Asking your record
Kennedy answers questions about the office's live record in plain English — which categories are breaching SLA, who wrote in about the bus line, what happened on a case last spring. He queries the record itself, not a stale export, and shows what he used. He lives in two places: the Kennedy tab — his full workspace — and a concierge you can summon from any screen (Esc dismisses him, Enter sends).
Ask operational questions, get operational answers: counts, lists, patterns, the case numbers behind them. He respects permissions — a staffer who can't see campaign history can't get it out of Kennedy either. Type / for commands — /case, /constituent, /draft, /summarize, /sla.
What he can reach
Kennedy grounds every answer in a fresh snapshot of the office plus a retrieval pass across your cases, events, bills, and meetings — so he's reading this office's record, scoped to what your role is allowed to see, not the open internet. On top of that he can call a small, named set of tools, each filtered per session to your permissions and the office's features:
NOTE — DRAFTS ONLY
A draft tool only ever writes a draft row — it never publishes or sends. In the chat it shows up as a card that reads “Saved as a draft — finish it in Comms ▸ …”, so the thing that leaves the office is always the one a person chose to send.
How far Kennedy's tools reach depends on the office's configured AI provider; with some providers he grounds his answers in retrieval alone. Either way the safety line is the same — read and draft, never send.
Drafts & approvals
Copilots inside each module do the repetitive writing: categorize inbound mail against the office's categories, draft replies in the office's voice, summarize a six-month thread before the constituent calls back.
Kennedy only ever drafts. Every artifact he produces waits in a pending state for a staffer to edit, approve, or discard — nothing reaches a constituent, a reporter, or the public feed on his own signature.
Every Kennedy action lands in the audit log with what he read and what he produced. If the reply was wrong, the trail shows exactly where the human review should have caught it.
Needs your okay
Anything Kennedy wants to do — send the NJDOL nudge, reply to a constituent — lands in the Needs your okay rail beside the conversation, labeled with what it is and when he queued it. Approve it and it sends; decline it and he notes why and keeps the draft.
The rail is the whole safety model in one column: Kennedy reads freely, drafts eagerly, and acts never. What he read and produced is in the audit log either way — and the → Open agents link shows any longer-running work.