Agentic activities · Automation

Wire a help desk auto-triage flow that drafts responses for you

Combine the help desk triage skill from Workshop 2 with an automation that watches a shared inbox or ticket queue and writes the first-draft response. The technician approves or edits before sending. The agent does the typing; the human does the judgment.

About 35 minutes. Everything you write stays in your browser.

A help desk technician spends a third of their day typing the same five responses (“Have you tried restarting?” “Can you confirm which device this is on?” “I’ve reset your password — please check your phone for the reset code”). An automation that drafts those responses and waits for your approval saves two hours a day. The pattern is the same as the job-search agent: agent drafts, human approves, sends.

This activity uses Make.com as the example because its free tier handles multi-step flows better than Zapier’s. The pattern transfers.

Pick a non-production sandbox

You are not going to wire this to your real corporate ticketing system today. Your employer’s ticketing system has compliance, security, and approval requirements you cannot satisfy in 35 minutes.

Instead, build the flow against a sandbox you control:

A personal Gmail account. Set up a label called “ticket-sandbox.” Forward fake tickets to it.
A Google Sheet with a “ticket” column. Same as the job-search agent.
A Trello or Notion board with a “New” column. Make has integrations for both.

Pick the lightest option you can stand. Sheet works fine.

Your sandbox

Not saved yet.

Build the Make scenario

In Make, click “Create scenario.” The shape:

Module 1 - Trigger. Watch the sandbox for new items.

Google Sheets: “Watch new rows.”
Gmail: “Watch new emails matching label.”

Module 2 - AI: classify the ticket.

OpenAI or Anthropic module (Make supports both).
Prompt: the classification prompt below.
Output: a category (password / hardware / software / network / other) and a priority (P1 / P2 / P3).

Module 3 - AI: draft the response.

Same AI module as 2.
Prompt: the response draft prompt below.
Input: the ticket + the classification.

Module 4 - Output: write back.

Update the sheet row, or label the email “Draft ready.”
Save the draft into the sheet’s “Draft” column.

Module 5 - Notify.

Email yourself or post to a Slack channel: “New ticket drafted, ready for review.”

Save the scenario. Run it once manually with a test ticket before turning it on.

The classification prompt (module 2)

Classification prompt

You are classifying a help desk ticket for an automated triage flow. I will give you the ticket. Do exactly two things and nothing else.

1. CATEGORY (pick one): password, hardware, software, network, access, other.
2. PRIORITY (pick one):
 - P1: user fully blocked AND time-sensitive (deadline named, revenue impact, multiple users affected).
 - P2: user significantly slowed but a workaround exists.
 - P3: inconvenience or cosmetic.

Rules:
- Pick exactly one category and one priority. Do not output "P2 or P3."
- If the ticket is a P1, justify in one sentence using the user's exact words.
- If the ticket is missing key information, default to P2 and note "needs clarifying questions" in the justification.

Format:
CATEGORY: ___
PRIORITY: ___
JUSTIFICATION: [one sentence]

Ticket:
{{ticket}}

The response draft prompt (module 3)

Response draft prompt

You are drafting the first response a help desk technician will send to a user. I will give you the ticket and the classification.

Rules:
- Under 100 words.
- Open with one sentence acknowledging what the user said, in their words.
- If the classification is "password" and the priority is P2 or higher, ask for the user's identity verification method available in your environment (callback, MFA push, manager confirmation). Do NOT promise to email a new password.
- For all other categories, ask the 2-3 most important clarifying questions, numbered, in plain English.
- One sentence on what happens next.
- Sign off with placeholder "[Your name]".
- No filler ("Thank you for reaching out," "We apologize for the inconvenience").
- No emoji.

Ticket:
{{ticket}}

Classification:
Category: {{category}}
Priority: {{priority}}
Justification: {{justification}}

Run three test tickets and grade the output

Add three test tickets to the sandbox. Use these or your own:

“Email is broken since this morning. I have a 2pm meeting and I need to send the agenda.”
“Printer in the back office says offline. Front office printers work.”
“Can you reset my password and send to my Gmail? I’m out of the office, urgent contract.”

Wait for the scenario to run. Open your sandbox. Read the drafts.

Notes on the three test outputs

Not saved yet.

The third ticket is the one that matters. The draft must NOT promise to send the password to the personal Gmail. If your draft does that, your prompt is too loose. Tighten the password rule and re-run.

Build the human review checklist

You will not check every drafted response forever. After two weeks of approving drafts, you will be tempted to skip review on the routine ones. Resist.

Write down what triggers a careful review.

Your review rules

Not saved yet.

Self-check: is this safe to deploy at work?

Check each one you can honestly say yes to. Saved to your browser.

I built and tested this against a sandbox. I have not wired it to my real corporate ticketing system without my employer's approval.
The agent never sends a response without my approval click.
On the password-reset test ticket, the draft does NOT promise to send credentials to a personal email.
The classification step gets P1 right (named deadline) and does not over-classify routine tickets.
I have a written rule for which drafts get a careful review versus a quick review.
If I were to propose this at work, I know who I would need approval from (security team, IT manager, compliance).
I sanitize ticket content (real user names, emails, asset tags) before any test data goes into a public AI.

What to watch for

Compliance is real. Most workplaces have policies about what data can flow through which third-party systems. A free Make scenario routes ticket text through Make’s servers and through the AI provider’s servers. That is a vendor relationship your employer probably has not approved. Build the proof of concept on a sandbox; pitch the production version through the right channels.
AI confidently writes wrong technical advice. “Try restarting your computer” is the AI’s default suggestion for almost everything. Watch for this in your draft outputs. Tighten the prompt to require the agent to ASK before suggesting restart.
The classification prompt is the cheap leverage point. A bad classification cascades into a bad draft. Spend more time on prompt 1 than on prompt 2.
Free tier limits. Make’s 1,000 ops/month sounds generous but a 4-step scenario at 25 tickets/day eats it in a week. If this is useful, you may need the paid tier or a different platform.
Never wire the agent to auto-send. Even with strict prompts, the agent will eventually generate a response that is wrong in a way you would notice in 5 seconds and the AI will not. The whole point of human-in-the-loop is that 5-second check.
Audit log. If you ever do propose this for production, your employer’s security team will ask about audit logs. Make and Zapier both keep run histories on free tiers; know how to pull them.