How to Draft Zendesk Macros With ChatGPT Without Exposing Data

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Draft Zendesk Macros With ChatGPT Without Exposing Data.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Draft Zendesk Macros With ChatGPT Without Exposing Data.

Work from redacted examples, never raw tickets. To draft Zendesk macros with ChatGPT without exposing customer data, strip names, emails, phone numbers, addresses and account numbers from 10 to 20 past replies, paste them in with model training off (or on a Business workspace) and ask for macros built on Zendesk placeholders. For one-off replies, describe the problem in your own words. A person approves everything.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

This works because a good macro never needs a real customer in it. Zendesk swaps in details such as {{ticket.requester.first_name}} and {{ticket.id}} when an agent uses the macro, so the text ChatGPT drafts can be completely anonymous. Live tickets are a different matter. There are now two ways to put AI on a ticket in place, OpenAI's Zendesk (Beta) plugin for ChatGPT and Zendesk's own writing tools, and they send customer data to different places under different rules. Pick the route before anyone starts pasting.

Which route to AI in Zendesk exposes the least customer data?

Pasting redacted text into ChatGPT exposes the least, because only what you paste leaves Zendesk, while the plugin and Zendesk's own writing tools handle live ticket data under different rules.

RouteWhat leaves your ZendeskList costBest use
Redacted text pasted into ChatGPTOnly what you paste, after redactionChatGPT Plus $20 a month; Business $25 a user a month (or $20 billed annually), from two usersBuilding and refreshing the macro library, tone guide and help-centre snippets
OpenAI's Zendesk (Beta) pluginTicket content and customer context the plugin readsNo separate charge listed by OpenAI; availability depends on plan and workspace settingsDrafting one reply to a complicated ticket with its full history
Zendesk's own writing toolsProcessed by Zendesk's AI; Zendesk says OpenAI will not use the data to train its modelsSuite or Support Professional and above (Suite Professional is $115 an agent a month billed yearly); Copilot add-on $50 an agent a month billed yearlyPolishing replies and macros inside the composer, every day

The plugin deserves a closer look because it is the newest and the easiest to misuse. OpenAI's plugin page lists it as "Zendesk (Beta)", developed by OpenAI, with interactive, read and write capabilities: it can investigate tickets and customer context, review queues and update Zendesk "only when requested". One of OpenAI's own example prompts asks it to summarise a ticket and "draft a concise customer reply for review without sending it". On ChatGPT Business, Enterprise and Edu, OpenAI says information accessed through apps is not used for training by default; on Free, Go, Plus and Pro it may be, if Improve the model for everyone is switched on. So the plugin is reasonable on a Business workspace, and a poor idea on a personal account with default settings.

For most small desks, the order I'd suggest is: build the macro library with redacted text first, because that's where most of the time savings sit, then decide whether agents need AI on live tickets at all. If they do, Zendesk's own tools keep the data in one system, which is simpler to explain to customers and to your own team. If you are still weighing the Zendesk side, whether Zendesk Copilot is worth $50 an agent works through that decision in detail.

What should you redact before ticket text reaches ChatGPT?

Before any ticket text reaches ChatGPT, replace names, ticket numbers, site addresses and serial numbers with bracketed labels, and delete email addresses, phone numbers, logins and payment details entirely. The rule of thumb: ChatGPT needs the shape of the problem and the shape of a good answer, never the person. For an illustrative solar installer's service desk, where tickets are full of site details, equipment serial numbers and monitoring logins, the redaction sheet looks like this:

In the ticketReplace withLater becomes in the macro
Customer's name, in greeting or signature[CUSTOMER]{{ticket.requester.first_name}}
Ticket number[TICKET]{{ticket.id}}
Agent's name and sign-off[AGENT]{{current_user.first_name}}
Email address, phone numberDeleteNothing: Zendesk already holds them
Site address, gate or key-safe codes, directions[SITE]Nothing: never in a macro
Inverter, battery or panel serial numbers[SERIAL]A request: "the serial number on the label"
Monitoring portal username or passwordDelete entirelyNothing, ever
Finance agreement, card or bank detailsDelete entirelyA link to your billing team or payment page
Photos, screenshots, attachmentsDon't uploadDescribe in words if needed
Internal notesDon't pasteNothing: they carry the most candid detail

Here is how one ticket reply looks before and after redaction. Before (illustrative):

