Glossary

Foundation model

A foundation model is a large neural network trained on broad, general-purpose data at scale and designed to be adapted to a wide range of downstream tasks through fine-tuning, prompting, or tool integration, rather than being built from scratch for any single application.

Before foundation models became practical, most production NLP systems were trained end-to-end for one task, such as intent classification or named entity recognition. Building and maintaining a separate model for each use case was expensive and slow. Foundation models changed the economics by establishing a single pre-trained base that accumulates broad linguistic and world knowledge during training, then passes that knowledge to downstream tasks with relatively little task-specific data. For customer service platforms, this means teams can deploy a capable support agent without labeled training corpora numbering in the millions of examples.

How foundation models work

Foundation models are trained through self-supervised learning on large corpora, most commonly using the transformer architecture with a masked-language or next-token prediction objective. During this pre-training phase, the model develops representations of language, reasoning patterns, factual associations, and, in multimodal variants, images or audio. After pre-training, the model can be adapted through several mechanisms.

Prompting: Task instructions are passed in the context window without any weight updates. This includes zero-shot and few-shot learning approaches.
Fine-tuning: A subset of the model's weights are updated on a curated task-specific dataset. Fine-tuning is the most common way to adapt a foundation model for a specific domain or tone.
Tool integration: The model is equipped with external tools via function calling or retrieval, letting it access live data without retraining.
Retrieval augmentation: The model is paired with a retrieval system to ground its responses in current, authoritative documents, addressing the knowledge cutoff inherent in any pre-trained model.

The term was formally introduced in the Stanford HAI paper On the Opportunities and Risks of Foundation Models (Bommasani et al., 2021), which established the conceptual framing now used across the industry.

Why foundation models matter for customer experience

The practical consequence of the foundation model paradigm is that generative AI for customer service becomes accessible to organizations that could not previously afford to build proprietary NLP systems. A team can select a foundation model appropriate for their latency and cost constraints, apply domain adaptation through fine-tuning or retrieval, and deploy a support agent within weeks rather than years.

The risk is dependency. When a service team's entire customer-facing AI stack sits on a single third-party foundation model, changes to that model, whether a new version with altered behavior, a pricing change, or a service disruption, affect the whole product. Evaluating model updates carefully before deploying them to production and monitoring for model drift after updates are standard practices for managing this risk. AI observability tooling is the mechanism teams use to detect when a model update has shifted output quality in a direction that affects customer outcomes.