Product · AI Actors
Give AI workloads an owner, a budget, and an off switch.
An AI actor is a person, application, or agent using AI. Assign its owner and purpose, then set approved models, budgets, and access controls for the traffic it routes through Tallin.
One actor record
Know who can change the workload.
This sample support agent has an owner, a purpose, and configured limits. Its company content policy remains separate from what the Gateway enforces.
- Owner
- Priya Nair · VP Support
- Purpose
- Draft support replies for employee review.
- Model scope
- Approved models on configured gateway routes.
- Monthly budget
- $2,000 · automatic cutoff on new routed requests at the configured threshold.
- Capture
- Metadata-only. Full-content capture is configurable.
- Content policy
- Customer PII prohibited by company policy. Content-level enforcement is on the roadmap.
Actor controls
Set limits before the workload runs.
- Scope
- Choose approved provider routes and models. Requests outside that scope can be blocked at the Gateway.
- Budget
- Set alert thresholds and an automatic cutoff for routed requests. Provider-direct spend is not stopped by gateway caps.
- Kill switch
- Disable one actor's Tallin-mediated model access. This does not stop tools or external systems the actor can reach independently.
- Shadow mode
- Simulate supported policy controls before enabling enforcement, then review what would have been affected.
Control history
See why a request stopped.
| Time | Actor | Trigger | Outcome |
|---|---|---|---|
| 09:41 | deploy-agent-04 | Budget alert threshold | Owner alerted |
| 10:07 | deploy-agent-04 | Budget limit reached | New routed requests paused |
| 10:24 | support-triage-agent | Model outside approved scope | Gateway request blocked |
| 11:15 | invoice-drafter-app | Owner used kill switch | Tallin-mediated model access disabled |
Connected provider APIs may observe calls outside the Gateway. Observation does not extend these controls to direct calls or tool and MCP side effects.
Start with one accountable workload.
Assign the owner, choose a gateway route, and test the model scope and budget before expanding.