Which signal is the weakest reason to escalate to a human?
Why D is the answer
The model reported a high confidence score for its own answer
Model-reported confidence is a poor control signal. Prefer explicit rules such as a person being requested, limits being reached, or policy requirements.
In detail
- Self-reported confidence from a model is a poor signal, because it may not reflect real accuracy.
- Strong triggers are explicit and easy to check: a request for a person, a limit reached, or a policy requirement.
- Prefer signals that you can check, not the model's view of itself.
Use case
An illustrative scenario, not a real incident.
A team escalates when the model reports low confidence, yet some wrong answers come with high confidence. Rules for requests for a person, turn limits and policy approvals prove more dependable.
See it in different ways
flowchart LR
A["Agent"] --> B{"Limit reached or person requested?"}
B -->|Yes| C["Handoff summary"]
C --> D["Human agent"]
B -->|No| AsequenceDiagram participant U as Customer participant A as Agent participant H as Human U->>A: I want a person A->>H: Summary, facts and steps taken H-->>U: Picks up the case
mindmap
root(("Escalation"))
n1["Strong"]
n2["Customer asks for a person"]
n3["Limit reached"]
n4["Policy requires it"]
n5["Weak"]
n6["Model-reported confidence"]Key point: Escalate on checkable rules.
Why the other options are wrong
- A. Policy requires human approval for this type of action
- A policy requirement is mandatory, so it is not weak.
- B. The customer asked for a person
- A direct request for a person is a strong trigger.
- C. The agent reached its turn limit without finishing the task
- A reached turn limit is a clear signal that the agent needs help.