For many CTOs, every week opens with the same question. The board wants to know when an LLM will run over the company data. Not next year, this quarter. A competitor is already doing it, so the story goes.
The question is understandable. The answer is less exciting than the question: if sources, definitions and permissions are not in order, there is no reliable road for this model to drive on.
A Ferrari needs a road
Everyone wants to run an LLM on the data warehouse, but you can't drive a Ferrari until the road exists. Platform first, model second.
A model is only as good as the foundation beneath it. Above that, it just gets faster and more expensive.
The image holds because an LLM has no knowledge of your organisation of its own. For current, organisation-specific answers about revenue, customers, stock or contracts, it must rely on the right business sources; general training knowledge is no substitute. When that data is fragmented, duplicated or contradictory, the answer is fluent and still wrong.
AI is the final piece: reliability starts in the layers beneath it.
The work that comes before AI
An AI application needs more than people who write good prompts. Analytics engineers and data platform engineers do the less visible work that comes before it: connecting sources, modelling data and monitoring quality.
They clean up messy pipelines, give the warehouse structure and tune clusters so the cloud bill stays under control. It is less visible than a demo, but it decides whether the demo ever reaches production.
What a foundation actually is
The word foundation often stays abstract. At Datahub it is concrete: a data and AI platform. That platform is the road every model drives on, and it consists of a few parts:
- One meaning. What is a customer, a revenue line, an active order? While three departments give three answers, the model will give three too.
- Reliable pipelines. Data arrives on time, complete and traceable. Errors are caught before a user catches them.
- A structured warehouse. A clear layered build, from raw source to usable gold table, so it is known where everything comes from.
- Governance. Who may see which data, and what a model may do with it. This is also the basis for private AI, where the data stays in your own hands.
- Controlled cost. Clusters and queries that run efficiently, so AI does not become a surprise on the invoice.
A usable answer starts with connected sources and depends on shared meaning.
Without a road, an expensive mistake
Putting an LLM on a messy environment does not give you a slow but honest assistant. It gives you a fast, persuasive speaker that delivers nonsense with the same confidence as facts. Expensive, and hard to check.
The fault may not lie in the model alone. Source selection, definitions and access rules can also be wrong. A better prompt does not repair a missing agreement.
One order, different answers
Take a planner asking: can we ship this order today? The ERP shows a stock position. The warehouse knows which goods have already been reserved. The quality team may have blocked a batch. Without those agreements, in stock does not yet mean available to ship.
The platform connects those records and makes the definition explicit: available to ship means available, released and not reserved. The calculation uses approved sources within the planner’s permissions. The model can then explain the result, including why something cannot be done.
That is a different division of work from asking a model to invent a number. The organisation owns the meaning; the platform executes the rules; the model helps the user work with the outcome. If a source is missing, that should become visible rather than disappear into a persuasive sentence.
A smaller first step
The reflex is to start with the model, because that is what the board sees. The smarter step is smaller: first look at how strong the foundation is. Which sources can be trusted, where is a single meaning missing, where is cost leaking away.
That is the thinking behind the foundation scan. A sharp view of what exists, what is missing and what has to be laid first. After that, the first use case can grow on trust instead of hope.
There is a strategic choice in this too. A foundation that stays in your own hands makes it possible to answer a large share of routine questions from the company brain with open-source or self-hosted models. How large that share is must be tested against your own questions, sources and quality requirements; it is not a fixed percentage or guaranteed outcome. The large frontier models stay valuable for research and business development, but the daily questions travel on your own road, with less dependence on an external model provider. That calls for a choice made up front, not afterwards.
The model is the easy part. The road is the work.
Start with one decision and build only the foundation needed to support it reliably.