Hi [customer's first name], thanks for sending the photo. Your inverter (serial [full serial number]) at [full site address] went offline on the 3rd, which matches when you said the new broadband router went in. I've logged into your portal (user: [customer's email]) and can see no data since. If you text me on [office mobile number] I'll walk you through reconnecting. Thanks, [agent's first name], Service Desk

After:

Hi [CUSTOMER], thanks for sending the photo. Your inverter ([SERIAL]) at [SITE] went offline on [DATE], which matches when you said a new broadband router was installed. I can see no monitoring data since. If you reply here, I'll walk you through reconnecting. Thanks, [AGENT]

Notice the redaction also removed things that weren't customer data but would have caused trouble: the claim to have logged into the customer's portal, and the office mobile number, which would end up baked into a macro used by everyone. Three quick checks before pasting anything: search the text for "@" (stray emails), for any run of digits longer than four (serials, phone numbers, order numbers) and for capitalised words that aren't product names (people and places). Strip email signatures and the quoted thread below the latest reply, because that is where phone numbers hide.

On a personal ChatGPT plan, two settings matter. Switch off Improve the model for everyone under Settings > Data controls, and do redaction sessions in a Temporary Chat: OpenAI says temporary chats don't appear in history, don't create memories and aren't used to improve its models, though they may be kept for up to 30 days for safety. That also stops ChatGPT's memory from quietly learning details about your customers. The general method behind this sheet, for any document, is in anonymising client data before you paste it into AI.

How do you keep every ChatGPT macro in the same voice?

A tone guide keeps every ChatGPT macro in one voice: write it once and paste it at the top of every drafting session, and macros from different sessions will sound like the same company. Without a guide, ChatGPT writes in its own default voice: warm, wordy and heavy on apology. Here is a filled-in version for the solar installer's desk:

TONE GUIDE: solar installer service desk (paste before every request)

Voice: plain, friendly, practical. We sound like the engineer who
fitted the system, not a call centre.
Length: under 120 words unless the customer must follow steps.
Greeting: "Hi {{ticket.requester.first_name}},"  Sign-off: "Thanks,
{{current_user.first_name}}"
Structure: what we know -> what we're doing -> what we need from you
-> when you'll hear from us.
Apologise once at most, and only for something we got wrong.
Never promise: generation figures, savings, payback, warranty
outcomes, visit dates we haven't booked.
Never diagnose a fault as safe. Any mention of burning smells,
smoke, scorch marks, water near equipment or tripping breakers
goes to the SAFETY macro written by the operations lead. Do not
draft alternatives to it.
Words we use: "system", "monitoring app", "service visit".
Words we avoid: "unfortunately", "per our policy", "kindly".

Two rules carry most of the weight. "Never promise generation figures" exists because output questions are one of this desk's busiest topics (40 "generation lower than expected" tickets a month in the illustration below), and an AI-written "your system should produce around X" is a promise the business can't keep through a cloudy month. And the safety rule keeps AI completely out of the one category where wrong wording could hurt someone. If you'd like a deeper template for keeping AI replies on-script, reply templates that keep AI replies on-script covers the structure.

Twelve macros for a solar installer's service desk, start to finish

The main worked example. An illustrative three-agent service desk at a solar installer handles about 300 tickets a month on Zendesk Suite Team. A tally of one month's tickets by tag showed five topics making up most of the volume: monitoring offline (55 tickets), "generation lower than expected" (40), battery behaviour questions (35), invoices and finance queries (30) and annual service bookings (25). The owner wants 12 macros: two or three per topic, covering the usual variants.

Step 1: pick and redact the source replies (about 45 minutes)

For each topic, pick three solved tickets where the agent's reply was good, ideally ones the customer rated well. Copy only the agent's final reply plus one line summarising the customer's question in your own words. Redact with the sheet above. Fifteen replies at two or three minutes each is roughly 45 minutes, and it is the step people skip. Don't.

Step 2: the drafting prompt

[Paste TONE GUIDE]

You are drafting Zendesk macros for a solar installer's service desk.
Below are redacted examples of good replies for the topic
"monitoring offline". [CUSTOMER], [AGENT], [TICKET], [SERIAL] and
[SITE] are placeholders for removed details.

Write 3 macros for this topic:
1. Wi-Fi or router changed
2. Offline with no known cause
3. Offline after a power cut

For each macro give:
- Title in the form Monitoring::Offline::[variant]
- "Use when" note for agents (one line, internal)
- Customer-facing comment using {{ticket.requester.first_name}},
  {{ticket.id}} and {{current_user.first_name}} where needed
- One tag in lower case with underscores

Rules: only use steps that appear in the examples. If a step needs
information the examples don't give (an app menu, a setting name),
write [CHECK: describe what's needed] instead of guessing.

