Lorixen.
EN

Building an FAQ from your inbox with AI: a worked example

Short answer: Your inbox already contains the questions customers actually ask. Export a few months of enquiries, strip personal details, use AI to group the questions and draft answers, then have someone who knows the business correct every answer before it goes on the site. The result is an FAQ built on real demand, which you can also use to brief staff or a chatbot. Google no longer shows FAQ rich results, so write it for customers, not for a search feature.

This article walks through the process using a worked, hypothetical example, so you can see what each stage produces and where the human checks belong.

The hypothetical business

A small bicycle repair and hire shop in Bristol answers 15 to 30 emails a week. The owner suspects most are repeats: opening hours on bank holidays, whether they fix e-bikes, how long a service takes, deposit rules for hire. The existing FAQ page has six questions written years ago. The goal: a useful FAQ and fewer repetitive emails.

Stage 1: export the raw material

The shop uses Gmail. Google's Takeout help explains how to download data from a Google account, including Gmail, which exports in a mailbox file format and preserves each message's labels in a special header. Most email providers offer a similar export, or you can simply copy a sample of messages into a document.

What the owner did: exported the "Enquiries" label covering the last six months, then pasted the text of the customer messages (not replies) into a spreadsheet, one per row.

Human check: remove names, email addresses, phone numbers, order numbers and anything personal before any AI tool sees the text. A quick find-and-replace plus a manual skim is enough for a few hundred messages.

Stage 2: let AI group the questions

The owner gave an AI assistant the anonymised messages with instructions:

  • Identify the distinct questions customers are asking.
  • Group similar wording under one question.
  • Count roughly how often each appears.
  • Quote one or two example phrasings for each group.
  • Use only the messages provided; don't add questions that aren't there.

AI vendors' own guidance supports this kind of grounding. Anthropic's documentation on reducing hallucinations suggests asking the model to extract word-for-word quotes before doing a task with long documents, and to cite supporting quotes for each claim, retracting anything it can't support. Asking for example phrasings serves the same purpose: you can see each group really came from the inbox.

What came back (illustrative): about 25 question groups. The top ones: e-bike repairs, turnaround time for a full service, hire deposit and ID, bank holiday hours, whether bikes can be dropped off before opening, puncture repair while you wait, and child seats on hire bikes.

Stage 3: draft answers, then fix them

Next, the owner asked the AI to draft short answers using a brief of facts: the pricing approach (no prices on the FAQ; link to the price list instead), hours, hire terms, what brands of e-bike they can and can't service.

Human check, the important one: the owner corrected every draft. Common fixes:

  • The AI said "most e-bikes" could be serviced; the truth was "we service mechanical parts on all e-bikes, but motor and battery faults go to the brand's dealer". Precise beats reassuring.
  • Turnaround was stated as "24 hours"; the real answer depended on parts and season. The final answer explained the usual range and how customers are kept updated.
  • A deposit answer used wording from another shop's typical terms, not this shop's. Rewritten from the actual hire agreement.

Anthropic's guidance also suggests explicitly allowing the model to say it doesn't have enough information. In practice, telling the AI "if the brief doesn't contain the answer, write NEEDS OWNER INPUT" produced placeholders instead of invented answers, which made review faster.

Stage 4: structure the page for people

The owner chose about 15 questions (the rest were rare or belonged elsewhere) and grouped them under headings: Repairs and servicing; Hire; Visiting the shop. Each answer is two to four sentences and links to the relevant page (price list, hire terms, booking form).

Google's Search Central updates log notes it removed documentation for the FAQ rich result because the feature is no longer shown in Google Search results, following a May 2026 announcement. So the page is designed for visitors: scannable headings, plain answers, links onward, no need to chase a search feature.

Stage 5: reuse the answers

  • Email templates: the same answers become saved replies for the team, keeping wording consistent.
  • Chatbot or AI phone assistant: if the shop adds one later, the verified FAQ is a good knowledge source because a human has already checked it.
  • Service pages: the most common questions (e-bike servicing scope, turnaround) also got a short section on the relevant service page.

Stage 6: keep it current

The owner set a quarterly reminder to skim the latest enquiries and update the FAQ. Any time a policy changes (hire terms, hours), the FAQ and saved replies change the same day.

What the example shows

  • The inbox is a better topic source than guessing.
  • AI is fast at grouping and drafting, but answers about your policies must come from you.
  • Anonymising first is quick and avoids putting customer details into tools unnecessarily.
  • A verified FAQ pays off several times: website, email replies and future automation.

FAQ

How many emails do I need to build a useful FAQ?

A few months of enquiries is usually enough to show patterns. Even 100 messages can reveal the top ten questions.

Is it safe to paste customer emails into an AI tool?

Remove personal details first and check the tool's data settings and terms. Many businesses use business plans or settings that limit how data is used.

Should the FAQ include prices?

Prices change, and an outdated FAQ answer causes friction. Linking to a single, maintained price list is usually easier.

Do FAQ pages still help SEO?

Google no longer shows FAQ rich results, but a page that genuinely answers common questions can still be useful to visitors and can be found for those questions like any other page.

Answering the same customer emails every week? Send us a sample of anonymised questions and we'll help you turn them into an FAQ and saved replies.

Sources

WhatsAppLet’s talk