Security
Usage metadata. Never your prompts.
Arbitir is designed to measure economic performance without turning AI governance into employee surveillance.
What is never read
Arbitir ingests token counts, model names, timestamps, and dollar amounts. It is designed so prompt content, completions, and conversation text are not required for spend attribution or the allocation model. There is no code path that reads them, and that is enforced in the schema rather than promised in a policy.
No connection can read a prompt, a completion, a conversation, a Slack message, a CRM note, a contact or an email body. Every credential is scoped to usage, cost, owners, deals and member lists. Where a service offers a broader key, the broader key is rejected at connect time.
Connections, credential reads and every change to AI capital are auditable events. The economic ledger and the decision workflow stay inspectable without making prompt surveillance a prerequisite.
How it is built
Four rules the product is held to.
Each of these is enforced in the database or the code, not in a document. They are the same rules the product's own help guide describes to customers.
- 01
Credentials in a vault
A key you connect is validated once against the live API and written to an encrypted vault. The application database stores the name of the secret, never the secret, and every read of it writes an audit row.
- 02
Tenancy at the database
Every table carries the organization it belongs to, and the database itself refuses a read or write across organizations before any application code runs. Every server action checks membership again.
- 03
An audit trail for everything
Every connection, every credential read, every budget change, every capital decision, every identity mapping and every export is a row naming who did it and when.
- 04
One credential writes
Every connection reads, with one exception: the Claude Enterprise spend limits key, which can set a member's monthly limit. It is used only when a capital request is approved or a temporary approval ends, and each use is on the ledger.
Subprocessors
The services that run Arbitir.
Arbitir Inc uses the following providers to deliver the service. Each handles only the data named beside it. Last updated September 14, 2026.
| Provider | Purpose | Data it handles |
|---|---|---|
| Supabase | Database, authentication, and the encrypted vault that holds connected credentials | All customer data: usage metadata, people and teams, KPI figures, budgets, the ledger, and credentials |
| Vercel | Application hosting, scheduled jobs, and page-view analytics on the marketing site | Requests to the application; page views and visitor counts on the marketing site, without cookies |
| Sentry | Error tracking | Error reports carrying record identifiers. No names, emails, or content |
| Axiom | Operational logs | Sync results and decision points carrying identifiers. No names, emails, or content |
| Resend | Transactional email: invitations, budget notices, weekly reviews, access confirmations | Recipient email addresses and the content of each notice |
The systems a customer connects, such as Anthropic, OpenAI, HubSpot, Salesforce and Slack, are the customer’s own accounts, read on the customer’s behalf with the customer’s credentials. They are not subprocessors of Arbitir.
Logs, roles, and support access
Operational logs carry identifiers, never names or emails. Errors from a provider are logged by code, and the provider’s own message is never shown in the product.
An owner or admin can connect providers, map identities, set budgets, import people, change settings and invite members. A member sees only their own record, and the database enforces it: their account cannot read anyone else’s figures.
Arbitir’s own staff hold owner access to every organization so support can see exactly what a customer sees. It is an ordinary membership rather than a back door: the same row-level rules apply, the same audit rows are written, and every action is recorded against the staff member’s named account.
Questions about security, or a report of a vulnerability, go to hello@arbitir.ai.
Start allocating AI spend toward what works.
We grant access to a select number of companies each month, based on fit. Describe how AI spend is governed today and we will respond within two business days.