Sensitive data

What does an AI do with sensitive information a customer slips into an email?

When a customer slips sensitive information into an email, a Muzau agent's AI reads it to reply, but keeps it out of what it learns. Health, religion, political opinions, origin, sex life, union membership, private life: a word list and the AI's own judgment stop them before the Brain, its memory. Nothing it draws from an exchange gets in without your "Keep" click. After a meeting, the same filters keep them off the customer's record.

Why a customer slips sensitive data into their emails

A customer moves a meeting and explains they are just out of hospital. Another apologizes for a long silence with a bereavement, or warns that a date falls on a religious holiday. They are not writing to be filed: they are writing to be understood, by someone they know.

With an AI that reads your exchanges, the risk is not that it reads this message: it has to read it to reply. The risk is that it turns it into knowledge, which would resurface in a draft or on the customer's record, in front of the team. Health and religion are among the data the GDPR protects most strictly. They are also confidences a customer does not want to see come back.

How the agent goes about it

Input

An email, an exchange or a meeting note

It does

Replies to the customer, then keeps sensitive topics out of what it learns

It delivers

Facts to approve and a customer record, sensitive topics left out

Two filters keep sensitive data out of the AI's memory

The Brain is what all your agents know about your company to reply: your offers, your lead times, your rules, your tone. Before a fact drawn from an exchange, a document, a meeting or one of your edits becomes knowledge, two filters run. A word list maintained in code spots health, political or religious opinions, origin, sex life, union membership and private life (an illness, a bereavement, a divorce). The AI that proposes the fact also judges: it catches the phrasings the list does not know. It is enough for one of the two to see a sensitive topic for the fact to be set aside, even when the customer wrote it themselves.

What the AI learns from an exchange waits for your "Keep" click

A fact that passes both filters is still only a proposal: nothing that comes from an exchange, a document or a meeting enters the Brain without your "Keep" click. If a phrasing had slipped past the filters, that is where you set it aside. When the Brain rereads the emails you sent to learn from your real replies, email addresses, phone numbers and references preceded by a "#" are masked before reading. Names stay readable. For other mistakes, see whether an AI that writes to your customers can get it wrong.

After a meeting, sensitive topics kept off the customer's record

After a meeting noted in Granola (see meeting follow-up), what the customer said themselves is filed on their record, in "Customers met": their need, their budget, their deadline, who decides, their objection. A sensitive topic is kept out of it by the same word list and by the AI's judgment. The record, though, fills in without a "Keep" click. Muzau does not keep the note itself: each item carries the meeting date and the link to the note. Your team's replies (a lead time, an offer) are proposed to the Brain, and the AI is instructed to leave out what is specific to the customer.

What to do with sensitive information a customer confides

With or without AI, the rule fits in one sentence: answer the request, note the consequence, leave the cause in the email.

• Answer the request, not the confidence. A human word is enough, without dwelling on the detail: "Thank you for letting me know, let's move it, no problem. Next Tuesday or Thursday, which suits you?"
• Note the consequence, not the cause. In HubSpot, Pipedrive or your tracking sheet, "unavailable until the end of the month, get back in touch then" is enough; the illness has no place there.
• Choose tracking columns that do not invite a confidence: the need, the budget, the deadline, who decides, rather than a "personal situation" box.
• Do not come back to the confidence in the next follow-up or recap. If the customer raises it again, follow their lead.
• When you hand an exchange to a colleague, summarize without the confidence. They need to know what to do, not why.
• If a confidence is already noted in a record or a shared file, delete it and keep its consequence.
• Do not copy a confidence into the memory of an AI tool or into an email template: at Muzau, what you write yourself in the Brain is not filtered.

If a chat also collects these messages on your site, see what to say in your privacy policy. For data hosting and how AI models use it, see the AI glossary.

What it does not do

• It does not guarantee that everything is spotted: a phrasing that escaped the word list and the AI's judgment could get through, to the Brain as to the record of a customer met; for the Brain, your "Keep" click remains the last check.
• It does not filter what you write yourself in the Brain: you chose it, with the text in front of you.
• It does not remove the information from the message received: the agent reads the exchange to reply, and the sorting applies to what becomes knowledge in the Brain or goes onto the record after a meeting.

Frequently asked questions

Can an AI remember the illness a customer mentions in an email?

Not in the Brain, the memory Muzau agents draw on to reply. Health is among the data set aside before it gets in, by a word list maintained in code and by the AI's judgment, even when the customer writes it themselves. And nothing drawn from an exchange gets in without your "Keep" click.

What happens to the note of a meeting read by an AI?

With Muzau and a Granola note, the note itself is not kept. The customer's record keeps what they said that is useful for what follows (need, budget, deadline, who decides), dated and linked to the note, sensitive topics left out. Your team's replies are proposed to the Brain and wait for your click.

What happens if sensitive data slips past the AI's filter?

If it comes from an exchange or a meeting, it stays a proposal: nothing enters the Brain, the memory of Muzau agents, without your "Keep" click, and you set it aside. The record of a customer met, though, fills in without a click: a phrasing that got past both filters could be filed there.

Are customers' addresses and numbers read when the AI learns from your emails?

When Muzau's Brain rereads your sent emails to learn from your replies, email addresses, phone numbers and references preceded by a "#" are masked before reading. Names written in plain text are not.

What should you do if a customer gives bank details by email?

Confirm them through another channel before using them, for example by phone, on a number you already had: this is how a transfer fraud starts. With Muzau, a message where an IBAN is spotted, in its text or as an attachment, is never handled alone, whatever your setting. An IBAN slipped into an image is not spotted.