Custom Field Classification (Response Workflows)

Response Workflows can now automatically set Zendesk custom fields based on ticket content and workflow context. This saves time, improves consistency, and ensures key data is captured without manual effort.

What makes this different?

Unlike manual field rules or basic automation, Stylo's Custom Field Classification sees the complete workflow context, not just the raw ticket:

• The generated response your workflow created
• Integration data fetched during the workflow (Shopify orders, CRM records, etc.)
• OCR results extracted from ticket attachments
• The original ticket content (subject, description, comments)

This means the Stylo makes smarter decisions because it has the same information a human agent would use. For example, instead of guessing order status from a customer's message, it can see the actual Shopify order data.

Common use cases

• Detect customer language from their message
• Classify sentiment (positive, neutral, negative)
• Identify issue type or product category
• Extract order status using integration data
• Set priority level based on urgency signals
• Capture dates (delivery, return, appointment)
• Tag account type (VIP, trial, enterprise)

How to set it up

1. Open Response Workflows and edit a workflow
2. Go to Ticket Actions
3. Click Add Action > Set Custom Field (AI)
4. Choose the custom field to populate
5. Write a classification prompt that tells the AI what to extract
6. Save the workflow

When the workflow runs, the Stylo analyzes everything available and automatically sets the field value.

What Stylo can see

Original ticket data

• Subject and description
• All comments and replies
• Ticket metadata (tags, brand, priority, etc.)

Workflow outputs

• The generated response
• Shopify order data (if fetched)
• Customer records from integrations
• OCR text from attachments
• Results from previous workflow steps

Existing custom fields

• Values already set on the ticket
• Account and organization data

This rich context enables accurate, intelligent field classification that adapts to each unique situation.

Real world examples

Example 1: E-commerce Order Status

Scenario: Customer asks "Where's my order?" and your workflow fetches Shopify data.

• Field: order_status (dropdown)
• Prompt: "Based on the ticket content and Shopify order data, select the current order status: pending, shipped, delivered, cancelled"
• Why it works: Stylo sees both the customer's question AND the actual Shopify fulfillment status, so it sets the field accurately even if the customer's description is vague

Example 2: Smart Priority Detection

Scenario: Automatically triage tickets based on multiple signals.

• Field: priority_level (dropdown)
• Prompt: "Analyze the customer's tone, issue urgency, and account value. Select priority: critical, high, medium, low. Critical = service outage or angry VIP customer"
• Stylo considers context humans might miss like an escalating tone combined with account status

Example 3: Language Detection

Scenario: Route tickets to language-specific teams.

• Field: customer_language (dropdown)
• Prompt: "Detect the primary language in the customer's message. Return one of: english, spanish, french, german, other"
• Why it works: Simple and reliable for routing logic

Example 4: Refund Request Detection

Scenario: Flag tickets that need finance team review.

• Field: refund_requested (checkbox)
• Prompt: "Set to true only if the customer explicitly requests a refund or chargeback. Mentions of returns without refund requests should be false"
• Why it works: Clear true/false condition prevents false positives

Writing effective prompts

General tips

• Be specific and concise: Focus on one clear task
• Define exact output format: Especially for dropdowns, list valid options
• Provide decision criteria: Explain how the AI should decide
• Give examples if helpful: "e.g., 'I want my money back' = refund requested"
• Set boundaries: "Only if...", "Never when...", "Always choose..."

What NOT to do

❌ "Figure out what the customer wants" (too vague)
✅ "Identify if customer requests: refund, exchange, or repair"

❌ "Set the right category" (doesn't define "right")
✅ "Choose category based on main topic: billing, technical, or general_inquiry"

❌ "Use good judgment" (subjective)
✅ "Rate urgency 1-5 where 5 = customer threatens to cancel or reports service outage"

Tips by field type

Text (single line)

Best for short, simple values like names, IDs, or single-word classifications.

Prompt example:
"Extract only the product SKU mentioned by the customer (e.g., ABC-123)"

Tips:

• Ask for plain text without formatting
• Specify length limits if needed
• Tell AI what to do if value is unclear ("return 'unknown'")

Multi-line text (textarea)

Best for summaries, notes, or explanations.

Prompt example:
"Summarize the customer's main issue in 1-2 sentences for agent handoff notes"

Tips:

• Specify desired length (sentences, words, characters)
• Use for context that agents need to read
• Good for "reason for return", "issue summary", etc.

Dropdown (single select / tagger)

Best for categorization where only one value applies.

Prompt example:
"Select the primary issue category. Choose exactly one: billing, shipping, product_quality, account_access, other"

Tips:

• List exact option names from your Zendesk dropdown
• Use "choose exactly one" to prevent confusion
• Provide decision criteria for each option
• Add fallback option like "other" or "unclear"

Advanced example:

Classify the ticket type:
- `billing_dispute`: Customer challenges a charge
- `billing_question`: Customer asks about pricing or invoice
- `payment_failed`: Customer can't complete payment
- `refund_request`: Customer wants money back
Choose exactly one.

Multi-select

Best when multiple values can apply simultaneously.

Prompt example:
"Select all problems mentioned by the customer: damaged_item, missing_item, wrong_item, late_delivery. Multiple selections allowed"

Tips:

• List exact option names from Zendesk
• Say "multiple selections allowed" explicitly
• Use when issues genuinely overlap (e.g., "damaged AND late")

Numeric (integer / decimal)

Best for counts, amounts, scores, or ratings.

Prompt example:
"Return the order total amount as a decimal number (e.g., 49.99). If not mentioned, return 0"

Tips:

• Specify units and decimal places
• Define fallback for missing values
• Use integers for counts, decimals for money
• Good for "quantity ordered", "satisfaction score 1-10", etc.

