Understanding Intent Predictions
An Intent represents a specific customer support request, such as tracking an order, requesting a refund, or updating payment details. The way an Intent is defined directly impacts how well an AI system predicts customer needs and provides accurate responses. As the workflows are triggered only when the successful prediction has been made, Intent accuracy is crucial for ensuring the correct Solve responses.
Key Factors Influencing Intent Prediction
Intent predictions rely on three main elements:
- Intent Name – The high-level label for the Intent.
- Intent Description – Additional context to refine AI understanding.
- Training Phrases – Examples of real customer queries that should trigger the Intent.
Each of these has different weights, meaning some contribute more significantly to the AI’s decision-making than others.
| Element | Required | Weight | Purpose |
| Intent Name | Yes | Low | Provides a clear goal for the Intent. |
| Training Phrases | Yes | High | Helps AI understand customer expressions and variations. |
| Intent Description | Optional | Low | Augments AI understanding but has less influence than Names and Training Phrases. |
Best Practices for High-Accuracy Intent Predictions
Use clear and specific Intent Names
- Ensure the name directly reflects the customer’s request (e.g., "Track Order Status" instead of just "Order Help").
- Avoid vague or overlapping names that could confuse the AI.
Ensure Training Phrases cover multiple variations
- Provide a diverse set of short, natural phrases customers might use.
- Avoid redundant or overly similar phrases.
- Example for "Track Order Status":
- “Where is my order?”
- “Track my shipment”
- “I can’t find my package”
Keep Intent Descriptions concise and complementary
- Should not conflict with the Intent Name or Training Phrases.
- Can include edge cases or additional details that refine AI behavior.
Avoid overlapping Intents
- Ensure similar Intents have distinct names, training phrases, and descriptions.
- Example: Instead of having both “Track Order” and “Order Status”, consolidate them into one well-defined Intent.
Regularly review and update Intents
- Monitor AI misclassifications and refine training data accordingly.
- Adjust training phrases to match emerging customer language trends.
Dos and Don’t with Examples
| Category | Good Example (✅ Well-Defined) | Bad Example (❌ Poorly Defined) | Why the Bad Example is Ineffective |
| Track Order Status (Bad Intent Name) |
Intent Name: Track Order Status Description: The customer wants to check the delivery status of an order. Training Phrases: - "Where is my order?" - "Track my package" - "Check order delivery status" - "My order hasn’t arrived" |
Intent Name: Order Help Description: Customer needs assistance with an order. Training Phrases: - "Order issue" - "Need help with my order" - "Check something on my order" |
❌ Bad Intent Name: “Order Help” is too broad—customers may need tracking, refunds, or modifications. ❌ Confusing AI: AI may misclassify between different order-related issues. |
| Cancel Subscription (Bad Intent Description) |
Intent Name: Cancel Subscription Description: The customer wants to stop a recurring subscription. Training Phrases: - "Cancel my subscription" - "Stop my membership" - "How do I unsubscribe?" - "End my monthly plan" |
Intent Name: Cancel Subscription Description: The customer has issues with payments, account settings, or wants to make changes to their subscription. Training Phrases: - "Cancel my subscription" - "Stop my membership" - "How do I unsubscribe?" - "End my monthly plan" |
❌ Bad Intent Description: Describes multiple concerns (payment issues, account settings), leading to AI confusion. ❌ Mismatch with Training Phrases: The AI may misclassify it under billing support rather than subscription cancellation. |
| Refund Request (Bad Training Phrases) |
Intent Name: Refund Request Description: The customer wants a refund for a recent purchase. Training Phrases: - "I need a refund" - "Request money back" - "Return and refund my order" - "Get my money back" |
Intent Name: Refund Request Description: The customer wants a refund for a recent purchase. Training Phrases: - "I need help with my order" - "Check my account balance" - "Resolve my payment issue" - "Fix my charge" |
❌ Bad Training Phrases: The AI won’t clearly recognize refund requests because phrases refer to general order/account/payment problems. ❌ Misclassification Risk: AI may trigger billing or order tracking intents instead. |
| Update Payment Method (Bad Intent Name) |
Intent Name: Update Payment Method Description: The customer needs to update their stored payment details. Training Phrases: - "Change my credit card" - "Update my billing info" - "Add a new payment method" - "Modify my card details" |
Intent Name: Payment Issues Description: The customer has questions about their billing or payment setup. Training Phrases: - "Change my credit card" - "Update my billing info" - "Add a new payment method" - "Modify my card details" |
❌ Bad Intent Name: "Payment Issues" is too vague—could refer to refunds, failed payments, or disputes. ❌ Confusion with Other Intents: AI may classify this under "Refund Request" or "Billing Inquiry." |
| Reset Password (Bad Intent Description) |
Intent Name: Reset Password Description: The customer is unable to access their account and wants to reset their password. Training Phrases: - "Forgot my password" - "Reset my login" - "Can’t log in, need a new password" - "Recover my account" |
Intent Name: Reset Password Description: The customer has login problems and needs help accessing their account. Training Phrases: - "Forgot my password" - "Reset my login" - "Can’t log in, need a new password" - "Recover my account" |
❌ Bad Intent Description: Too general—"login problems" could also mean account lockout, two-factor authentication issues, or username recovery. ❌ Overlaps with Other Intents: AI may misclassify this under broader login-related issues rather than password resets. |
| Billing Inquiry (Bad Training Phrases) |
Intent Name: Billing Inquiry Description: The customer has questions about their invoice or charges. Training Phrases: - "Why was I charged?" - "Explain my bill" - "Unexpected charge on my account" - "Help with my invoice" |
Intent Name: Billing Inquiry Description: The customer has questions about their invoice or charges. Training Phrases: - "Check my order status" - "I need a refund" - "My payment failed" - "Track my shipment" |
❌ Bad Training Phrases: Includes phrases unrelated to billing (tracking orders, refunds, failed payments). ❌ Misclassification Risk: AI may trigger the wrong intent, leading to customer frustration. |
Comments
Article is closed for comments.