"Foundation model" used to mean one thing: a very large network, pretrained on a very large slice of the public internet, that other systems build on. GPT, Claude, Gemini, models of general human knowledge, rented by everyone and owned by no one who uses them.
An enterprise foundation model inverts that. It is a transformer pretrained on a single company's own data, its transactions, orders, claims, sessions, payments, that becomes the base layer for every predictive question that company asks. One model, one company. It doesn't know the internet. It knows you.
Definition. An enterprise foundation model is a neural network pretrained with self-supervision on one organization's historical event data, producing a shared representation of its customers, accounts, and operations, from which downstream predictions, churn, risk, fraud, lifetime value, are derived by lightweight fine-tuning.
The insight: your data is already a language
The reason language models work is not something special about words. It is that language is a sequence with structure: a grammar, long-range dependencies, patterns that repeat with variation. To predict the next word, a model is forced to learn all of it.
Enterprise data has exactly the same shape. A customer's history, a card swipe, a deposit, a login, a support ticket, a payment, is a sequence of events with its own grammar. Salary arrives, rent leaves, spending follows a weekly rhythm, and deviations mean something: a life event, an emerging risk, an intent to leave. A company's full history is a corpus of millions of these sentences.
Pretraining on that corpus with the same objective that built LLMs, predict the next event, forces the model to internalize customer segments, seasonality, product relationships, and risk dynamics. Nobody labels anything. The structure is already in the data; the objective extracts it.
What it replaces
The status quo in enterprise ML is one model per question. A churn model, built by one team over two quarters. A credit-risk model, built by another team on different features. A fraud model, a propensity model, an LTV model, each with its own pipeline, its own feature store entries, its own drift and maintenance burden. The expensive part of each project is the same work done over and over: turning raw history into features a small model can digest, and losing most of the signal in the process.
A foundation model collapses that stack. The pretraining run does the expensive representation work once, on the raw sequences, with nothing thrown away. Every downstream question becomes a small head fine-tuned on top, hours to days of work, not quarters, and every one of them inherits everything the base model already knows.
Each new prediction stops being a project and becomes a fine-tune.
How one gets built
1. Tokenize
Every event, a transaction, an order, a claim, is encoded as tokens: its type, its attributes, its amount, its timing. A customer's history becomes a sequence; the company's history becomes the training corpus. No feature engineering, no aggregation into monthly averages. The raw behavior is the training data.
2. Pretrain
A transformer is trained to predict the next event in every sequence. This is self-supervised: the labels are the data itself. To get good at this one task the model must learn who each customer is, what they are likely to do, and when a sequence has gone off-script, which is precisely the knowledge every downstream model needs.
3. Adapt
Downstream predictions are lightweight heads fine-tuned on the frozen or lightly-tuned base: churn, credit risk, fraud, LTV, propensity. Every entity, customer, account, product, also gets an embedding, a vector that drops into existing systems: lookalike audiences, recommendation retrieval, anomaly scores.
What it's not
It is not an LLM fine-tuned on your documents, and it is not RAG. Those techniques give a language model access to your text. But the signal that runs an enterprise mostly isn't text, it's structured events, and the questions that matter about them are predictive, not conversational. We've written a separate essay on why the two kinds of models are complementary rather than interchangeable.
It is also not a shared industry model that pools data across companies. The defining property of the category, and the reason enterprises can deploy it on their most sensitive data, is one company, one model: trained inside your cloud, on your data alone, producing weights you own.
Why the category exists now
Three curves crossed. Transformer training became cheap enough that a single-company pretraining run costs a rounding error of an ML team's annual budget. The research matured, self-supervised learning over event sequences now reliably beats feature-engineered baselines on tabular-adjacent tasks that resisted deep learning for a decade. And enterprises watched LLMs demonstrate, vividly, what pretraining does: a general representation that makes every downstream task easier.
The conclusion we draw at Lyon is the one we build toward: within a few years, every serious company will run on a model of itself. The only question is whether it takes them years to build one or weeks to deploy one.
Frequently asked questions
How is an enterprise foundation model different from an LLM?
An LLM is pretrained on public text and knows the internet. An enterprise foundation model is pretrained on one company's private event data and knows that business. LLMs generate and reason over language; enterprise foundation models predict behavior and produce embeddings for downstream models like churn, credit risk, and fraud.
What data does it train on?
Structured event streams: transactions, orders, payments, claims, sessions, support tickets, any timestamped record of what customers and accounts did. Each event is encoded as tokens in a sequence, and the model is trained to predict the next event.
Does the data leave the company to train the model?
No. With Lyon, training and inference run inside the customer's own cloud. The data never crosses the company's perimeter, the weights belong to the company, and no data is pooled across customers.
How much data do you need?
Less than most teams expect. Millions of events, a scale most mid-size enterprises already have in their transaction, order, or claims history, is enough for pretraining to beat feature-engineered baselines. More history helps, but the threshold is not hyperscaler-sized.
What predictions can be fine-tuned from it?
Churn, credit risk, fraud and anomaly detection, lifetime value, product propensity, demand forecasting, segmentation, and next-best-action, each as a lightweight head on the shared model, typically in days rather than as a months-long standalone project.