Your Business FAQ: Building It From Client Emails With AI, Without Personal Data
By Anis Hammouche·November 11, 2026
Your client emails already contain your business FAQ: the questions that keep coming back, the hesitations, the misunderstandings. AI can group them by theme and propose a first version of answers. Before that, personal data has to go. Replacing names is not enough: a public FAQ must contain nothing that lets someone recognize a person.
Why this method works
A useful FAQ answers the questions your clients really ask. You already know them, but they are scattered across your inbox. Rereading two hundred messages by hand takes hours, and the rarer questions get forgotten.
Here, AI sorts and groups, a repetitive reading task. It does not know what is true for your business. That is what separates a useful FAQ from one that promises what you cannot deliver.
Anonymize first
The CNIL distinguishes anonymization from pseudonymization. Anonymization makes any identification of a person practically impossible, and the process is irreversible. Pseudonymization replaces direct identifiers with aliases, but the data remains personal, because re-identification is often still possible.
In practice, a pseudonymized client message remains personal data. For an FAQ, you must therefore go further. The CNIL describes three criteria for judging whether a dataset is anonymous: a person cannot be singled out, data about the same person cannot be linked across records, and information about someone cannot be inferred with near certainty. Source: the CNIL page on the anonymization of personal data.
In practice, remove:
- names, first names, email addresses, phone numbers and postal addresses;
- order, contract, invoice numbers and internal references;
- exact amounts and dates that could point to a person;
- situation details (place, company, project) that would make the client recognizable.
Replace them with generic markers: "the client", "a services company", "a renovation project". Then check the result: if someone close to the client would recognize the situation, it is not anonymous.
| Part of the message | What you do with it |
|---|---|
| Name, email, phone | Remove |
| Order or invoice number | Remove |
| Project or situation detail | Generalize, or remove |
| The question itself | Keep, since it is what the FAQ is about |
| Tone and emotion | Keep if it helps the answer |
The method, step by step
- Gather the messages. Take a set period, for example the last three months. Export them into a single file, without sending them to a tool before anonymizing.
- Anonymize. Apply the rules above. Reread a sample to check that no identifier remains.
- Ask for a grouping. AI groups the questions by theme and counts occurrences. You get a list of candidate questions.
- Write the answers from your real policy. Give the AI your exact conditions: deadlines, prices, terms. Without these, it invents.
- Reread each answer. Check figures, deadlines and commitments against your documents. Remove any answer you cannot keep.
Here is an instruction that limits drift. It contains no real data.
Here is a list of anonymized questions received from clients (no personal data).
1. Group them by theme and state the number of questions per theme.
2. For each theme, propose one FAQ question worded the way a client would ask it.
3. Do not invent any deadline, price or condition. When information is missing, write [TO COMPLETE].
The [TO COMPLETE] markers matter. They force you to fill in the details that commit your business.
What to avoid
An FAQ generated without rereading tends to go wrong on the same points: delivery or response times, guarantees, refund conditions, the scope of services. These are commitments. A wrong answer in a public FAQ becomes a promise the client can hold you to.
Avoid vague answers too. "It depends on your situation" answers nothing. If the answer really depends on the situation, say which criteria matter, and point to a contact.
Finally, do not publish a client testimonial taken from your emails, even anonymized, without their explicit consent. A quote can be recognizable to its author.
Frequently asked questions
Is it enough to replace names to anonymize? No. Replacing names is pseudonymization. The message remains personal data, because other details can identify the person. Remove dates, amounts and precise situations too.
Can I send my raw emails to an AI so it anonymizes them? You would then be sending them to a tool before any anonymization, which is exactly what we want to avoid. Anonymize first, on your own machine, then send the result. For sensitive data, avoid tools you do not control.
How many questions does a useful FAQ need? There is no official figure. Start with the questions that come back at least two or three times. A question asked once can stay in your emails.
Do I need to update the FAQ? Yes. An FAQ becomes wrong when your conditions change. Reread it whenever your prices, deadlines or offers change, and at least once a quarter.
