Build a Customer Support Intake Form With AI Routing

The category dropdown on a customer support form is a tax that everyone pays and nobody benefits from. The user has to translate their problem into your taxonomy ("is this a billing issue or an account issue?"). They get it wrong constantly. The support team has to re-route mis-categorized tickets, which adds delay and frustrates everyone. The dropdown exists to help routing, but it actively makes routing worse.

AI fixes this. The user describes their problem in natural language. An AI node reads the description, classifies the ticket, scores urgency, identifies the right team, and routes it — all before any human sees the form. The category dropdown disappears. Routing accuracy goes up. Time-to-first-response drops because hot tickets don't sit in a triage queue.

This is one of the highest-leverage AI applications in a customer-facing workflow. It is also one of the easiest to ship because the workflow architecture is simple, the prompts are forgiving, and the failure modes are recoverable (a misrouted ticket gets re-routed; nobody loses money).

This post walks through the full build — form structure, AI prompt, routing logic, and integration patterns for the major ticketing systems.

Why "Just Use a Dropdown" Doesn't Work

Three failure modes show up in every dropdown-based support workflow.

Users guess wrong. Studies of support form analytics show 25-40% of tickets land in the wrong category on first submission. Users see the categories your team uses internally and try to match — but they don't know the team's taxonomy, so they go with their gut. "My password isn't working" might be filed as "account," "technical," "login," or "other," depending on which category the user looked at first.

Categories don't capture priority. A category tells you which team should look at the ticket. It doesn't tell you whether to look now or in three days. "My website is down" and "I have a feature request" both file under "technical" but require very different responses.

Dropdowns add friction. Every dropdown is a small decision the user has to make. Forms with 4 dropdown selectors have ~15% lower completion rates than equivalent forms with 1 free-text field. Users abandon support forms more than any other form type, often because the form itself is annoying.

The AI alternative collapses categorization, prioritization, and routing into one step that happens automatically. The user describes the problem; the AI handles everything downstream.

What You Will Build

The finished workflow:

Support Form (subject + description + email + optional file upload)
  ↓
AI Text Generation (classify + prioritize + extract context)
  ↓
Branch (route by category + urgency)
  ├── Billing: Slack #billing + Zendesk billing queue
  ├── Technical urgent: PagerDuty + Slack #engineering-urgent + Linear ticket
  ├── Technical non-urgent: Linear ticket + email auto-response
  ├── Account: Slack #support + Zendesk account queue
  ├── Feature request: Linear feedback project + email "thanks for the suggestion"
  └── General: Zendesk general queue

The form is shorter than what the user is used to. The AI handles the work the dropdown used to do. Time-to-first-response on urgent issues drops from "whenever the on-call sees the email" to "within 60 seconds of submission." Time-to-resolution drops because tickets land in the right queue with the right priority and the right context.

Step 1: Build the Form

The form has fewer fields than a traditional support form, on purpose.

Required:

• Email (required, validated)
• Subject (text, 1-line summary)
• Description (textarea, the main content)

Optional:

• Attach screenshot or file (for visual bug reports)
• Customer ID (auto-filled if logged in, optional otherwise)

Removed (the AI handles these):

• Category dropdown
• Priority dropdown
• Sub-category nesting

If you have an authenticated user, also auto-capture: account tier, plan, account creation date, last activity. This metadata feeds into the AI's prioritization — a billing question from a $50K/year customer is different from one from a free-tier user.

For the form-design principles around minimal field count and how it affects completion, see 12 form design tips to double your completion rate. The general rule applies extra-strongly to support forms: friction is the enemy of getting help.

Step 2: Add the AI Classification Node

After the form submission, drop a Text Generation node onto the canvas.

Provider: OpenAI GPT-4.1-mini or Anthropic Claude Haiku — both work well for classification at low cost. The full-size models work too but the marginal accuracy gain is small for this use case.

Temperature: 0.1 — very low temperature for classification. You want consistent outputs for similar inputs.

Prompt:

You are a customer support routing expert. Classify the following support
request and produce structured output that downstream routing can act on.

Customer information:
- Email: {{supportForm.email}}
- Account tier: {{supportForm.accountTier|default:"free"}}
- Plan: {{supportForm.plan|default:"unknown"}}

Support request:
- Subject: {{supportForm.subject}}
- Description: {{supportForm.description}}

Categories:
- billing: payments, refunds, invoices, subscription changes, pricing
- technical: bugs, errors, broken features, login issues, performance
- account: account access, profile changes, data exports, deletion
- feature_request: requests for new features or improvements
- general: anything that doesn't fit above (questions, feedback, etc.)

