Set up automatic ticket tagging for Freshdesk

This guide walks you through setting up an agent that tags every new Freshdesk ticket as it arrives, and sets the ticket's Type (Freshdesk's own category field, like Question or Billing) when nobody set one. Use it if a large share of your helpdesk queue is noise, such as vendor notifications, newsletters, out-of-office replies, and cold sales pitches, and you want that separated from real customers without doing it by hand.

You import a ready-made workflow, fill in one box, connect three logins, and add one rule in Freshdesk. The workflow runs in n8n, a browser-based automation tool with nothing to install.

---

When to use this

Use this if:

• A meaningful share of what lands in your Freshdesk queue isn't a real customer ticket
• Tickets sit untagged and untyped until someone manually sorts them
• You want tagging to happen consistently, the moment a ticket arrives, not in a daily batch

This agent only tags and sets Type. It never replies to a customer, adds a public note, or changes any other field.

---

How it works

1. A ticket arrives in Freshdesk.
2. Freshdesk pings n8n to say a ticket came in.
3. n8n fetches the full ticket.
4. A CustomGPT.ai agent reads it and answers with tags, and a Type if the ticket has none.
5. n8n checks that answer against your own list, then writes the tags and Type onto the ticket.

The agent reads and decides. n8n makes every change to the ticket; the agent has no access to Freshdesk itself.

---

What you need

---

Step 1: Decide your tag vocabulary

Write out every tag the agent is allowed to use, along with what each one means. Both the agent and the workflow check answers against this exact list: the agent can only pick from it, and the workflow discards anything that isn't on it.

Two categories of tag matter here.

One deciding tag belongs on every ticket: action-required or no-action-required. This is the tag that separates real customers from noise. The rest of this guide calls it the disposition.

Topic tags work differently. A ticket can have several, or none.

Two things worth getting right now:

Reuse the tags your team already uses. If tickets have been tagged by hand until now, keep the agent consistent with that history, so a label doesn't start meaning something different once automation takes over.

Don't put two tags on the same idea. If both billing and billing_issue are on the list, the agent has no consistent way to choose between them, and one topic ends up split across two labels. Pick one spelling and stick to it.

You'll use this list twice: once in the agent's instructions in Step 3, and once in the workflow's settings in Step 5.

---

Step 2: Create the agent

1. Create a new agent in CustomGPT.ai.
2. Give it your tag list document as its only knowledge source.

3. Pick a current Claude or GPT model.
4. Turn Markdown off in the agent's response settings.

The agent itself doesn't need any tools or app connections. n8n just sends it the ticket text as a plain message, which is why each answer comes back in about ten seconds.

---

Step 3: Write the agent's instructions

Go to Personalize > Agent Persona and paste the instructions below, replacing everything in [SQUARE BRACKETS] with your own values. The tag list and Type list here must match what you put in the workflow's settings in Step 5, exactly.

You classify [YOUR COMPANY] support tickets from Freshdesk. You do two things: assign tags, and set Type when it is missing. You never do anything else.

You receive the ticket as text in the prompt: ticket id, subject, description, requester email, source, and current Type. Judge from that text.

Judge INTENT, not keywords. A newsletter saying "cost per seat" is not a billing ticket. A webinar invite saying "save your seat" is not a sales inquiry. A cold pitch titled "Partnership Proposal" is not a partner inquiry.

Step 1: Disposition (always exactly one)

action-required: a human must read this and do something.
no-action-required: safe to close unread. Nobody owes anyone a response. Covers vendor status mail, newsletters, autoresponders, bounces, cold outreach, SEO spam, and your own automation alerts.

Anything from a real customer about our product is action-required, including short, angry or unclear messages. WHEN UNCERTAIN, CHOOSE action-required. A real customer misfiled as noise is the worst error you can make.

OUR OWN AUTOMATION ALERTS ARE ALWAYS no-action-required. Tag them internal_alert and nothing else.

Step 2: Topic tags

Choose from this closed list ONLY. NEVER invent a tag.

Noise: third_party_notification, newsletter, auto_reply, spam, bounce, vendor_outreach, internal_alert
Content: [YOUR TOPIC TAGS, e.g. bug_report, account_issue, billing_issue, api_issue, integration_issue, feature_request]

BE PRECISE, NOT GENEROUS. Add a tag only if it is plainly true of this ticket. Zero topic tags is a valid answer.

