AI privacy and usage
What the AI provider sees, what it never sees, and how tokens, costs and the Orcabase AI credit are counted for your organization.
When you ask the agent a question, parts of the conversation are sent to an AI model. This page says exactly what, who receives it, and how usage is counted.
What the AI model sees
- Your questions, and the conversation so far.
- The names of your data sources, schemas, tables and columns.
- SQL, and the definitions of your data models and metrics.
- Small samples of results. A result is cut to at most 30 rows and 16,000 characters before the model sees it. The agent is told to add things up in SQL, so it rarely needs raw rows.
- Names of saved queries and dashboards, when it looks them up.
What it never sees
- Database passwords, service account keys and other credentials.
- Data from other organizations.
- Full results. Charts in the chat are drawn from up to 1,000 rows that go to your browser, not to the model.
Who receives it
Two companies handle each request:
- The gateway (OpenRouter or Vercel AI Gateway) passes the request on.
- The model’s maker (for example Anthropic, OpenAI or Google) runs the model.
Their terms decide how long requests are kept and whether they’re used for training. With your own key, your own agreement with the gateway applies, including any data settings you choose there. With Orcabase AI, requests go through Orcabase’s OpenRouter account.
Note
Chats are stored in Orcabase, with their charts, and only the person who started a chat can open it.
How usage is counted
Every call to a model is recorded: which model, who asked, in which chat, how many tokens went in and out, and the cost. Tokens are how AI providers measure text; a word is roughly one to two tokens.
- In a chat: the header shows the chat’s total tokens, and its cost when the provider reports one. Tokens spent on an answer that failed or was stopped count too.
- In Settings → Agent: the Usage card shows this month’s calls, tokens and cost (or share of credit) by model, and by member. On Orcabase AI it also shows today’s, this week’s and this month’s credit.
- On your own key, costs are shown in dollars, and your gateway bills you directly. Its own dashboard is the final word on what you owe.
Orcabase AI credit
On Orcabase AI, everyone in your organization shares one credit. So that one busy day can’t use it all, it has three limits:
| Limit | Resets |
|---|---|
| Daily | Every day at 00:00 UTC |
| Weekly | Every Monday at 00:00 UTC |
| Monthly | On the 1st of each month, at 00:00 UTC |
- Usage is shown as a percentage of each limit. Once half of a limit is used, the chat header shows a badge, like “62% of today’s credit”, which turns amber at 80%.
- When a limit is used up, the agent pauses and says which limit and when it resets. It picks up again after that. An answer already in progress can go slightly over.
- How much credit your organization has depends on your plan. To change it, message us, or switch to your own key in Settings → Agent.
Keeping costs down
- Certified metrics make answers shorter: the agent asks for a metric by name instead of exploring tables.
- Pick a smaller model for simple questions. Every answer says which model it used, so it’s easy to compare quality.
- Steps per question caps a single question, and the agent asks before going further.
- Start a new chat for a new topic. Long chats send more history with each question.
Something unclear or missing? Troubleshooting covers the common errors, or message us.