EXAMPLES:
[paste the redacted replies]

Step 3: what came back, and what was fixed

An illustrative first macro from the reply:

Title: Monitoring::Offline::Router changed
Use when: customer has a new router or Wi-Fi password since the
system went offline.
Comment:
Hi {{ticket.requester.first_name}},
Thanks for letting us know. When a router or Wi-Fi password changes,
the monitoring unit loses its connection, and the system keeps
generating as normal while the app shows no data.
To reconnect, open the monitoring app and go to Settings > Network >
Reconnect, then enter your new Wi-Fi details. If that doesn't work
within 24 hours, reply to this message quoting ticket {{ticket.id}}
and we'll book a remote session.
Thanks,
{{current_user.first_name}}
Tag: monitoring_offline_router

Mostly usable, with three fixes. First, "Settings > Network > Reconnect" appears in none of the examples: ChatGPT invented a menu path despite the rule to use only steps from the examples, and the installer fits two brands of inverter with different apps. It became two macros, one per brand, with the real steps written by the engineer. Second, "within 24 hours" committed the desk to a response time nobody had agreed. It became "if that doesn't work, reply here and we'll arrange a remote session". Third, "the system keeps generating as normal" is a reassuring claim the agent can't verify for every site, so it became "your panels usually keep working while monitoring is offline; we'll confirm once it's reconnected". The other eleven macros needed similar edits, mostly removing small promises.

Step 4: build and test in Zendesk (about 60 minutes)

  1. In Admin Center, click Workspaces, then Agent tools > Macros, and click Create macro.
  2. Paste the title. The double colons put it in nested categories, which Zendesk's macro list uses for filtering, so agents can find "Monitoring" and drill down.
  3. Set Available for to Me only while testing.
  4. Click Add action, choose Comment/description and paste the comment text. Add a second action to add the tag.
  5. Create a test ticket with yourself as the requester, apply the macro, and check every placeholder renders as your own name and ticket number.
  6. Once it reads correctly, change Available for to All agents. Sharing with a single group (Agents in group) needs Suite Growth or Support Professional and above, so on Suite Team it is all agents or just you.

For the simplest topics you may not need ChatGPT at all. When an agent has already written a good reply, Zendesk can turn that ticket into a macro directly: open the ticket, choose Create as macro from the ticket options menu, and rewrite the comment so it no longer refers to that one customer. Nothing leaves Zendesk, and the rewrite is the same redaction discipline applied in place.

Total for 12 macros in this illustration: 45 minutes redacting, 30 minutes prompting across five topics, about 60 minutes of fixes and engineer input, and 60 minutes building and testing. Call it three and a quarter hours, against a full day the last time the desk wrote macros by hand, and the redacted versions now sit in a folder ready for next quarter's refresh.

How do you reply to a live ticket without pasting it into ChatGPT?

Describe the problem in your own words instead of pasting the ticket into ChatGPT. Macros cover the routine, but the awkward ticket, the angry customer or the odd fault is where agents are most tempted to paste the whole thread. Three safer patterns, in order of preference.

Describe the problem instead of pasting it

[Paste TONE GUIDE]
A customer's battery shows 0% every morning since a firmware update
two weeks ago. They have restarted it twice. They are frustrated
because this is their third contact. We have booked an engineer
visit for next week and want to explain what the engineer will check.
Draft a reply under 120 words. No promises about the outcome.

Illustrative output: "Hi [first name], I'm sorry this has taken three contacts, that's not good enough. Thank you for restarting the battery twice; that rules out the simple fixes. Our engineer will visit next week to check the battery settings since the update, the connection to your inverter and how it's charging overnight, and will fix whatever they find…" The fix: the date. "Next week" became the booked date, typed by the agent, and "will fix whatever they find" became "will explain what they find and the next step", because some faults need a part ordered. Not one detail identifying the customer ever left Zendesk.

