Costs
What Copilot Studio agents cost, in Copilot Credits and money, and how each agent tracks against its budget.
Estimated is every agent action KEY2AI allowed to run, at Microsoft's rate. Billed is what Microsoft charged, loaded from Microsoft.
The estimate leaves out what KEY2AI cannot see (answers, knowledge grounding, agent flows), so billed is usually higher.
Agents
| Agent | Cost centre | Actions | Estimated | Billed | Monthly budget (credits) | Budget used | Where it goes |
|---|
By cost centre
| Cost centre | Agents | Estimated | Billed |
|---|
Billed, by kind of use
| Feature | Credits |
|---|
Estimated, by tool
| Tool | Credits |
|---|
Loads of billed credits
| When | Source | By | Days | Rows |
|---|
Redaction
The KEY2AI gateway sits between agents' tools and your systems. It masks sensitive values in what an agent receives
(for example
[PERSON_1], [AU_TFN_1]) and puts real values back when the agent sends a mask on to another system through the gateway.
Only counts and data types are recorded here: never the content, never the values.Systems behind the gateway
| Route | Goes to | Calls | Items masked | Restored | Not inspected | Stopped | Data types masked |
|---|
Recent calls
| Time | Route | Call | Result | Masked | Agent | Check time |
|---|
Shadow AI
AI services staff use outside approved tools, reported by the KEY2AI browser extension.
Prompt text is never collected: only the service, the event and the types of sensitive data. Trends: Grafana → KEY2AI – Shadow AI & agents.
| Service | Status | Events | Users | Sensitive items | High risk |
|---|
Recent events
| Time | Service | Event | Sensitive | Action | Risk | User (pseudonym) |
|---|
Copilot agents
Every tool call a Copilot Studio agent plans is checked by the KEY2AI checkpoint before it runs.
Shadow records what would have been stopped. Enforce stops it. Pause stops every action of that agent.
No prompt text or values are kept: only the decision, the data types and counts. Trends: Grafana → KEY2AI – Copilot checkpoint.
Waiting for approval 0
Actions listed under an agent's approval tools stop here until someone decides. Approving lets the same request run when the user tries again.
| Request | Agent and action | Requested by | What it carries | Asked | Decision |
|---|
Recent decisions
What would have been stopped, by rule
| Rule | Would stop | Stopped |
|---|
Agents
| Agent | Tool calls (7 days) | Would stop | Sensitive data seen | Mode | Pause | Last change |
|---|
Recent decisions
| Time | Agent | Tool | Decision | Rule | Data types | Time taken |
|---|
Multi-agent orchestration
Router → specialist → reviewer. Hand-offs never raise privilege.
Needs the KEY2AI Reference Agent (demo).
MCP gateway
Endpoint
http://localhost:8810/mcp. Tools this role can see: