Feedback, Explanations, and Control

The Feedback Triad

Effective human–AI interaction rests on three connected mechanisms: feedback from the system to the user, explanations of system behaviour, and controls that let the user intervene.

Without feedback, the user is blind. Without explanations, the user cannot calibrate trust. Without controls, the user is helpless. These three elements must be designed together.

Feedback: What Happened

Feedback confirms that an action was taken and communicates the outcome. For agentic systems, feedback must be:

• Timely — delivered within the user’s attention window. Batch summaries work for low-urgency actions; real-time notification is needed for impactful changes.
• Specific — “I archived 3 emails” is better than “I cleaned your inbox.” Include which emails.
• Actionable — link to the outcome so the user can review, undo, or modify.

Feedback type	When to use	Example
Toast notification	Quick, low-importance actions	"Meeting moved to 3 PM"
Activity log	All actions, available on demand	Timestamped list with before/after
Digest	Batch of low-urgency actions	"Here is what I did this morning"
Exception alert	Action outside normal bounds	"I declined an invite on your behalf"

Explanations: Why It Happened

Explanations serve two purposes: building trust and enabling correction. A good explanation answers:

• What the system did.
• Why it made that choice (which inputs drove the decision).
• What if — counterfactuals that help the user understand boundaries (“If the amount were over $500, I would have asked first”).

Explanation Fidelity

Not every decision needs a full explanation. Match explanation depth to impact:

Control: Intervention Mechanisms

Users need multiple ways to intervene:

Control	Scope	Granularity
Undo	Single action	Reverses the last operation
Pause	Agent-wide	Suspends all autonomous actions until resumed
Override	Single decision	User changes the agent's choice mid-task
Recategorise	Pattern-based	"This is not spam" — retrains future behaviour
Revoke permission	Capability-wide	"Stop accessing my calendar"

The Undo Problem

Undo in an agentic system is harder than in traditional software because actions may have irreversible side effects (a deleted account, a sent message). Design for:

• Soft actions first — prefer tentative operations where possible (draft mode, hold before send).
• Compensating actions — if undo is impossible, offer a compensating action (“Re-send invitation to the person you removed”).
• Confirmation for irreversible actions — require explicit user confirmation before any irreversible operation.

Feedback Loops in Learning Systems

When agents learn from user feedback, every correction trains the model. This creates a feedback loop that can amplify biases or drift over time.

Designing the Critique Interface

Users should be able to tell the agent not just what went wrong but why:

• One-click corrections — “Wrong category” with a dropdown of alternatives.
• Free-text critique — “Next time, prioritise the CEO’s calendar over internal meetings.”
• Preference teaching — “I prefer Thursday for recurring meetings.”

Store preferences explicitly so users can review and edit them later, rather than encoding them as opaque model weights.

Key Takeaways

• Feedback, explanations, and control form a triad — all three must be present.
• Match explanation depth to action impact.
• Provide layered controls: undo, pause, override, and permission revocation.
• Design compensating actions for irreversible operations.
• Make user feedback visible, editable, and disconnected from model internals when possible.

---

Next: Chapter 5 — Safety, Robustness, and Failure Modes