Use the Zendesk (Beta) plugin, but only on a Business workspace

When the history genuinely matters (a long thread with three engineers' notes), the plugin can read it in place. Set it up deliberately: install it from the Plugins directory in ChatGPT, review the permissions it asks for, and under the app's permission settings choose Always ask rather than Allow read actions or anything broader, so ChatGPT asks before every read or action. OpenAI's help page warns that Allow all actions carries elevated risk because actions may run without another confirmation. Then use a prompt in the spirit of OpenAI's own example: summarise the ticket, list what's been tried and what's missing, and draft a reply for review without sending it. The agent copies the approved draft into Zendesk. Availability depends on your plan and workspace settings, and your workspace admin can turn the app off from the admin console.

Use Zendesk's composer tools for tone

If you're on Suite or Support Professional or above, the enhance writing tools sit in the ticket composer and the macro editor: Expand, Simplify, Make more friendly and Make more formal, with Write a custom prompt on the Copilot add-on. For an agent who has written a correct but curt reply, one click on "Make more friendly" does the job without anything leaving Zendesk. Admins control which groups get it under AI > Agent copilot > Writing in Admin Center.

What should you check before a ChatGPT-drafted macro goes live?

Before a macro goes live, check that placeholders render on a test ticket, that someone who does the work has checked the steps, and that nothing promises more than policy allows. Every macro gets the same check, filled in here for the "Monitoring::Offline::Router changed" macro:

  • Placeholders render on a test ticket: yes, own name and ticket number appeared.
  • Steps checked by someone who does the work: engineer rewrote the reconnection steps per inverter brand.
  • No promise outside policy: "within 24 hours" removed; "keeps generating" softened.
  • No customer data in the text: searched for "@" and digit runs; none.
  • Safety wording untouched: not a safety topic; SAFETY macro not edited.
  • Tag added for reporting: monitoring_offline_router.
  • Audience set: moved from Me only to All agents.
  • Problem tickets: not used on Problem tickets; if it were, placeholders escaped (see below).

How can a macro leak data or mislead despite a careful process?

A careful process can still leak data: on a Problem ticket, a reply using {{ticket.requester.name}} sends the requester's name to every linked ticket, according to Zendesk's macro documentation.

These are the failures that turn up in practice, each with the symptom that gives it away.

  • A macro broadcasts one customer's name to many. Zendesk's macro documentation warns that on a Problem ticket, placeholders render when the macro is applied, so a reply using {{ticket.requester.name}} sends the problem ticket requester's name to every linked ticket. The symptom is a confused customer replying "who is this?". The fix Zendesk gives is an escape character before the placeholder: \{{ticket.requester.name}}.
  • Customer wording survives in the macro. A redacted example said "as you mentioned, the scaffolding firm left the cable loose", and ChatGPT kept that phrasing in a general macro. Every customer then received a sentence about scaffolding. Search new macros for anything that only made sense in one ticket.
  • Phone numbers sneak through in signatures. Agents redact the reply but paste the quoted thread underneath, signature and all. Only ever copy the reply itself.
  • Memory remembers customers. On personal plans with memory on, details pasted in normal chats can be remembered. Temporary Chat avoids this; on a Business workspace, inputs also stay out of model training by default.
  • The plugin gets broad permissions "to save time". Someone switches the app to allow actions without asking, and a draft lands on a ticket unchecked. Keep it on Always ask and keep drafts out of the ticket until a person pastes them in.
  • Made-up troubleshooting. Invented app menus, settings and error-code meanings read perfectly. The [CHECK: …] rule in the prompt surfaces most of them; an engineer's read catches the rest.

For the wider question of what your team should never paste anywhere, whether it's safe to put customer data into ChatGPT sets out the categories and the plan differences.

What does ChatGPT cost compared with Zendesk's built-in AI for three agents?

For the illustrative three-agent desk, ChatGPT Plus for the person who writes macros costs $20 a month and ChatGPT Business for two seats $50, against $180 a month to move three agents from Suite Team to Suite Professional and a further $150 for Zendesk's Copilot add-on, on yearly billing.

A quick illustrative sum for the same desk, using list prices checked in November 2026:

OptionExtra monthly costWhat it gives the desk
ChatGPT Plus for the person who writes macros$20Macro drafting with redacted text; Temporary Chat for one-off replies
ChatGPT Business, two seats, monthly billing$50No training on business data by default; the Zendesk (Beta) plugin where available
Move Suite Team ($55) to Suite Professional ($115), yearly billing$180 (3 x $60)Enhance writing tools in Zendesk, with Copilot features pooled at 5 uses per agent a month (15 in total)
Copilot add-on on top of Professional$150 (3 x $50)Custom prompts, suggested first replies, macro content suggestions from your own ticket data

The pattern is clear. Building and refreshing a macro library is a few hours a quarter, so $20 a month of ChatGPT with disciplined redaction is the cheapest way to do it. AI help on every live reply is a daily, per-agent job, which is what Zendesk's own tools are priced for, and they keep the data in one place. Note how small the Professional allowance is without the add-on: 15 pooled uses a month across three agents is a taste, not a workflow. With the add-on, Zendesk's macro content suggestions analyse how your agents actually resolve tickets and propose new macros or edits, with links to the tickets behind each suggestion, which overlaps with the manual method above. A desk already paying for Copilot may not need ChatGPT for macros at all.