Step 3: Type, only when missing

Type is MISSING when: it is empty/unset, OR it is "Other" and the source is not "Portal".
Type is PRESENT when: it is any specific value, OR it was set through the Portal.

If Type is PRESENT, do not touch it: type_is_missing = false and type = null. This holds even when the existing Type looks wrong, because a human chose it.

Allowed Type values: [YOUR FRESHDESK TYPES, SPELLED EXACTLY AS FRESHDESK SPELLS THEM]

TYPE MUST AGREE WITH DISPOSITION. If disposition is no-action-required, Type may ONLY be [YOUR NOISE TYPES, e.g. Other, Newsletter].

Output

Reply with exactly these six tags and nothing else. No preamble, no markdown fences, no text outside the tags.

<disposition>action-required or no-action-required</disposition>
<tags>comma-separated topic tags from the closed list, or empty</tags>
<type_is_missing>true or false</type_is_missing>
<type>the Type value, or null when type_is_missing is false</type>
<confidence>high or medium or low</confidence>
<reasoning>one or two sentences</reasoning>

Do not put the disposition tag inside <tags>. <tags> is for topic tags only.

You have no ability to modify Freshdesk and must never claim to have done so. You never set Priority, Group, Status or assignee, and you never write a reply to a customer.

Two things worth understanding about why the instructions are written this way:

The full tag list is written out, not summarized. Tell the agent to "pick something sensible" instead, and it starts inventing tags of its own, leaving your reporting full of near-duplicate labels for the same idea.

The reasoning line stays free text on purpose. Everything else is a short labelled answer so a stray sentence in one line doesn't break the rest. Without reasoning, a wrong label gives you nothing to go on when you're checking its work.

---

Step 4: Import the n8n workflow

Download the workflow file from [GITHUB LINK].

In n8n, open your list of workflows, choose Import from File, and select the JSON file you downloaded. The workflow lands fully built and labelled, so there's no dragging, connecting, or laying out left to do. Each step shows up as a box on the canvas, called a node.

What each section does:

The canvas has two starting points, a settings box, and four labelled sections, with instructions written directly on the page. The workflow can start two ways. Day to day, it's the Freshdesk webhook. To test it by hand instead (Step 7), use Test With One Ticket and type in a ticket number. Either path leads into the same settings box, then a Preflight Check step that halts the run with a plain-English message if a setting is missing.

Section 1: Read the ticket. Pulls the ticket and turns it into a message for the agent, skipping it first if the sender is on your own alerting list.

Section 2: Ask the agent. Opens a new conversation with your agent and hands it the ticket.

Section 3: Check the answer. Reads the agent's reply, drops any tag that isn't on your allowed list, and merges what's left with the tags the ticket already had.

Section 4: Write back to Freshdesk. Always writes the tags, and sets Type only if the ticket had none. If Freshdesk rejects the update, this section falls back to applying just the tags and leaves a private note explaining why - that's the lower path in the diagram below.

---

Step 5: Fill in the settings

Open the box called SETTINGS - EDIT THIS NODE. the only one in the whole workflow you'll actually edit.

Match Freshdesk's spelling for allowed_types exactly, punctuation included. If the agent suggests a Type that isn't on this list, the workflow silently drops it, so a typo here shows up as "Type never gets set," not as an obvious error.

If a setting's missing, the run halts at Preflight Check before it ever reaches Freshdesk or the agent, and spells out exactly what to fix. For example:

SETUP NEEDED - 3 things to fix before this can run: (1) Set freshdesk_domain in the SETTINGS node to your Freshdesk subdomain, the part before .freshdesk.com (2) Set customgpt_agent_id in the SETTINGS node to your CustomGPT agent number, the digits in the agent URL (3) No ticket number...

---

Step 6: Connect three logins

Three separate credentials, each doing a different job. n8n stores them so you don't have to paste a login into every step.

n8n already has built-in support for Freshdesk, so setting up that credential is just two fields: your API key and your domain.

n8n calls this credential type "Freshdesk API"; the name you give the credential itself, shown above, is "Freshdesk account."

