement in a numerical model is worth caring about — turns out to be as hard as the technical one.

The Security Architecture Worth Noting

One detail in this conversation that enterprises in regulated industries will want to understand: Fundamental's deployment model is built around what Fraenkel calls "confidential compute." The model is encrypted at both the architecture and weights layers, and deployed fully within the customer's own environment. The customer's data never leaves their infrastructure.

This architecture exists because regulated enterprises — banks, insurers, energy companies — can't send sensitive operational data to an external model endpoint. That's not a preference; it's a legal constraint in most jurisdictions. If Fundamental's system required the data to travel, the addressable market would shrink considerably. The design choice is therefore both a security feature and a business strategy.

The Larger Pattern

Here's what this conversation is really about, underneath the product discussion: a company is betting that the AI industry has been building the wrong half of the brain.

The LLM wave was real and consequential. But language models are, at their core, pattern-matchers over text. The enterprise decisions that move markets — trading positions, credit approvals, inventory levels, demand forecasts — are not text problems. They're numerical inference problems, and they've been treated as second-class citizens in the current AI boom.

That's not because they're unimportant. It's because they're hard to make visible, hard to demo, and historically assumed to be "solved enough" by existing tools. Fraenkel's bet is that solved enough isn't actually good enough, and that there's a category of prediction problems — old problems, in many cases, running for decades — where significantly better models will translate directly to measurable revenue or risk outcomes.

That bet might be right. It also might underestimate how entrenched the existing tooling is, how long enterprise procurement cycles run, and how hard it is to displace a methodology that's been baked into institutional workflows for twenty years.

The compelling demos and the messy production deployments tell different stories. Fraenkel knows this — it's why his standard advice to prospects is to skip the benchmarks and run the test themselves. For enterprises evaluating this category, that's probably the right instruction regardless of which vendor is delivering it.

---

Bob Reynolds is Senior Technology Correspondent at BuzzRAG.