Give your AI assistant the context, not the card numbers

Share the context, hold back the values

The notes worth giving an assistant are the working ones: the offer you’re drafting, the vendor call, the runbook with the staging credentials in it. They’re also the notes with a salary, a card number or an API key somewhere in the middle. Until now the choice was all or nothing — connect the assistant and it reads everything in the note, or keep the note out of reach and lose the context.

Dexi Guard is the third option. Switch it on for a connected app and that app reads a copy of each note with the restricted values replaced — <CREDIT_CARD_1> where the card number was, <COMPENSATION_1> where the salary was. Everything around them stays: who the call was with, what was decided, what happens next. Your own view of your notes never changes, and nothing is removed from them.

What gets hidden

• Payment and bank details — card numbers, IBANs, routing numbers and labelled account numbers. Free-standing numbers are only treated as a card or IBAN when their checksum is valid, so order numbers and version strings are left alone.
• Government identifiers — US Social Security numbers and ITINs.
• Credentials — API keys, access tokens, private keys, passwords in password = … lines and inside connection strings.
• Pay figures — a money amount on the same line as a word like salary, bonus, compensation or “per year”.
• Sensitive properties — the value of a note property whose name says what it is: salary::, bonus::, ssn::, card::, password::.

A Strict level adds email addresses and phone numbers. The same value always gets the same placeholder within a note, so an assistant can still tell two cards apart and follow one through the text.

It can’t search for what it can’t see

Hiding a value in the text isn’t enough if the assistant can still ask “which notes mention 185000?” and get an answer. So a guarded connection doesn’t only read the de-identified copy — it searches it. Keyword search, pattern search and search by meaning all run against the copy, never your original, and so do tag filters. A search for a hidden value finds nothing.

The same care applies to writing. A guarded assistant can add notes and append to existing ones as usual, but it can’t replace the text of a note that has hidden values in it — writing back the placeholders would overwrite the originals.

One policy for the whole team

On a team, an owner or admin sets one policy that covers every member’s notes, and can require Dexi Guard for everyone — each person’s connected apps are then guarded, and members can’t switch it off. The policy is where a team adds what only it knows:

• Blocked words — a client’s name, a project codename. Hidden wherever they appear.
• Property rules — hide any property by name (client*, margin), or replace it with a stable code so an assistant can group notes by client without learning who the client is.
• Hide the whole note — for teams that would rather an assistant never see a note with restricted values at all.

A preview box in Settings shows any text the way a guarded assistant would receive it, before the policy is saved.

How to turn it on

1. Connect an assistant, if you haven’t — see connect an AI assistant to Dexi.
2. Switch on Dexi Guard for that app: on the approval page when you connect it, or any time in Settings → Connected apps → Edit access. It applies on the app’s next request; nothing needs reconnecting.
3. Let it check your notes. A note becomes visible to guarded apps once its de-identified copy exists. New and edited notes are checked within moments; a large library is worked through in the background, and Settings shows the progress.

Three controls on every connection

Dexi Guard sits beside the two controls every connected app already has. They answer different questions, and they combine.

What Dexi Guard doesn’t do

Guard finds values by their shape and the words around them. That makes it dependable for the things that have a shape, and it means there are things it won’t catch:

• It does not detect names, postal addresses or company names, and it can’t recognise a confidential fact written as ordinary prose. Add the specific ones you care about as blocked words.
• It errs on the side of hiding pay: an amount next to a word like “annual” is replaced even when it’s a budget rather than a salary.
• It covers apps connected over MCP. A share link or a published feed shows a note as written to whoever you give it to.
• Bases aren’t available to a guarded app, because a base filters and totals on the real property values.

Treat it as a safeguard for the values that most often leak by accident, not as a guarantee that nothing sensitive is in a note.

Frequently asked questions

Keep reading