Why Panel

The Why Panel shows exactly how AI arrived at its decision. This transparency helps you trust (or correct) AI actions.

Accessing the Why Panel

From Approval Queue

1. Open an approval item
2. Click "Why?" or "See reasoning"
3. Panel slides in from right

From Ticket Detail

1. Open any ticket
2. Find an AI action in the timeline
3. Click "Why did AI do this?"

What the Why Panel Shows

Intent Recognition

How AI interpreted the customer's request:

Detected Intent: Refund Request
 
Customer is asking for a refund on order #12345.
They mentioned the product arrived damaged.
 
Confidence: 95%

Information Sources

What data AI used:

Reasoning Steps

The logic chain:

1. Identified request as refund for damaged item
2. Found refund policy in knowledge base
3. Checked order status via Shopify
4. Verified within 30-day return window
5. Determined eligible for full refund
6. Generated response with refund steps

Confidence Score

How certain AI is about this decision:

Alternatives Considered

Other approaches AI evaluated:

Alternatives:
1. Offer replacement instead of refund (rejected: customer specifically asked for refund)
2. Escalate to human (rejected: straightforward policy case)
3. Request photos first (rejected: policy allows refund without photos under $50)

Using the Why Panel

To Build Trust

As you review Why Panels:

• See AI follows your policies
• Understand its decision process
• Build confidence in its capabilities

To Improve AI

When AI gets it wrong:

• Identify where reasoning failed
• Add missing knowledge base content
• Adjust training examples

To Learn Patterns

Notice patterns in AI decisions:

• Common question types
• Frequent policy applications
• Areas needing more KB content

Example Why Panels

Example 1: Simple Response

Intent: Business Hours Inquiry
Confidence: 99%
 
Sources Used:
- Knowledge Base: "Business Hours" article
 
Reasoning:
Customer asked about business hours.
Found exact answer in knowledge base.
Generated direct response with hours.
 
No alternatives considered - straightforward query.

Example 2: Refund Decision

Intent: Refund Request (Damaged Item)
Confidence: 87%
 
Sources Used:
- Customer message: Mentioned damaged item
- Shopify: Order #12345 - $75.00
- Knowledge Base: "Return Policy" article
- Customer History: 2 previous orders, no returns
 
Reasoning:
1. Customer reports damaged item
2. Order found: $75.00, delivered 5 days ago
3. Within 30-day return window
4. Refund policy allows for damaged items
5. No previous abuse of return policy
6. Eligible for full refund
 
Alternative Considered:
- Request damage photos first
  Rejected: Policy only requires photos for items over $100

Example 3: Escalation

Intent: Account Closure + Legal Threat
Confidence: 72%
 
Sources Used:
- Customer message: Contains "lawyer" and "cancel my account"
- Customer History: VIP customer, 3 years
 
Reasoning:
1. Customer mentions potential legal action
2. Also requesting account closure
3. VIP customer with significant history
4. Combination warrants human attention
5. ESCALATING to human review
 
Alternatives Considered:
- Auto-respond with standard closure process
  Rejected: Legal mention triggers escalation policy

Interpreting Confidence Scores

High Confidence (90%+)

AI is very sure. Usually:

• Clear, simple request
• Exact match in knowledge base
• Standard scenario

Medium Confidence (70-89%)

AI is confident but used some inference:

• Multiple possible interpretations
• Partial knowledge base match
• Some ambiguity resolved

Lower Confidence (50-69%)

AI is uncertain:

• Ambiguous request
• Missing information
• No clear KB match
• May need human review

Very Low Confidence (<50%)

AI usually escalates these:

• Very complex issue
• Contradictory information
• Sensitive topics
• Outside training scope