AI safeguards
Privacy mode
The first time you sign in, you choose AI on or privacy mode, and you can change it any time in Account. In privacy mode nothing you do is sent to any AI model, and nothing in My Work is either. Anything already running stops, and your files at Anthropic are deleted. With AI on, every tool has its own switch.
Your switches cover you. Teammates who have AI on can still use it on work your team shares. On a team, an agent runs only when both the task's owner and the person who put it to work allow it.
What each tool reads
Each tool gets only the data its job needs. Each switch's ? card in Account says the same.
- Suggestions
Suggests links, impact, effort, and an owner for a task, and which agent fits each task in My Teams' lists.
Reads The task and others near it: titles, the first 160 characters of notes, owners, projects, due dates, status, impact and effort, and the names of people who could own it. For the agent: each listed task's title, the first 400 characters of its notes, and its files' names.
- Brain Dump review
Turns new notes into tasks for you to check.
Reads That new text, today's date, your name, your teams with their projects and people's names, and the titles of up to 150 open tasks, so it skips duplicates.
- Ask
Answers questions about your work on the Map.
Reads Only what it looks up as it answers: tasks you can see, with their owners, dates, status, and links.
- Update drafts
Drafts a team's weekly update for you to edit.
Reads What changed that week: tasks finished, moved, or added, with their titles and owners.
- Plan with Specificity
Plans projects with you. It drafts a plan from an imported file, turns a brief into a draft project, writes a brief from what you describe, and drafts sections of your docs. It asks up to 5 questions first, and you review everything before anything is made or changed.
Reads For an import, the file's titles, notes, and paragraphs (owners and dates stay out). For a doc you draft in or turn into a project, the doc's whole text, names and dates included. For a new brief, what you type and your answers. Nothing of it is kept, docs labeled No AI are never read, and anything with a sensitive number is refused before Claude sees it.
- The six agents
Document Writer, Proofreader, Research Analyst, Deck Designer, Spreadsheet Builder, and Tracker Updater.
Reads Its task (title, notes, files, and earlier work), the tasks it waits on and their handoffs, your team's and project's tasks (titles, owners, dates, and status), your Library (instructions, writing samples, templates, and reference files), and the Docs where it works (never a No AI doc; an Internal doc only for its own team's agents). The Document Writer saves the documents it writes as docs. The Proofreader also reads the document it reviews, the Deck Designer your PowerPoint template, and the Tracker Updater the tracker it updates.
- Web search
Lets the Research Analyst search the web. It has its own switch.
Sends Search queries to a public search engine through Anthropic. It's instructed to search only for the general question, never names or internal details.
Some work never goes
- Tasks and docs labeled No AI are never sent to an AI model, whatever your switches say.
- Suggestions, Brain Dump review, Ask, Update drafts, and Plan with Specificity never read tasks labeled Restricted.
- Tasks labeled Internal or Restricted never use the web, nor does an agent's run once it has read an Internal doc.
Where AI work goes
Specificity's AI tools use Anthropic's Claude. Anthropic says it deletes API inputs and outputs within 30 days, unless they're flagged for breaking its usage policy (then it can keep them up to 2 years), and that it doesn't train on them by default (how long data is kept; model training).
When an agent works with files, they go to Anthropic's Files API and code execution container while it runs. We delete them when the run ends, or at once when you turn AI off, and an hourly check removes any left behind. The container can keep data up to 30 days (code execution).
AI output can be wrong. Check it before you rely on it. The privacy policy has the rest.