If you want help choosing between ChatGPT and Zendesk's own AI for your support desk, I offer one-to-one AI consultations for small businesses.

How do you check that your macros earn their place?

Check after a month which macros agents actually use: tag each macro, report on usage in Zendesk Explore, and retire or rewrite the ones that are ignored.

  • Usage by tag. Because each macro adds its own tag, a simple report in Zendesk's Explore (Zendesk publishes a recipe for reporting on macros using tags) shows which macros are used and which are ignored. Retire or rewrite any macro used fewer than three or four times in a month.
  • Reopen rate on macro-solved tickets. If tickets solved with "Monitoring::Offline::Router changed" reopen more often than average, the steps don't work for customers. That usually means a missing step, not bad tone.
  • Agent edits. Ask agents which macros they always edit before sending. Heavy editing means the macro is too specific or too vague; send it back through the drafting prompt with the edited versions as new examples, redacted first.
  • Customer comments. Read satisfaction comments on macro-answered tickets. "Felt like a template" on a sensitive topic tells you that topic needs a personal reply, not a better macro.

A month in, the desk in this example should be answering most monitoring and booking tickets with approved macros, keeping ChatGPT for the occasional described-not-pasted reply, and holding the redacted examples ready for the next refresh.

If you export macro usage counts to a spreadsheet, The Complete Excel Course with AI Integration on Coding Liquids covers the spreadsheet side.

Questions about ChatGPT, Zendesk and customer data

Does OpenAI train on Zendesk tickets that ChatGPT reads?

It depends on the plan. OpenAI says it does not use information accessed from apps to train models by default on ChatGPT Business, Enterprise and Edu. On Free, Go, Plus and Pro it may, if the Improve the model for everyone setting is on. Switch it off under Settings, Data controls, or use a Business workspace.

Can ChatGPT post replies straight into Zendesk?

It can, but a person should approve every reply first. OpenAI lists its Zendesk (Beta) plugin with read and write capabilities and says it updates Zendesk only when requested. For customer replies, keep a person in the loop: set the app permission to Always ask, ask for a draft without sending, and paste the approved text into the ticket yourself.

Which Zendesk plans include AI writing tools?

Suite and Support Professional plans and above. Zendesk's help pages say its generative writing tools (Expand, Simplify, Make more friendly, Make more formal) come with those plans, with a monthly usage allowance. The Copilot add-on, $50 an agent a month billed yearly, adds custom prompts, suggested first replies and macro content suggestions.

Should agents keep using ChatGPT once macros are built?

For routine tickets, no: the point of the macro library is that agents apply approved wording inside Zendesk. Keep ChatGPT for the occasional awkward reply, described in your own words, and for the quarterly macro refresh. That keeps the volume of anything leaving Zendesk small and easy to police.

How do I stop customer data reaching ChatGPT when drafting macros?

Redact the examples first. Strip names, emails, phone numbers, addresses and account numbers from 10 to 20 past replies, and ask ChatGPT to build macros on Zendesk placeholders, which fill in the customer's name and ticket number only when an agent applies the macro. Turn model training off or use a Business workspace, and have a person approve every macro.

Further reads

Sources: OpenAI's Zendesk (Beta) plugin page, OpenAI help pages (Connected apps in ChatGPT; Plugins in ChatGPT; Data controls in ChatGPT), chatgpt.com/pricing, openai.com/business/chatgpt-pricing, zendesk.com/pricing, Zendesk help articles on creating macros, macros from tickets, placeholders, organising macros, enhance writing, the Copilot add-on and macro content suggestions. Checked 17 November 2026.

Want a second pair of eyes on your setup?

Bring the process you're trying to fix. On a 1:1 call we'll work out which AI tools should handle it, which parts stay manual, and how to roll it out without disrupting the team.

Book a 1:1 call with me