Urgency levels:
- urgent: customer is blocked or experiencing data loss / security issue
- high: significant feature broken, blocking workflow but workaround exists
- medium: moderate issue, can wait 24 hours
- low: minor issue, question, or non-blocking feedback

Output JSON:

{
  "category": "<billing|technical|account|feature_request|general>",
  "subcategory": "<more specific descriptor, free-form>",
  "urgency": "<urgent|high|medium|low>",
  "sentiment": "<frustrated|neutral|positive>",
  "summary": "<1-2 sentence summary of the issue>",
  "suggestedNextStep": "<what should happen first, in 1 sentence>",
  "extractedDetails": {
    "errorMessage": "<exact error if mentioned, or null>",
    "browserOrPlatform": "<if mentioned, or null>",
    "stepsToReproduce": "<if described, or null>",
    "affectedFeature": "<feature name if identifiable>"
  },
  "redFlags": []
}

Urgency calibration:
- A free-tier user reporting a feature they can't use → medium
- An enterprise customer reporting the same → high
- Anyone reporting data loss, billing fraud, or security issue → urgent
- Anyone using language like "I'm canceling," "this is unacceptable," "lawsuit" → at least high, add to redFlags

Be precise about subcategory. "Login issue" is more useful than just "technical."
"Refund request for double charge" is more useful than just "billing."

A few notes on this prompt:

The customer-tier conditional logic is the hidden multiplier. Without it, AI treats every ticket the same. With it, the same complaint from an enterprise customer routes faster than from a free user. This isn't favoritism; it's appropriate ROI-based prioritization.

The extractedDetails block is what saves support agents the most time. Instead of reading the description to figure out what error message the user got, what browser they're on, and what they were trying to do, the agent sees a structured summary. They go straight to the diagnosis.

The suggestedNextStep field is a recommendation, not a directive. The agent can ignore it. But for new agents or for after-hours triage, the suggestion saves a few minutes of "what do I do with this?" thinking.

Step 3: Add the Branching Logic

Drop a Branch node after the AI Text Generation node. The branch reads the AI output and routes accordingly.

Branch path 1: Billing

• Condition: {{aiClassify.category}} equals billing
• Actions:
  • Create ticket in Zendesk billing queue with priority {{aiClassify.urgency}}
  • Post to Slack #billing with summary
  • Send auto-response email to user with expected response time

Branch path 2: Technical (urgent)

• Condition: {{aiClassify.category}} equals technical AND {{aiClassify.urgency}} equals urgent
• Actions:
  • Trigger PagerDuty incident
  • Post to Slack #engineering-urgent with full extracted details
  • Create Linear ticket with priority Urgent
  • Send auto-response with "we're on it" message

Branch path 3: Technical (non-urgent)

• Condition: {{aiClassify.category}} equals technical
• Actions:
  • Create Linear ticket with priority based on urgency
  • Post to Slack #bugs (low-noise channel)
  • Send auto-response

Branch path 4: Account

• Condition: {{aiClassify.category}} equals account
• Actions:
  • Zendesk account queue
  • If {{aiClassify.subcategory}} contains "deletion" or "GDPR," route to compliance team
  • Auto-response

Branch path 5: Feature Request

• Condition: {{aiClassify.category}} equals feature_request
• Actions:
  • Create entry in Linear "Customer Feedback" project (not the bug project — different team)
  • Send auto-response: "thanks for the suggestion, here's how we use feedback"

Branch path 6: General

• Condition: catch-all for anything not matching above
• Actions:
  • Zendesk general queue
  • Auto-response

The branching logic is similar to the AI lead qualification form and AI vision document verification workflows — score, classify, branch, differentiate. The architecture transfers across use cases. What changes is the prompt content. For broader context on these branching patterns, see workflow automation best practices.

Step 4: Configure the Auto-Response Templates

Auto-response emails are not "form submitted, we'll get back to you." They are differentiated by category and urgency, and they reference what the AI extracted.

For urgent technical issues:

Subject: We're on it — incident #{{ticketId}}

Hi {{firstName}},

We received your report about {{aiClassify.summary}} and have escalated it
to our on-call engineering team. You can expect a response within 30 minutes.

Reference: {{ticketId}}

In the meantime, here's what we know so far based on your description:
- Issue: {{aiClassify.summary}}
- Affected: {{aiClassify.extractedDetails.affectedFeature}}
- Severity: Urgent

If your issue is resolved before you hear from us, please reply with "resolved"
so we can close the ticket.

