Skip to content
RezolveITDocumentation

Governance — guardrails, budgets and the external gateway

AI Assist ships with its own governance: content guardrails, a spend budget, and a usage/audit trail. An organization that already runs its own AI gateway can instead hand all of that governance to the gateway.

Guardrails

Section titled “Guardrails”

Every AI Assist input and output is screened against a policy with three categories:

Category What it catches Default action
Toxicity Offensive or profane content Block
Prompt injection Override / exfiltration attempts embedded in a message Block
Data leakage (PII) Personal data — identifiers, card numbers, API keys, bulk emails Flag (redact)

Each category can be independently enabled or disabled, and its action switched between Block (a safe refusal is shown, the underlying content is never displayed) and Flag (the matched part is redacted but the rest of the reply is still shown). Admins can also add custom toxicity keywords — extra plain words matched on word boundaries, on top of the built-in list.

Every blocked or flagged turn is recorded as a guardrail event, so it’s visible to a later review, and the same policy is what actually runs at request time — there’s one editor for it, reachable both from the AI Control Center’s Guardrails tab and the AI Assist Governance page, never two copies that could drift apart.

Budget

Section titled “Budget”

A monthly spend limit (in USD) can be set with one or more alert thresholds (percentages of the limit) that notify admins as spend approaches it. A limit of 0 means “track spend, don’t cap it.” Once a hard limit is set and month-to-date spend meets or exceeds it, further AI Assist requests stop calling the live model — they get a fixed message asking an admin to raise the limit, rather than continuing to spend.

Usage and audit

Section titled “Usage and audit”

Every AI Assist turn — question, skill run, or gateway call — is logged with its input, output, provider, latency, and (for the assistant) the scope decision and guardrail outcome. That data feeds the AI Control Center’s Conversations and Performance tabs, and a separate AI Assist Governance surface holds the interaction log and usage-anomaly alerts an admin can review.

The external AI gateway

Section titled “The external AI gateway”

Instead of calling model providers directly, an organization can route all AI Assist traffic through its own AI gateway — a self-hosted, OpenAI-/Anthropic-compatible proxy such as LiteLLM. This is configured from the AI Gateway card at the top of the Model providers tab:

  1. Turn on the master switch and choose a vendor (LiteLLM or a generic OpenAI-compatible gateway).
  2. Set the gateway’s base URL and which request format to send (Anthropic- or OpenAI-shaped).
  3. Paste the gateway’s key — it’s stored in the same encrypted vault as the native provider keys, never anywhere else, and never seen by the browser.
  4. Optionally map RezolveIT’s model ids to the gateway’s own virtual model names; anything unmapped falls back to a configured default model, or passes through unchanged.
  5. Use Test connection to confirm the gateway responds.

What changes when it’s on

Section titled “What changes when it’s on”

Turning gateway mode on hands governance to the gateway entirely:

  • Native guardrails, budgets, and usage logging are disabled — the request bracket that would screen input/output, check the budget, and log usage is skipped, because the gateway is expected to enforce its own guardrails, budgets, and logging.
  • The admin console tells you so. The Guardrails and Budget tabs (and the older Governance scope/log surface) switch to a read-only state with a banner explaining that governance now lives in the gateway — settings aren’t deleted, so turning the gateway back off restores them exactly as they were.
  • The native per-provider credential controls collapse to read-only on the Model providers tab, with a note that they aren’t used while the gateway is active.
  • The Overview tab’s routing line switches from the native-provider status to naming the active gateway and its host.
  • Fails safe. A half-configured gateway (turned on but no base URL) falls back to native routing rather than silently sending requests nowhere.

This is a deliberate trade-off: the organization takes on responsibility for configuring equivalent guardrails, budgets, and logging inside their own gateway. RezolveIT’s “Test connection” can confirm the gateway is reachable, but it can’t verify what the gateway’s own governance is configured to do.

Section titled “Related”