Foundation model: the base model behind your assistant
Definition
A foundation model is a large, broadly trained AI model that serves as the basis for many tasks, from understanding text through summarising to coding. It is trained once on huge amounts of data and then used for concrete purposes, for example as the engine of an assistant. You rarely build from scratch, you sit on top of such a base model and only adapt the use.
Why the Term Matters
The foundation model is the engine, not the finished car. It brings language understanding and broad general knowledge, but it does not know your company. For a usable assistant you combine the base model with your knowledge via RAG, with clear instructions and with permissions. Understanding this, you pick the model by task, cost and data protection and do not confuse it with the whole solution.
When It Is Relevant
- An assistant or agent is set up new and needs an engine.
- Tasks differ in difficulty, language and pace.
- Cost per request and answer quality must be balanced.
- Data protection or on-premise operation dictate the model choice.
When Not
When a ready standard product fully covers the task, you barely need to worry about model choice. And the debate over the perfect model is not worth it while knowledge, permissions and prompt are not cleanly in place, since the bigger lever usually sits there.
Industries That Use It
- Software houses: a base model with strong code understanding for development help.
- Public authorities: a model that can run in a data-protection-compliant way.
- Translation and localisation: a model with wide language coverage for many target languages.
- Retail: a low-cost model for high volumes of simple requests.
Typical Use Cases
- The engine behind a personal AI assistant.
- The base of a company brain connected via RAG.
- A strong model for demanding analysis, a cheap one for bulk tasks.
- A locally runnable model for sensitive data.
- A multilingual model for international teams.
Decision Guide
| Signal | Choose foundation model by | Alternative |
|---|---|---|
| Demanding tasks | quality and context size | a small model struggles |
| High volume, simple | cost per request | an expensive model is waste |
| Sensitive data | on-premise operation | avoid open cloud |
| Only standard use | a ready product | no own choice needed |
Example
A team picks two models for its assistant: a strong one for complex analysis and a cheap one for simple, frequent questions. Both are supplied with company knowledge through the same company brain. So quality stays high where it counts and cost stays low where volume rules. This describes the method, not an asserted metric.
Service Connection
We make the model choice when building an AI assistant and in custom GPT development, always in interplay with the company brain.
Sources
This framing describes common practice around foundation models. Which model fits depends on task, cost, language and data protection, and must be checked case by case.
