Skip to content
Open Bluedeer

Choose what PII the AI sees in a workspace

How the owner or an admin of a shared workspace sets what happens to names, contact details, card numbers, passwords and other PII before it reaches the AI, where the rules apply, what admins see, and what the checks can't do.

In a shared workspace, Bluedeer can look for personally identifiable information (PII) in what is about to go to the AI, such as names, email addresses, card numbers and passwords, and handle each kind the way the workspace chooses. Personal workspaces don't have these checks. In a shared workspace, the checks are part of Bluedeer Enterprise.

People in the chat always see the real values. A value hidden from the AI is swapped for a placeholder that names its kind, such as a name or an email address, just before the message goes to the AI, and put back in the answer before anyone reads it. The same value keeps the same placeholder for the whole chat, so the AI can still follow who is who.

Where to find it

On the web: select Workspace in the sidebar, then Settings, and open the Compliance tab.

On the phone: on the home screen, tap the button at the top right, then Manage workspace, and tap Compliance under Rules.

Everyone in the workspace can open it and read the rules. Only the owner and admins can change them.

Part of Enterprise

PII checks are part of Bluedeer Enterprise. On the Team plan, the Compliance tab shows them locked, with an Enterprise badge, and the rules can't be changed. If checks are on when the workspace leaves Enterprise, they keep working. Without Enterprise the owner or an admin can make them stricter or turn them off, but not loosen them: a kind of PII can move up from Allow to Highlight or Hide from AI, and another place can be checked. Moving a kind down, or unchecking a place, needs Enterprise, and those choices say why when you hover over or focus them. Turning checks off is always allowed, but turning them on again needs Enterprise.

The same tab holds Where AI is processed, which is also part of Enterprise. See Choose where a workspace's AI is processed.

Turn checks on or off

Censor PII at the top turns the checks on or off for the whole workspace.

  • On: from that moment, Bluedeer checks new messages and connector results, as chosen under Applies to, and handles each kind of data as set below.
  • Off: new messages and connector results go to the AI as they are. Anything that was already hidden stays hidden. Bluedeer asks you to confirm before it turns checks off. Your choices are kept, dimmed, for when you turn checks on again.

Checking a message usually takes a fraction of a second and runs while Bluedeer prepares the reply, so most replies start just as quickly. Very long pasted text can add a little time. While checks are off, new messages aren't checked and no time is added.

Turning checks on or off, and changing a rule, each add a line to the workspace's Activity.

Messages sent earlier

Checks apply to each message when it is sent, and the result stays with that message:

  • A message sent while checks were on keeps what was hidden in it. Turning checks off later doesn't show those values to the AI.
  • A message sent while checks were off is never checked later, even after checks are turned on again. It goes to the AI as it was sent.
  • A chat can therefore contain both kinds of messages if checks were turned off and on during the chat.

This also holds for Applies to: messages or connector results sent while that kind wasn't checked aren't checked later.

A chat continued from a share link, or imported from another app, into a workspace with checks on is checked once when it is brought in.

We recommend turning checks on when you create the workspace, so every message is checked from the start. On Enterprise, a new shared workspace starts with checks on. On the Team plan, checks start off and can't be turned on.

Choose what the AI gets for each kind of data

Under PII, each kind of data has one of three choices:

  • Allow: goes to the AI as it is.
  • Highlight: goes to the AI, and is marked for the person who sent it, so they notice what they shared.
  • Hide from AI: swapped for a placeholder before the AI sees it, and put back in the answer.

The kinds are grouped:

  • Secrets: API keys and passwords; bank, IBAN and card numbers; ID and passport numbers.
  • People and contact details: names; emails and phone numbers; addresses; IP addresses.
  • Other: personal web addresses; dates tied to a person, such as a birthday; other identifiers, such as usernames.

On Enterprise, a new shared workspace starts with checks on, Secrets set to Hide from AI, and everything else set to Allow.

Select Save after changing a rule. New messages follow it from then on. From a chat's next message, the rules also apply to its older messages that were sent while checks were on. Messages sent while checks were off stay as they were sent.

Applies to

Choose where Bluedeer looks:

  • Typed messages: what people write or paste in a chat, and what they say in a voice message.
  • Connector results: what apps such as email or a CRM send back to the AI.

A voice message counts as a message: what the person said is checked, and values are hidden from the AI, just as in typed text. Only the sound itself isn't checked.

Files aren't checked. What is attached to a message, and what the AI reads from Files or Drive, goes to the AI as it is: documents, PDFs, images, videos, audio and pictures from pages. This also covers files sent earlier in a chat, and nothing is hidden in a file, even a value that is hidden in a message.

Results the AI gets after someone approves an action, such as an app's data or a shared location, are checked as connector results when they arrive. A scheduled task's result posted into a chat is checked when it is posted.

Mask keys and passwords in tool results

Where Bluedeer offers it, the Compliance tab also shows Mask keys and passwords in tool results. While it is on, keys, tokens and passwords in what tools and commands return are replaced with a placeholder such as [hidden: password], before the AI reads them or the chat keeps them. The AI can't use a hidden value either. Choose which kinds are masked, such as known key formats, private key blocks or passwords in connection strings; at least one kind stays on while masking is on. It works only while Censor PII is on, and only the owner or an admin can change it. If Bluedeer stops offering it, it can still be turned off. Without this option, nothing in tool and command results is masked, including values people saved with Secure input.

Activity

The owner and admins see counts for the last 30 days: how many values were hidden from the AI, how many were highlighted, and how many messages couldn't be fully checked, with a bar for each kind of data. Members don't see these counts.

Bluedeer never records the values it finds, only where they were in a message and what kind they were.

What stays the same

  • The chat's history stays as it was written. People read their messages and the AI's answers with the real values.
  • Only the person who sent a message sees what was hidden or highlighted in it. Others in the chat, shared links and exports see the chat as usual.
  • Tools the AI uses, such as sending an email or adding a calendar event, get the real values, so they work as before.
  • Web search, reading a web page and creating images never receive a hidden value. If the AI tries, that step doesn't run and the AI is told why.

What the checks can't do

  • Detection of PII is not perfect. Most of it is found, but some can be missed, especially secrets written in an unusual form. Company names are not checked.
  • Files aren't checked for PII. This includes images, videos, audio and scanned documents, and the text Bluedeer reads out of a document. What someone says in a voice message is checked, as typed text is.
  • If the checks can't finish in time, the message is still sent. Passwords and keys are still hidden, the sender is told the message was checked more lightly than usual, and it is counted under Couldn't check. Later messages in the chat send it to the AI with the detected values hidden.
  • Text sent while checks were off is never checked, and the AI has already read it. Text sent before a rule was made stricter has already reached the AI in earlier replies; the stricter rule applies to it from the chat's next message.