— Support Team

For feature requests:

Subject: Got your idea — {{aiClassify.subcategory}}

Hi {{firstName}},

Thanks for sending this in: "{{aiClassify.summary}}"

Feature requests like this go directly to our product team. We don't promise
a response on every one, but we read all of them and use them to inform
our roadmap. If we ship this or something similar, we'll email you.

If you want to add more context, just reply to this email.

— Product Team

For billing:

Subject: Re: {{supportForm.subject}}

Hi {{firstName}},

We received your billing inquiry: "{{aiClassify.summary}}"

A billing specialist will reply within 1 business day. Most billing issues
are resolved within 24 hours.

If this is urgent (e.g., a duplicate charge or fraudulent activity),
reply to this email with "URGENT" in the subject line and we'll prioritize.

— Billing Team

These templates are not generic. They reference the AI's understanding of the issue, set realistic response expectations based on category and urgency, and give the user a path to escalate if needed. For the underlying email-templating mechanics, see setting up email notifications for form submissions.

Step 5: Handle the Edge Cases

Real-world support traffic includes things the prompt doesn't anticipate. Here are the patterns that come up.

Spam or off-topic submissions. The AI usually catches these and sets redFlags. Add a branch path that routes red-flagged tickets to a "review" queue rather than the standard queues.

Multi-issue submissions. A user describes three problems in one ticket. The AI picks the most prominent for category, but a sub-summary captures all three. Configure your routing to handle this — typically by routing to the highest-priority team and letting them split the ticket if needed.

Non-English submissions. GPT-4.1 and Claude Sonnet handle 30+ languages well; Haiku and mini variants are weaker on non-English. If you support multilingual users, either upgrade to a stronger model for the classification node, or add a language-detection step that routes non-English submissions to a human-language-capable queue first.

Hostile or abusive language. Add explicit prompt instructions: "If the user uses profanity, threats, or harassing language, route to general queue with redFlags: ['hostile_language'] and do not generate a personalized auto-response. Use a neutral template."

Existing customer with prior tickets. If you can pass prior-ticket context to the AI, do it. The prompt can include "Previous tickets from this customer: [...]" and the classification becomes much smarter — repeat issues get higher priority, the AI can reference past resolutions in the suggested next step.

Image-based issues. Users sometimes paste a screenshot but type "see attached" with no description. Add a Vision node that runs first if the description is empty or near-empty. The Vision node extracts the visible content (error messages, UI state) and feeds it into the classification prompt. For the broader vision pattern, see how to use AI vision in forms to verify uploaded documents — same primitive, different application.

Numbers You Can Expect

Realistic numbers from teams that have shipped AI-routed support intake:

• Routing accuracy: typically 92-96% on the AI's first classification, vs. 60-75% on user-selected dropdown categories.
• Time to first response: drops 60-80% on urgent tickets because they bypass the triage queue and hit the right team directly.
• Triage time per ticket: drops from 2-4 minutes (read, categorize, prioritize, route) to under 10 seconds (read the AI's summary).
• Customer satisfaction (CSAT): typically lifts 5-15 points because users get faster, more relevant responses.
• Cost per ticket: $0.005-0.02 in AI inference cost. Negligible.
• Setup time: 4-6 hours for the first build, mostly in tuning the prompt for your team's specific category boundaries.

What This Connects To

The intake form is the front door to the support workflow. AI classification turns it from a triage bottleneck into a routing system. Once the routing works, the rest of the support workflow gets to be smarter:

• Automated first-pass responses can handle simple categories (billing FAQ, password reset) without ever escalating to a human.
• AI-drafted reply suggestions help agents respond faster to the tickets that do escalate.
• AI-powered survey analysis of post-resolution CSAT surveys closes the feedback loop. See AI-powered survey analysis for that workflow.
• Pattern detection across tickets surfaces emerging issues before they become outages — same theme-extraction techniques as the survey analysis post.

For broader context on AI in form workflows, see 7 ways AI is changing form builders in 2026 and AI vs. manual data entry. For the workflow-builder primitives, see the AI nodes overview and workflow automation best practices. Adjacent posts in the AI series:

• AI-powered lead qualification form — the same architecture for sales
• AI chatbot form for 24/7 lead qualification — conversational alternative
• Auto-generate forms from a PDF using AI — for converting paper-based intake to digital
• AI vision for document verification — when intake includes documents

The dropdown is gone. The user describes the problem. The right team gets it within 60 seconds. That is the new normal.