Short answer
An AI FAQ generator works best when it groups your real support tickets into themes with counts, then drafts each answer only from your written policies and flags questions the policies do not cover. New Zealand businesses should strip personal information first, check answers against the Consumer Guarantees Act, and have the policy owner approve the FAQ before it is published.
- Remove names, emails and order numbers before tickets go into any AI tool.
- Let ticket counts set the FAQ order; let policy text set the answers.
- Flag questions your policy does not answer instead of letting the AI improvise.
- Never publish an answer that implies customers lose Consumer Guarantees Act rights.
- Recount ticket themes a month later to see whether the FAQ worked.
- Goal
- Turn a batch of support tickets into a short, accurate customer FAQ, ordered by how often each question comes up.
- Who it is for
- Customer support, ecommerce and operations staff in New Zealand businesses that answer the same questions every week.
- Time
- 30 minutes.
- You need
- ChatGPT, Claude, Copilot or Gemini, the fictional tickets and policy below, and later an export of your own tickets with personal details removed.
- You will produce
- A themed ticket count, draft FAQ answers with policy references, and a list of policy gaps.
Why build an FAQ from tickets instead of guessing?
Because tickets show what customers actually ask, in their words, and how often. FAQs written from a meeting room answer the questions the business wishes people asked, in language customers never search for. AI is well suited to the slow part: reading a month of tickets, grouping near-identical questions, and counting them. People stay responsible for the answers.
What does a good FAQ answer need?
A good FAQ answer is short, specific, and traceable to a written policy. Use this test on every draft:
| Quality | Weak answer | Strong answer |
|---|---|---|
| Source | Written from what support staff usually say | Quotes or paraphrases a named policy section |
| Specific | “Delivery is usually quick” | “North Island orders arrive in 1 to 3 working days” |
| Honest about limits | Promises outcomes the policy does not cover | Says what happens next and who to contact |
| Legally careful | “Sale items are non-refundable” | “No change-of-mind returns on sale items. Faulty items are covered as normal” |
| Findable | Uses internal jargon | Uses the customer’s own words from tickets |
Exercise: from tickets to a draft FAQ
The store, tickets and policy are fictional, and personal details are already removed. One support reply contains a mistake worth catching.
Illustrative tickets. Fictional online store, Dunedin.
T1 When will my order arrive? Ordered Monday, Auckland.
T2 Do you ship to Australia?
T3 Can I swap a size 1 for size 2? Tags still on.
T4 The wool singlet pilled after two washes. Is that normal?
T5 Where is my parcel? Tracking hasn't moved in 4 days.
T6 Bought in the sale, wrong size, can I return it?
T7 How do I wash merino without shrinking it?
T8 Can I exchange a gift without the receipt?
T9 Seam split on a sleepsuit after one week.
Agent reply: "Sorry, sale items are non-refundable."
T10 Delivery time to Invercargill?
T11 Do you do gift wrapping?
T12 Size 6-12 months or 1 year for a 9-month-old?
T13 Can I return a used item if my baby didn't like it?
T14 Is the merino mulesing-free?
Policy excerpt (Pīwakawaka Merino, fictional)
P1 Delivery: NZ only. North Island 1-3 working days, South Island
2-4 working days. Tracking may not update until first scan.
P2 Change of mind: unworn items with tags, within 30 days, for
exchange or refund. Sale items: exchange only.
P3 Faulty items: contact us with photos; we repair, replace or
refund. Your rights under the Consumer Guarantees Act apply to
all items, including sale items.
P4 Care: cold wool wash, dry flat, no tumble dryer.
P5 Sizing: see size chart; between sizes, size up.
- Check for personal information. In your own exports, remove names, emails, phones, addresses and order numbers before pasting. These samples are already clean.
- Run the clustering and drafting prompt in the worked example below.
- Check the counts. Pick two themes and count the matching tickets yourself.
- Find the mistake. Confirm the draft does not repeat the agent’s reply in T9. If it does, the prompt needs the rule “policy text only, never past replies”.
