Skip to main content
Omni’s AI features consume tokens each time they process a request. The Token Tracking dashboard, available in the Analytics section of Omni, helps you understand how tokens are being used across your organization.

​
Requirements

Organization Admin permissions are required to access the Token Tracking dashboard.

​
What are tokens?

Tokens are the unit of measurement for AI model usage. Every interaction with an AI feature in Omni — such as asking a question, generating a summary, asking AI to make data model changes, or using the MCP Server — consumes tokens. Token usage varies depending on several factors:
  • Task complexity — More complex questions, multi-step analysis, or larger datasets use more tokens than simple lookups. Each message in a session carries prior context, so long running sessions can compound.
  • Data model and context size — Omni sends context from the data model to the LLM to improve answer accuracy (field descriptions, ai_context, synonyms, etc). Larger models use more tokens.
  • LLM model — More capable models (e.g., Sonnet-class) consume more tokens per request than lighter models (e.g., Haiku-class). See AI model settings to configure tiers.

​
Using the Token Tracking dashboard

To access the Token Tracking dashboard, click Analytics in the left navigation of the Omni app and select the Token Tracking dashboard. The dashboard displays token consumption over time and can break down usage in various ways, such as by feature and user.

​
Optimizing token usage

  • Adjust the LLM model tier. See AI model settings.
  • Trim unused fields, descriptions, or ai_context from the data model. Reducing context may lower per-call token counts, but can increase the number of turns needed to get a good answer. The goal is optimizing token efficiency overall.
Set up alerts to be notified when nearing your token limit. See Usage analytics for more information.

​
FAQ

Omni’s AI is an agentic system. For a given prompt or request, it may fire off multiple tool calls to get the answer or complete the task. For example, asking “show me revenue by region” might involve the AI searching the data model for the right fields, building the query, creating a visualization, and summarizing results — each as a separate step. Each of these steps is logged individually in the tracking data, which is why a single user prompt can appear as several rows.
These are “tool calls” — intermediate steps where the AI determines which action to take next. They don’t have a user-facing prompt because they are part of the AI’s internal reasoning as it works through a request. See the question above for more context on why these occur.
Token usage can vary between identical prompts due to differences in conversation context (earlier messages in the session) or the AI choosing a different path to arrive at the answer.
Yes, MCP calls do consume tokens. The granular tracking data for an individual MCP call may show 0 tokens, but the token usage is logged under the associated query tool call.

​
Next steps