Chats and the AI Assistant

Every workspace includes an AI assistant that works over the data in that workspace. Its defining feature is on the new-chat screen: “Sensitive spans are detected and redacted before anything reaches the model.” Each message you send is analysed against the workspace’s chat analysis plan, sensitive spans are replaced according to the chat transformation plan, and only the redacted payload is sent to the AI provider — you see your original text, with the detections highlighted.

Starting a chat

There are three ways in:

The Chat button in the top bar (⌘J), then New chat.
The ask-bar on a workspace card on the All workspaces page — type a question or pick a starter prompt.
The chat prompts on the workspace Overview.

The Start a conversation form takes your opening prompt — describe the task, paste the text, or ask a question. An optional Model & options disclosure lets you pick an Agent, a specific Model (only models whose provider has an active credential are listed; leave blank and the agent picks), a pinned Credential, and a Temperature. Choose Start chat.

The conversation view

A chat has three columns:

Chats sidebar (left) — all your chats in this workspace, grouped by day, searchable, with a + for a new chat. Each entry shows the chat’s name and a short summary; the name and summary are generated automatically, and the chat owner can regenerate them with the Re-summarise action in the header.
Message thread (centre) — the conversation. Replies stream in live; a typing indicator shows while the assistant is working.
Info bar (right) — details for the chat or a selected message (below).

The composer works as you’d expect: Enter to send, Shift+Enter for a new line — and, as the helper text under it notes, every message is “scanned for PII before it reaches the model”.

Seeing what was redacted

Your messages render with inline highlights over every span the detectors found — names, addresses, identifiers, whatever the workspace’s categories cover. Two controls on each message bubble go deeper:

Show redacted view — flips the bubble between what you wrote and the literal redacted payload that was actually sent to the model, substitution tokens and all. Flip it back at any time.
Show info — loads the message into the right-hand info bar, with a Markers panel listing each detected span (linked to its marker and category) and a Metrics panel for the message.

At the chat level, the info bar’s Markers panel groups every detection in the conversation by category, and Metrics summarises the whole chat. Clicking an inline highlight focuses that specific detection.

Detection is also not a one-shot pass: after each turn, a background deep scan re-analyses the conversation with a heavier detector chain (using the workspace’s deep-scan plans) and additively redacts anything it finds in earlier turns. A span that was initially missed can therefore become highlighted — and redacted — shortly after.

Privacy and oversight

Chats are private to their owner. Only you can open your conversations, post messages, re-summarise, or delete them.
Oversight is separate and read-only. Users granted the chat-view permission can see chat activity, exposure metrics, and per-message detection info through the workspace’s AI dashboard — this is how the workspace measures residual risk in AI usage.
Activity is audited. Starting a chat and posting messages emit audit events, visible to administrators in the audit log.

Which model answers?

Each chat shows the model it uses in its header (“Default” means the agent picks). What’s available is controlled per workspace: providers and models are linked to the workspace under AI → Providers / Models, and a provider link carries the default credential and plans. If the model you need isn’t listed when starting a chat, its provider has no active credential or the model isn’t linked — ask whoever manages the workspace.