AI Models

In this view, you get an overview of all language models that you have available, as well as manage which language models should be available to use in your organization’s Intric environment. As an administrator, you can control exactly which models should be available in different parts of the platform. By then connecting specific models to different security classes, you can ensure that more resource-intensive or specialized models are only used where they are really needed.

Different types of models

For the platform to function optimally, four different types of models are used with specific areas of responsibility:

Model Type	Description
Completion (Language Models)	Functions as the system’s “brain”. Used to generate text, conduct dialogue and solve logical tasks based on the user’s instructions.
Embedding (Text Embedding)	Converts information into mathematical vectors to enable search in knowledge libraries (RAG). Rule: Only one active model per security class.
Transcription (Transcription)	Converts speech to text. Makes it possible to index audio files and interact with the assistant via voice.
Image (Image Generation)	Generates images from text descriptions. Image models can be selected per assistant in the assistant settings. Admins control which image models are available in the organization.

Sub-processor column

Each model in the table now includes a Sub-processor column. This shows which company directly processes data when that model is used.

Scenario	What is shown
Model accessed via an external API (e.g. OpenAI, Anthropic)	The respective model vendor
Model hosted by Intric on our own infrastructure	Intric’s underlying infrastructure provider

From a GDPR and data handling perspective, the sub-processor is the entity that receives and processes your organization’s data when the model is called. Use this column when reviewing your Data Processing Agreements (DPA) or when deciding which models are appropriate for a given security classification.

Token costs for custom models

For standard vendor models, Intric already knows the pricing and applies it automatically to your usage and cost reporting. For custom models — such as self-hosted models or organization-specific deployments — Intric has no way of knowing the price, so you set the token cost yourself directly on the model’s configuration in this view.

Recommended format and considerations:

• Enter the cost per token in the fields Intric provides, matching how your provider or internal cost model prices usage (for example, separate input and output token costs where applicable).
• Use the same currency and basis as the rest of your cost reporting so figures remain comparable across models.
• Enter your actual negotiated or internal rate rather than a public list price, so internal follow-up reflects real cost.

How it affects reporting: Once a token cost is set, it is applied to that model’s consumption wherever usage and cost are shown — including the Usage and Cost control views. This lets organizations that run custom setups keep cost follow-up aligned with their own model configuration.

Reviewing accuracy: Periodically compare the cost reported for a custom model in Cost control against your provider invoice or internal cost model. If they diverge, revisit the configured token cost and update it.

Limitations and assumptions: Cost reporting for a custom model is only as accurate as the value you enter. A model with no configured cost, or with an outdated value, will lead to under- or over-reported figures for that model, so keep these values current as your pricing changes.

Model selection – how should I think?

When choosing which models should be active for different assistants or security classes, base your decision on the following principles:

• Don’t overcomplicate it: Most modern language models are general-purpose and handle most everyday tasks very well.

• Speed vs. Reflection ability: For the personal assistant and simpler tasks, use a fast model. This provides a smoother user experience where responses come directly.

  • For complex tasks, such as advanced legal analysis, medical review, or heavy problem-solving, choose a model with high reflection ability, even though these may feel somewhat slower.

• Cost efficiency: By limiting access to the most advanced (and often more expensive) models to specific security classes where they are actually needed, you can optimize your organization’s resource usage.

You also need to consider what types of personal data might possibly be processed by the assistants being used. Read more about how this is clarified to users by using Security Classifications.