An assistant that already knows the work around your request.
Andy is not a chatbox bolted onto a sidebar. Andy lives inside TAO with the customer thread, the workflow, the calendar event, and the invoice already in view — and drafts the next operational step from real state.
What makes Andy different from a chatbox
A blank chatbox asks you to bring all the context. Andy already has it. The customer thread, the workflow run, the prior decisions — Andy reads them before drafting, and shows you exactly what it read.
- No more “forget what I said earlier” — Andy works from the operational record, not a 100-message scrollback.
- No copy-paste of context. Andy is already inside the workspace the context lives in.
- No invisible actions. Every write goes through the Approvals queue with a one-line “why”.
What Andy does on a typical day
Andy drafts the next operational step for whatever you are working on. Sometimes that is a reply. Sometimes it is a calendar move. Sometimes it is a follow-up task that nobody else noticed was missing.
- Drafts email replies in your voice, with the right thread and customer attached.
- Proposes calendar events with the right attendees and duration.
- Creates follow-up tasks before they slip — and owns saying it might be wrong.
- Surfaces stalled workflows so they do not quietly age out.
Operator-controlled memory
The brain is yours. You decide what Andy can see, what it should pin, and what it should never store.
- Per-workspace memory boundaries.
- Pin facts so they outlast the originating thread.
- Mark anything as “forget this” — Andy removes it from its working memory.
- Audit log of what Andy read and what it drafted, queryable any time.
For technical readers
Andy is model-agnostic: TAO routes between large generalist models for drafting and smaller specialist models for routing, summarisation, and entity extraction. Every model call is wrapped in a retrieval step that fetches the relevant slice of the company brain for the current task; we do not throw the whole workspace at the model.
Significant writes (outbound email, calendar writes, accounting writes) require explicit human approval. We treat that as a hard rule, not a configurable default — operators can opt into auto-approve for low-risk classes, never the inverse.
Want to see this running on the apps you already use? Apply for the beta, or tell us what your team is trying to run from one place.