How to Build an AI First-Response Generator for Tickets

An ai first response generator for support tickets cuts your first response time from hours to seconds. I build these for teams where the initial reply is usually a template anyway. The AI reads the ticket, pulls from your knowledge base, and drafts a personalized response that either solves the issue or asks the right follow-up question.

Agents review and send. They do not start from blank.

What You Need Before Starting

• A ticketing system with API access
• Your knowledge base or FAQ content
• Python 3.8+ with the Anthropic SDK
• Historical resolved tickets for quality reference

Step 1: Load Your Knowledge Base

Structure your support content so the AI can reference it:

import json

def load_knowledge_base():
    with open("support_kb.json") as f:
        articles = json.load(f)
    return "\n\n".join([
        f"Topic: {a['title']}\nAnswer: {a['content']}"
        for a in articles
    ])

Step 2: Build the Response Generator

import anthropic

client = anthropic.Anthropic()

SYSTEM_PROMPT = """You are drafting a first response to a support ticket for [Your Company].

Rules:
- Address the customer by name if available
- Directly answer their question if the knowledge base has the answer
- If you need more info, ask ONE specific clarifying question
- Keep responses under 200 words
- Professional but warm tone
- Never make up information not in the knowledge base
- End with a clear next step

KNOWLEDGE BASE:
{kb}"""

def generate_first_response(ticket):
    kb = load_knowledge_base()

    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=400,
        system=SYSTEM_PROMPT.format(kb=kb),
        messages=[{
            "role": "user",
            "content": f"""Generate a first response for this ticket:

Customer: {ticket.get('customer_name', 'Customer')}
Subject: {ticket['subject']}
Message: {ticket['body']}"""
        }]
    )

    return response.content[0].text

Step 3: Auto-Draft on New Tickets

Hook into your ticketing system webhook:

from flask import Flask, request, jsonify

app = Flask(__name__)

@app.route("/new-ticket", methods=["POST"])
def new_ticket():
    ticket = request.json

    draft = generate_first_response(ticket)

    save_draft(ticket["id"], draft)

    return jsonify({
        "ticket_id": ticket["id"],
        "draft": draft,
        "status": "draft_ready"
    })

Step 4: Let Agents Review and Edit

Present drafts in a simple review interface. The agent can approve, edit, or reject:

@app.route("/review-draft/<ticket_id>", methods=["POST"])
def review_draft(ticket_id):
    action = request.json["action"]  # approve, edit, reject

    if action == "approve":
        draft = get_draft(ticket_id)
        send_reply(ticket_id, draft)
        log_review(ticket_id, "approved", draft)

    elif action == "edit":
        edited = request.json["edited_text"]
        send_reply(ticket_id, edited)
        log_review(ticket_id, "edited", edited)

    elif action == "reject":
        log_review(ticket_id, "rejected", get_draft(ticket_id))

    return jsonify({"status": "processed"})

Step 5: Track Approval Rates

Measure how often agents approve vs edit to gauge quality:

def get_quality_metrics():
    conn = sqlite3.connect("drafts.db")
    stats = conn.execute("""
        SELECT action, COUNT(*) as count
        FROM draft_reviews
        WHERE reviewed_at > datetime('now', '-7 days')
        GROUP BY action
    """).fetchall()

    total = sum(s[1] for s in stats)
    approved = next((s[1] for s in stats if s[0] == "approved"), 0)

    return {
        "total_drafts": total,
        "auto_approved": approved,
        "approval_rate": round(approved / total * 100, 1) if total else 0
    }

What to Build Next

For tickets with a 90%+ approval rate pattern (like password reset requests), switch to auto-send. The agent reviews after instead of before. That is where the real time savings stack up.

Related Reading

• AI in Customer Service - first response automation in the full support picture
• AI-Powered Reporting That Actually Gets Read - reporting on response quality and speed
• Minimum Viable Automation - start with drafts, graduate to auto-send