Checkbox (true/false)

Best for yes/no conditions or flags.

Prompt example:
"Set to true only if the customer is a repeat complainant (3+ tickets in 30 days with negative sentiment). Otherwise false"

Tips:

• State condition very clearly
• Use "only if" to prevent false positives
• Specify edge cases ("unknown = false")

Date / DateTime

Best for capturing specific dates mentioned in tickets.

Prompt example:
"Extract the customer's requested delivery date in YYYY-MM-DD format. If no date mentioned, return null"

Tips:

• Specify exact format (YYYY-MM-DD, MM/DD/YYYY, etc.)
• Tell AI how to handle relative dates ("next Tuesday" = calculate actual date)
• Define fallback for missing dates
• Use DateTime fields for appointments with times

Tags / ID fields

Best for selecting from a fixed set of IDs or codes.

Prompt example:
"Identify the product category using these codes only: CAT-A (electronics), CAT-B (clothing), CAT-C (home goods), CAT-D (other)"

Tips:

• Provide complete mapping of codes to meanings
• Use strict list with "these codes only"
• Good for SKUs, region codes, team IDs, etc.

How it works in a workflow

Stylo Custom Field Classification runs automatically when the workflow executes. Here's a typical flow:

1. Workflow starts (ticket created or updated)
2. Fetch integration data (e.g., Shopify order lookup)
3. Generate response to customer
4. Stylo sets custom field and sees all above context
5. Add tags based on field value
6. Set ticket status to "Pending"
7. Post response to ticket

The action can be placed anywhere in the workflow and will use all context available at that point.

Combining with other actions

Custom Field Classification works alongside any other ticket actions:

Example workflow:

1. Fetch Shopify order
2. OCR scan of attachment
3. Generate response
4. Set custom field "order_status" (AI), sees all above
5. IF order_status = "delivered" THEN add tag "delivered"
6. IF order_status = "cancelled" THEN set priority "high"
7. Post response to ticket

You can even use the AI-set field value in conditional logic for subsequent actions.

Troubleshooting

Stylo picks the wrong value

Problem: Field is set but the value doesn't match the ticket
Solution:

• Make your prompt more specific with clear criteria
• Add examples of each option: "urgent = customer threatens to cancel"
• Check if the AI had the right context (was integration data fetched?)
• Review recent tickets and refine prompt based on patterns

Field stays empty

Problem: Workflow runs but field is never populated
Solution:

• Prompt may be asking for information that isn't in the ticket
• For dropdowns, ensure option names in prompt match Zendesk exactly
• Try adding a fallback option like "unknown" or "needs_review"
• Stylo may be unable to determine a clear answer from ambiguous content and make a judgement not to respond

Inconsistent results

Problem: Same type of ticket gets different field values
Solution:

• Add constraints: "Always choose exactly one option"
• Be more explicit about edge cases: "If unclear, choose 'other'"
• Provide decision tree: "If X mentioned, choose A. If Y mentioned, choose B"
• Consider if variation is actually correct (context differences)

Dropdown shows unexpected value

Problem: Field is set to a value that doesn't exist in Zendesk
Solution:

• List exact option names from Zendesk in your prompt (case-sensitive)
• Add "use these options only" to your prompt
• Check that Zendesk field hasn't been modified since workflow was created

Best practices

✅ Test before deploying

• Run workflow on sample tickets representing different scenarios
• Check field values are accurate
• Adjust prompts based on results

✅ Use objective criteria

• Stylo works best with clear, factual conditions
• Avoid subjective judgment ("seems urgent", "appears angry")
• Use measurable signals ("keywords: refund, cancel, chargeback")

✅ Provide fallback options

• Always include "other", "unknown", or "needs_review" for dropdowns
• Tell Stylo what to do when information is missing
• Prevents workflow failures from ambiguous content

✅ Leverage workflow context

• If you're fetching Shopify data, reference it in prompts
• Use the response draft to guide classification
• The more context, the better the decisions

✅ Keep prompts focused

• One field = one task
• Don't ask Stylo to "figure out" multiple things at once
• Create separate actions for separate fields

❌ Don't use for external validation

• Stylo can't check live inventory or verify account balances
• Not suitable for "does order exist?" (use integration lookup instead)
• Best for interpretation and categorization

❌ Don't rely on it for critical compliance

• Use manual review for legal/financial decisions
• AI is very good but not 100% perfect
• Consider adding "requires_review" flag for sensitive classifications

Limitations

• Works best with clear criteria – Subjective or nuanced judgment may vary
• Depends on available context – Can only use what's in the ticket and workflow
• Not for real-time validation – Can't check external systems (use integrations for that)
• May not set value when uncertain – Ambiguous tickets may result in empty fields
• Language-dependent – Best performance with English; other languages supported but may vary

Getting help

Need assistance crafting effective prompts?

• Send us your field names and types – We'll provide tailored prompt suggestions
• Share sample tickets – We can help you tune prompts for your specific use cases
• Review existing prompts – We'll analyze what's working and suggest improvements
• Request custom examples – Tell us your workflow and we'll build prompt templates

Contact support with "AI Custom Field" in the subject line and we'll help you get the most out of this feature.

FAQ

Q: Can I use this for multiple fields in one workflow?
A: Yes! Add multiple "Set Custom Field (AI)" actions. Each runs independently with access to the full workflow context.

Q: What happens if the AI is uncertain?
A: If the AI cannot determine a clear answer, it may leave the field empty rather than guessing. Make your prompts more specific or add fallback options like "unknown" to handle these cases.

Q: Can I override Stylo - set values manually?
A: Yes, agents can always manually edit custom fields after the workflow runs.

Q: How fast does this run?
A: AI classification typically adds 1-3 seconds to workflow execution time, depending on field complexity.