LLMs vs. RFM: why ChatGPT can't predict your customer's next move

Somewhere in your company, someone has already tried this: paste a customer’s data into ChatGPT or Claude and ask, “Is this customer good or bad?” The answer comes back fluent, confident, well-structured… and statistically worthless.

The problem isn’t that the model is bad but that it’s the wrong kind of model for the question.

The two halves of the enterprise brain#

Think of your company’s data as a brain with two halves.

One half is unstructured: documents, contracts, emails, support tickets, and meeting notes. For this half, the AI revolution has already arrived. Large language models read it, summarize it, and answer questions about it.

The other half is structured: transactions, registrations, payment histories, product catalogs, ownership relationships, and clickstreams. This is the operational core: the ground truth of the business, its blueprint. It’s also where the decisions that actually move money live: who gets credit, which transaction is fraud, which customer is about to leave, which lead will convert. For this half, the revolution never showed up. There was no model you could simply point at your database and ask, “What happens next?”

That asymmetry is the gap this article is about.

Why language models fail here#

An LLM is trained to predict text from text. That makes it remarkably good at anything that lives in language: reasoning, summarizing, drafting, explaining. But default risk, fraud, and churn are not properties of language. They are properties of structure and time: who is connected to whom, how behavior changed, in what order events happened, how patterns in millions of similar histories played out.

When you ask an LLM a predictive question about your business, it has none of that signal. It has never seen your data distribution, can’t propagate information through your entities’ relationships, and has no notion of your base rates. So it does what it was built to do: produce plausible text. Plausible is not the same as calibrated, and in credit and fraud, the difference is measured in money.

Text-to-SQL lets a language model translate a question like “how many customers churned last quarter?” into a database query and return the answer. But look closely at what it’s doing: retrieving and aggregating what already happened. Every text-to-SQL answer is a fact about the past.

A decision needs the opposite. “How many customers churned last quarter?” is a query. “Which customers will churn next quarter?” is a prediction. No amount of SQL, however cleverly generated, can select over the future. You have to forecast the future before you can query it.

The traditional answer, and why it doesn’t scale#

Until now, enterprises answered predictive questions the manual way: hire a team of data scientists, engineer features by hand, assemble training datasets, build and validate a bespoke model, then deploy and babysit it.

It works. It’s also slow and expensive: commonly twelve months from question to production model. And it’s twelve months per task: the churn model doesn’t help you with fraud, the fraud model doesn’t help you with credit limits, and behavior often differs enough by segment or region that you end up building variants of the same model over and over.

The result is that most predictive questions inside a company simply never get asked. The cost of answering them is too high.

A foundation model for structured data#

What changed for text can change for structured data: instead of building one model per task from scratch, pre-train a single model on a massive corpus, then adapt it cheaply to each specific problem.

That’s the idea behind relational foundation models (RFMs), a new class of models designed from the ground up for structured business data. Rather than reading sequences of words, an RFM represents the entities of a business (customers, companies, products, transactions) and the relationships between them as a graph, and learns from millions of historical trajectories how outcomes unfold across that graph over time.

The payoff mirrors what LLMs did for text. The heavy lifting (learning which relational patterns are predictive) is done once in pre-training. Answering a new predictive question no longer requires months of feature engineering and model construction; the pre-trained model already carries most of the signal, and a comparatively light adaptation step tailors it to the task. Out of the box, models of this class match what expert teams build by hand over months. Fine-tuned on a specific use case, they surpass them.

What this looks like at Avra#

We build this class of foundation model in two complementary forms. Together they cover the two ways structured business data becomes a prediction — and the approach generalizes: it isn’t tied to any one industry, database, or market.

The first is a Graph Foundation Model (GFM): a foundation pre-trained on a large temporal knowledge graph of an economy — its companies, individuals, and the relationships between them. It learns how an entire market is wired, then powers customer-specific downstream models for credit risk, fraud, growth, and monitoring, delivered through an API into the decision systems customers already run. The same recipe can be pre-trained on any economy for which that graph can be assembled.

The second is our Relational Foundation Model (RFM): a schema-agnostic foundation model built to read a client’s own relational database directly. Instead of learning a fixed table layout, it reads table and column names as language, so the same pre-trained weights adapt to a marketplace’s transaction tables as readily as a bank’s loan book — in any market, without a bespoke research project per client.

The two are complementary. A GFM contributes the outside view: what the wider economy reveals about an entity. An RFM contributes the inside view: what a company’s own data reveals about its customers and their behavior. And they compose — a GFM’s embeddings can feed into an RFM as an additional signal, so a single prediction can draw on both at once.

Because each foundation already understands the structure and history of the domain it was pre-trained on, both models perform even where a given customer’s own data is thin: a company that just opened, a customer with no history in your book, a counterparty you’ve never seen. Their neighborhood in the graph has been seen before.

Complementary, not competitive#

So should you cancel your ChatGPT contract? Obviously not. LLMs and RFMs are not rivals. They are the two halves of the same brain.

LLMs predict text from text. RFMs predict outcomes from business data. A complete decision workflow usually needs both. Consider renewals: a relational model estimates which contracts are at risk and which retention offer each customer is most likely to accept; a language model then drafts the outreach, explains the reasoning to an analyst, or powers the conversation. Take away the language model and you lose the interface with humans. Take away the relational model and the whole workflow is guesswork wearing a confident tone.

This matters even more as companies adopt agents. Today’s agents can retrieve documents, search the web, write code. What they can’t do is make grounded decisions about your business, because deciding well means forecasting the outcome of each action you’re considering, and that forecast has to come from your structured data. Agents get real decision-making capability the moment they can call on accurate predictions. That’s the piece that doesn’t exist off the shelf. And building that piece, for whatever market a business operates in, is what we do.

The question to ask#

Predictions exist to become decisions. Every meaningful decision, whether to extend the credit, block the transaction, or make the offer, is a bet on a forecast, whether that forecast comes from a model or from someone’s gut.

So the question for 2026 isn’t “which AI should we adopt?” Half of your company’s brain already has its answer. The question is what you’re doing about the other half.