Reviewed September 24, 2026. Features depend on your enabled modules, permissions, and configured integrations.
Tip: permissions
Rosie needs the ai feature and the use_rosie permission (roles that had view_analytics when it was split out got it automatically). Every tool she reaches checks your own permission for that desk: drafting comms needs manage_comms, texting also needs the texting feature, and the working set is gated by run_agents. Your campaign's own AI provider does the thinking; the model comes from your integration settings.
She lives in two places
Rosie has one brain and two front doors. The full-page Rosie tab is a three-column operator console: your conversation history on the left, the chat in the middle, the working set of things awaiting your okay on the right. The concierge is a small launcher that floats in the corner of every other view, so you can ask a question without leaving the turf map or the call-time queue. Same conversation, same data, same rules, in a smaller window.
Type / for commands, and the chat is threaded and saved per campaign, so yesterday's question is still there this morning.
Ask your own data
Rosie answers from your campaign's live data, not a general model of the world. Ask her to find voters, count a universe, check field progress, or pull the campaign and fundraising numbers, and she reaches for the right tool and shows you what she used. The answer comes back as a structured card you can act on (a voter list you can open, a count you can save as a universe, a stat block, a donor summary), not a paragraph you have to re-key.
Because she works from live data, the numbers match what you'd get by hand, and because she works within your permissions, a staffer who can't export voters can't get an export out of Rosie either. She can make the small edits a staffer makes in passing (log a contact, set a support level, add a note, create a task), but only when you hold the permission for it, and each one is written to the audit log. Nothing leaves the building on her say-so.
Drafting across the desk
Rosie doesn't just read. Ask and she'll draft a press release, an email or SMS blast, a blog post, a social post, an editorial, or a rapid-response statement, each written on message with your pillars and brand, and each handed back as a real draft you open and edit. Nothing sends inline; the Communications guide covers exactly what "approve" does for each channel.
✶ Rosie: a model from plain English
In the Voters ▸ Scores builder, Rosie drafts a voter model as a set of toggleable assumptions, each with a confidence you can dial, and trains it on your real canvass data as a cancelable job, then shows what she assumed against what the data learned. The full walkthrough is in the Voters guide.
She reaches further into the desks than she used to. She can break down doors, calls, and texts by day, week, canvasser, or outcome; follow donor activity; report on the Modeling desk and explain why a particular voter scored the way they did; and read past election results for a town or precinct. Ask her to plan a canvass and she stages the event, the turf cut, and the invites as a card you approve step by step (the Field guide has the details).
Request a briefing from recent campaign activity, or configure a daily briefing agent separately. Opening Rosie does not automatically schedule a briefing every morning.
The working set: she drafts, you decide
Anything Rosie wants to actually do (send a blast, publish a post, run a longer task) lands in the working set beside the chat, and as a quiet nudge on the concierge. It's backed by the real pending-approvals queue: each item is labeled with what it is and when she queued it. Approve it and it happens through the same compliant path you'd use by hand; decline it and she keeps the draft and notes why.
That's the whole safety model in one column: Rosie reads freely, drafts eagerly, and sends nothing on her own signature. What she read and what she produced is written to the audit log either way, and AI requests are logged separately with the provider and model that answered, so nothing she does is a black box.