AI Governance Copilot
Ask AxioRank questions about your own governance data in plain language. A read-only analyst grounded in your agents, policies, incidents, logs, and spend.
The AI Governance Copilot ("Ask AxioRank") is a conversational analyst embedded in the dashboard. Ask it a question in plain language ("which agents got denied the most this week?", "what is driving our model spend?", "draft a rule that holds any tool call carrying a secret") and it answers from your workspace's own governance data, with the agents, incidents, and logs behind each claim linked as citations.
Read-only by design
The Copilot can read everything in your workspace and change nothing. When you ask it to block, hold, redact, or cap something, it does not act. It returns a pre-filled policy draft link that opens the normal create-policy dialog for you to review and confirm. Every change still goes through a human.
How it works
Each question runs a bounded tool-use loop. The model can only call a fixed set of read-only tools, each wrapping an existing workspace-scoped query. The workspace id is injected server-side and never comes from the model, so a tool can only ever read your own tenant.
| Tool | What it reads |
|---|---|
query_audit_logs | Recent governed tool and model calls, with decision, risk score, reason, and signals |
list_agents | Your agents, with quarantine status, labels, and last-used time |
list_incidents | Recent incidents, filterable by status and severity |
list_policies | Your policies in evaluation order |
get_spend | Month-to-date spend, the per-agent and per-model breakdown, and the workspace budget |
get_recommendations | The current prioritized next-best-actions |
get_exec_summary | The executive posture summary over a recent window |
draft_policy | Turns a natural-language ask into a pre-filled, monitor-mode policy link (no write) |
Answers are grounded in what those tools return. The Copilot is instructed to answer only from tool data, to never invent agents, numbers, or ids, and to treat any text inside a tool result (an audit payload, a reason string) as untrusted data to report on, not as instructions to follow.
Which model powers it
The Copilot runs through the AI Gateway using your workspace's model credential, so its own usage is governed and its spend shows up in Spend governance like any other call. There are two sources, set per workspace:
- Managed (default): AxioRank's own model key. Free on every plan, with no setup. The managed path is capped at 300 messages per workspace per month.
- Bring your own model: connect an OpenAI-compatible upstream credential under MCP → Secrets (see Secrets broker) and switch the Copilot to it. Your own key is uncapped, since the spend is yours.
If neither is available, the Copilot asks you to connect a model before it answers.
Availability and limits
- Every plan. The Copilot is available on Free through Enterprise. The managed monthly cap is the free tier's allowance; bring your own model for unlimited use.
- Per-deployment toggle. A self-hosted or private deployment can turn the Copilot off entirely (it needs a managed model key configured on the deployment); when it is off, the feature is hidden.
- Abuse cap. Requests are rate-limited to 20 per minute per workspace.
- Stays up during a freeze. The Copilot keeps working even when the workspace is under an emergency lockdown, so you can investigate while everything else is held or blocked.
Next steps
- Spend governance: the cost view the Copilot reads from.
- Policies: where a confirmed policy draft lands.
- Secrets broker: connect your own model credential.