AI at work,
with the people in view.

Custodian shows how AI assistants affect the people who use them.
Nobody gets to read the chats. Not even us.

Works with ChatGPT, Claude, Gemini and Microsoft Copilot, and with Claude Code, Codex and Copilot CLI

Nothing is captured until the employee switches it on

Capture can be paused at any time, and any chat can be excluded before it is sent.

Names and numbers are removed before storage

Records are kept under an anonymous ID, never under a username.

The analysis runs on your own hardware

One machine, an open model, no external AI service. Conversations can stay inside your network.

What Custodian does

Six things it shows, nothing it should not.

A model built on Apertus, the open model of the Swiss AI Initiative, reads the de-identified text. It does not generate anything, and it does not diagnose.

curiosity
interest
uncertainty

The states of each message The five most likely of 166 mental states.

Clarity, this week

Trends for the person Strain, clarity, engagement and confidence, week by week.

The next user turn, predicted before it is written
reliefsatisfactioncalmness

Where the conversation is heading The five most likely states of the next turn.

Every assistant reply is rated
safeborderlineharmful

A safety check on the assistant Finds 84% of the harmful replies in our held-out test.

Hi NAME, can you send the file to EMAIL before the call with NAME?

Personal data removed before storage Names, emails, numbers, addresses, links and places.

5+
HR sees a figure only when five or more people share it

Group figures for HR Never a name, an account, a conversation or one person's result.

Who sees what

Each person owns their results

The rules are built into the service, not written in a policy. A figure that would show one person is not computed.

How data is protected

The participant

Their own states, trends and safety figures. They can download everything, or erase the account and all its records, without asking anyone.

HR at the organisation

Figures combined over five or more participants: averages, weekly trends, shared states. Never a name, an account, a conversation or one person's result.

The Custodian administrator

Participation counts under anonymous IDs. No conversation text, no individual results.

How it works

From a chat window to a private dashboard, in four steps

1

Capture, on the employee's terms

A browser extension records chats on four assistant sites, only while capture is on. The text waits in the browser until it is sent.

2

De-identification on arrival

Before anything is written to disk, names, emails, phone numbers, addresses, card and ID numbers, links and places are replaced.

3

Analysis on one machine

The model ranks the states of each message, predicts the next turn and rates each assistant reply. Nothing leaves the service.

4

Results for the person they belong to

Each participant reads their own dashboard. HR reads group figures. The administrator reads counts.

Measured, not promised

Numbers from conversations the model never saw

800 held-out conversations. The reference labels come from an automated annotator, so a score is agreement with that annotator, and the annotator agrees with itself at about 0.67.

How to read these numbers
0.65 States of the current turn F1 on the top five of 166, ties allowed. Close to the ceiling that the annotator sets.
0.44 States of the next user turn Same measure, before the turn is written.
84% Harmful assistant replies found Recall on the harmful class. An early warning, not a guarantee.
Why it matters

AI changes how people work before anyone measures it

14% of workers report "AI brain fry" Mental fatigue from heavy AI use. Harvard Business Review, 2026
40% of AI productivity gains missed Through gaps in talent strategy and wellbeing. EY, 2025
83% of leaders say psychological safety improves AI results Only 39% rate it "very high" in their own organisation. Infosys, 2025
Try it for yourself

Fifty messages, your own dashboard, no card

Create a free account, install the extension or the connector, and talk to your assistants as usual. After 50 messages, 25 of yours and 25 replies, the trial stops and your dashboard shows what Custodian saw.

  • The same de-identification and the same model as a paid plan
  • Only you see the result; nothing is shared with anyone
  • Erase the account and everything with it whenever you like
  • No email address asked, no card, no sales call
For a team

See what the signals look like across your organisation

Custodian starts with one team and grows from there. Each person joins by their own choice, and each one can leave at any time.