The Reply Loop

Description

GOAL
Build a workflow that automatically searches the FAQ knowledge base for every new support ticket, creates a personalized reply draft from it, and places it in the ticket for the responsible agent to review. No automatic dispatch to the customer.

Adapt the structure to the concepts of your automation tool, for example nodes, steps, zaps or agent actions. While doing so, preserve the trigger logic, the sequence and the data flow.

TRIGGER
Type: webhook or direct integration with the ticketing system (Zendesk, Freshdesk or Intercom), fires on every new ticket.

STEPS
1. Load ticket data: subject, message, requester name and metadata such as queue and assigned agent.
2. Search the FAQ: query the knowledge base using the core terms of the request, preferring semantic search over exact keyword matches. Collect the most matching entries together with content and links.
3. AI step: create the reply draft, using the prompt below exactly.
4. Save the draft: as an internal comment or draft in the ticket, no automatic dispatch to the customer.
5. Output: notify the responsible agent in Slack or Teams, with a ticket link and a note about the draft that is ready.

PROMPT (use exactly as written)
FAQ reply generator:
"""
You are a professional support agent for [COMPANY]. Analyze the following customer inquiry and create a helpful, friendly reply draft.

Customer inquiry:
Subject: [TICKET_SUBJECT]
Message: [TICKET_MESSAGE]
Requester name: [CUSTOMER_NAME]

Relevant FAQ entries:
[FAQ_RESULTS]

Task:
Create a reply draft that
1. addresses the inquiry directly and calls the requester by name,
2. uses the information from the FAQ entries but tailors it to the specific situation,
3. is written in the tone [TONE], for example professional and friendly, casual or formal,
4. names concrete next steps or solutions,
5. includes a matching closing and signature.

Format:
Output only the finished reply draft, without additional explanations or meta commentary.
"""

DATA FLOW
Step 1 to step 3: subject, message and requester name.
Step 2 to step 3: the most matching FAQ entries.
Step 3 to step 4: the finished reply draft.
Trigger and step 1 to step 5: assigned agent and ticket link.

ERROR HANDLING
No FAQ matches: create the draft anyway, but flag it clearly as without an FAQ basis so the agent can decide. This also signals a gap in the knowledge base.
Requester name not extractable: use a neutral fallback, for example "Hello there", instead of an incorrect salutation.
Draft not saved: check the write permissions for comments and drafts in the ticketing system.
Notification not received: check the channel ID or webhook.
Request in a different language: create the draft in the language of the request.

PLACEHOLDER CONVENTION
All placeholders in square brackets are variables. Mapped dynamically are [TICKET_SUBJECT], [TICKET_MESSAGE], [CUSTOMER_NAME] and [FAQ_RESULTS], as described in the data flow. Not dynamic are [COMPANY] and [TONE]: set these manually before activation. List all placeholders that require manual configuration separately at the end.

SELF CHECK BEFORE COMPLETION
Check, before you report the workflow as finished:
1. Are all steps connected correctly, and is the draft stored exclusively internally, without being sent to the customer?
2. Is the prompt inserted in full and unchanged, and are all dynamic placeholders mapped?
3. Does the FAQ search use semantic matching, and are the matches passed to the prompt correctly?
4. Are all points listed that require manual configuration: the ticketing system webhook and write permissions, access to the knowledge base, the notification channel, [COMPANY], [TONE] and the LLM access?

Working prompts

You are a professional support agent for [COMPANY]. Analyze the following customer inquiry and create a helpful, friendly reply draft.

Customer inquiry:
Subject: [TICKET_SUBJECT]
Message: [TICKET_MESSAGE]
Requester name: [CUSTOMER_NAME]

Relevant FAQ entries:
[FAQ_RESULTS]

Task:
Create a reply draft that
1. addresses the inquiry directly and calls the requester by name,
2. uses the information from the FAQ entries but tailors it to the specific situation,
3. is written in the tone [TONE], for example professional and friendly, casual or formal,
4. names concrete next steps or solutions,
5. includes a matching closing and signature.

Format:
Output only the finished reply draft, without additional explanations or meta commentary.

JSON export