Your domain goes in twice - here, and again in the settings box from Step 5. Mismatch the two and the login still works, but n8n queries the wrong Freshdesk for the ticket, which shows up as a "not found" error rather than anything that points at the domain.

CustomGPT.ai has no built-in integration in n8n, so this one you build by hand: call it Authorization and set the value to Bearer, a space, then your CustomGPT API key.

You'll reuse this same generic form for the webhook secret in Step 8, just with a different name and value.

Then open each step that needs a login and pick it from the dropdown. For a Freshdesk step, choose Predefined Credential Type > Freshdesk API:

For a CustomGPT step, choose Generic Credential Type > Header Auth > CustomGPT API Key:

If you see an "unauthorized" error, the step it happened on tells you which key is wrong. See Troubleshooting below.

---

Step 7: Test it

Enter a ticket number in Test Ticket ID and run the workflow. Pick a ticket you're fine with changing, since this genuinely edits it - you don't need the Freshdesk rule from Step 8 for this test.

The agent's side of each run shows up in the CustomGPT.ai dashboard as its own conversation. Here's its answer for a marketing email - tags, Type, confidence, and reasoning:

And here's what lands on the ticket itself - one deciding tag, two topic tags, and a Type:

Three signs confirm it worked:

1. The run completed with no error.
2. New tags landed on the ticket, plus a Type if it lacked one.
3. No private note appeared. Silence means success.

---

Step 8: Go live

In Freshdesk, go to Admin > Workflows > Automations > Ticket Creation and add a rule.

Freshdesk requires at least one condition on a rule, so give it one that runs regardless of priority - anything that matches every ticket created:

For the action, choose Trigger webhook, set the request type to POST, and paste your n8n webhook address into the URL box. n8n shows you that address on the Freshdesk Webhook step. Use the production one, not the test one.

Leave Requires authentication off; the secret goes through a different channel, covered next.

Turn on Add custom headers and put your webhook secret there. Set Encoding to JSON, Content to Simple, and check the Ticket ID placeholder so Freshdesk passes n8n the ticket number.

This header name has to match the one on your Freshdesk Webhook Secret credential in n8n. Use X-Webhook-Secret in both spots - it's how n8n tells your real Freshdesk apart from someone who simply guessed the webhook address.

Flip the workflow to active in n8n, and you're live.

---

Things to watch

Only ticket text reaches the agent, and only the start of it. Attachments are never sent, and the ticket description is capped at roughly 2,500 characters. On an unusually long ticket, the agent judges only the beginning of it.

A failed call to your agent stops the run, with no note. Unlike a refused Freshdesk update, a failed or timed-out call to your CustomGPT.ai agent, for example a bad API key, or a run that outpaces your plan's usage, has no fallback today. Nothing gets written to the ticket, and nothing explains why. Check the run in n8n if a ticket seems to have been skipped.

Tags get combined, not replaced. A normal ticket update overwrites the whole tag list with whatever you send it, so sending only the agent's tags would silently wipe out anything a colleague added by hand. This workflow avoids that by reading the ticket's existing tags first and writing both sets together.

The workflow enforces the tag list, not just the agent. The agent's instructions aren't the only safeguard - anything it returns that isn't on your list gets dropped before the write happens. Those drops get reported, so you can decide whether to add the tag on purpose or tighten the instructions instead.

An empty required field can block an update that never touched it. Freshdesk re-validates every required field on each change, so it can refuse the whole update over a blank field you had nothing to do with - this shows up mostly on tickets that are already closed. When that happens, the tags still go through on their own, and a private note names the field that caused the rejection. Give that field a default under Admin > Ticket Fields to prevent it, or turn off Explain What Failed if you'd rather skip the notes.

A Type someone chose is never overwritten. The workflow only counts Type as missing when it's blank, or when it's Other on a ticket that didn't arrive through the customer portal. Email tickets have no Type picker at all, so Other there is just the default value, not a choice anyone made.

Confirm your webhook address actually works. Some tools display one address while listening on a different one. Run a test before wiring it into Freshdesk - otherwise the rule quietly sends tickets into a dead end.

---

Troubleshooting

---

Related articles

• Create a new agent
• Use Persona Builder to set up your agent's persona
• How Persona Builder works