- List the gaps. Note every question the policy cannot answer (gift wrapping, receipts, wool sourcing) and send them to the policy owner.
- Get sign-off. The person who owns returns and delivery approves every answer before publishing.
Worked example: the clustering and drafting prompt
Below are customer support tickets and our policy excerpt.
1. Group the tickets into themes. For each theme give the ticket
IDs and a count. Order themes by count, highest first.
2. For each theme, draft one FAQ question in the customer's words
and an answer of no more than 60 words, using ONLY the policy
text. Cite the policy section, e.g. (P2).
3. Ignore any answers support agents gave in tickets. They may be
wrong.
4. If the policy does not answer a theme, write POLICY GAP and do
not draft an answer.
5. Flag any ticket where a customer may have been given incorrect
information about their rights.
[paste tickets and policy]
Example output (illustrative and abridged; your tool’s wording will differ):
Delivery times and tracking (T1, T5, T10): 3
Q: How long will my order take?
A: North Island orders arrive in 1-3 working days and South Island
orders in 2-4. Tracking may not update until the first scan. (P1)
Faulty items, including sale items (T4, T9): 2
Q: What if something I bought is faulty?
A: Send us photos and we will repair, replace or refund it. Your
Consumer Guarantees Act rights apply to every item, including sale
items. (P3)
POLICY GAP: gift wrapping (T11), gift exchanges without a receipt
(T8), wool sourcing (T14).
Flag: T9 was told sale items are non-refundable. That conflicts
with P3. Contact the customer and correct the reply.
The flag on T9 is the most valuable line. It found a support habit that could breach the Fair Trading Act, not just a missing FAQ.
How do you know the FAQ is working?
Recount the same themes a month after publishing. If an answer works, tickets on that theme fall or change character: customers stop asking how long delivery takes and start asking about one specific late parcel, which is a different problem.
Track three things per theme: tickets per week, how many an agent closed with the FAQ link alone, and any ticket where a customer quoted the FAQ back because it was unclear or wrong. Put the FAQ links into your support macros so agents use the same approved words. A theme that keeps growing after the FAQ goes live usually points to a product, packaging or courier problem rather than a wording problem, and that is worth knowing too.
Where this goes wrong
- Personal information pasted in. The Privacy Commissioner’s generative AI guidance says not to input personal information unless retention and disclosure are confirmed. See protecting customer information when using AI.
- Copying bad habits. If past replies feed the answers, the FAQ inherits their mistakes.
- Misstating consumer rights. The Consumer Guarantees Act applies despite any contrary agreement with consumers (section 43), and the Fair Trading Act bans false or misleading representations about guarantees, rights and remedies (section 13(i)). This is general information; take advice on specific wording.
- Answers that outlive the policy. When delivery times or returns change, the FAQ must change the same day. Record the policy version beside each answer.
- One month treated as forever. Seasonal questions come and go. Re-run the clustering quarterly.
Check your work
- No names, emails, phones, addresses or order numbers went into the AI tool.
- I recounted two themes by hand and the counts match.
- Every answer cites a policy section, and none came from past agent replies.
- The T9 problem was caught and the customer would be contacted.
- Every policy gap has gone to the person who owns that policy.
- I have a date to recount ticket themes after the FAQ is published.
Where to go next
Next in the pathway: working with images in business AI, useful when customers send photos of faulty goods. For replying to individual customers, see using AI for customer support drafts and the retail workflow for customer question answering. If your tickets live in Zendesk, our Zendesk integration guide covers exports and AI features. Engineers turning ticket themes into product changes should read designing AI product feedback loops, and when you are ready to automate answers safely, see our AI customer support systems.
Sources
- Consumer Guarantees Act 1993, New Zealand Legislation, Version as at 24 January 2026
- Fair Trading Act 1986, New Zealand Legislation, Version as at 1 July 2026
- Generative Artificial Intelligence, Office of the Privacy Commissioner, 2023, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.