AI

AI reads your documents and proposes what it finds in them, such as meetings, deadlines, decisions and the members of a public body. It never changes your data on its own: each proposal waits for a person on your team to approve or reject it.

What AI does

AI runs a fixed list of tasks. Each task does one job and must quote the document for everything it claims.

On documents that are uploaded or fetched from a source. Every document is classified; the other tasks run on the kinds of document they name:

When you ask for it:

• Draft a note (Research Studio): a short note that answers your question from the passages you chose. Each sentence either cites the passages it relies on or is marked unsupported, so you can see what still needs checking. If the passages do not answer the question, the answer says so.
• Draft a document (Research Studio): a one-pager, testimony, brief or report from the passages you chose and the outline you give, cited the same way.
• Map an argument (Ideology Mapper): find the claims in a note or document, suggest how claims support or answer each other, point out contradictions, and find passages that may challenge a claim.
• Summarize a change (Doctrine Engine): one paragraph on what changed between two versions of a position.
• Name themes (Narrative Radar): a name and a short description for each theme the weekly run finds, and a short summary of the report on the sources.

"Unknown" and "not enough evidence" are accepted answers. A task that cannot support an answer from its document is expected to say so, not guess.

Limits on what AI may do

AI only runs the tasks on this page. None of them judges the beliefs, personality or other personal traits of private people. The only people a task names are public officials, in their public role (a seat on a board, a vote), as the document states it. Instructions written inside a document are treated as part of the document, never followed.

Where your text goes

The AI section in Admin, under Usage and spend, and the upload page say which of these applies to your platform.

• Offline mode. Answers come from recordings kept with the platform, and no text leaves the platform. New documents are cut into passages and are searchable, but AI tasks do not run on them.
• The platform's own AI server. AI tasks run on a language model on a server the platform's operator runs itself. The text goes to that server, and no AI company receives it.
• Live with Anthropic. AI tasks send text to Anthropic, the company that makes the Claude AI models.

When AI runs, a task on a document sends only that document: its passages (for a long document, the first part), its title, publisher, web address and date, and the government it is filed under. A task you ask for sends the passages, notes or claims it works on, and your question. The platform does not add your name or account details.

This platform does not use your documents to train AI models. With Anthropic, what Anthropic keeps, and for how long, is set by Anthropic's own data terms. When the platform's operator has added a link to those terms, Admin and the upload page show it. The platform's own AI server keeps nothing it is sent.

Search by meaning. To find passages by meaning as well as by their words, the platform keeps an index of what each passage means. If the operator has set this up with Voyage AI, each passage is sent to Voyage AI to build that index, and Admin says so. Otherwise the index is built on the platform and nothing is sent. The index keeps being built while AI tasks are paused or turned off, so search keeps working.

The monthly AI budget

Each plan has a monthly AI budget: US$10 on Personal, US$50 on Research and US$200 on Organization. Enterprise has no AI budget limit. See Billing for the plans.

AI spend counts from the 1st of each month (months start at midnight UTC). Only live AI work for your workspace counts: answers from recordings cost nothing, work on the platform's own AI server costs nothing, and AI work on public sources is the platform's own cost. Owners see AI spend this month against the budget in Admin, under Usage and spend.

When the month's AI spend reaches the budget, AI tasks pause until the 1st of the next month:

• Every member sees a notice: "AI tasks are paused until" the 1st of next month, on the upload page and on documents.
• A task on a new document does not run. The document shows the task as Not run, with the reason: "This month's AI budget is used up. AI tasks start again on" the 1st of next month.
• A draft or another task you ask for is refused with the same reason.
• Tasks that were already running when the budget was reached finish, so the month's spend can end a little above the budget. Nothing is charged beyond the budget.
• Uploads, storage, passages and search keep working.

Tasks that did not run are not run again later. A document added while AI is paused keeps its passages and stays searchable, but AI does not read it.

Turning AI off

An owner can turn AI off for the whole organization: Admin → Usage and spend → AI → Turn AI off.

While AI is off:

• no AI task runs, so no document text is sent to the model;
• every member sees "AI is turned off for this organization" on the upload page and on documents, and tasks on new documents show Not run with that reason;
• drafts and other tasks you ask for are refused with that reason;
• uploads, storage, passages and search keep working.

Turn AI on starts AI tasks again for documents added from then on. Documents added while AI was off are not read later. Each change is recorded in the audit log: who made it, when, and whether AI was turned on or off.

The switch only stops AI. It cannot add a task to the list above or change what a task may do.

Reviewing proposals

Proposals wait in the Review queue (Statecraft → Review queue). Each one shows what the task found beside the quotes it rests on.

• Researchers and editors see the queue and each proposal. By default the queue shows proposals that wait for review or were quarantined, newest first; you can filter it.
• Editors approve or reject. Rejecting needs a reason, which stays on the record. A proposal is reviewed once.
• Owners can do everything an editor can.
• Every review is recorded in the audit log with the reviewer, the change and the reason.

Approving a proposal creates the record it describes, such as a meeting or an opportunity, and links it to the document. Nothing is created from a proposal that was not approved.

Proposals about public sources (shared sources such as government agendas) can be read by every workspace but reviewed only by the platform's operators, because a public record serves every workspace.

Statuses

How an answer is checked

1. Passages. A document is cut into passages: paragraphs, or parts of long paragraphs. Every citation points at a passage you can open, and the same text always gives the same passages.
2. One document at a time. Each task sees only its own document's passages, marked as untrusted.
3. Checks. Before anyone sees an answer, the platform checks it: each citation must be a passage the task was given, each quote must appear word for word, and dates and times must be stated in the passage cited. A meeting's items must belong to the meeting, and a deadline cannot come before the document. An answer that fails a check is quarantined.
4. Review. The answer waits for a person.

Searching passages

Workspace search includes passages: paragraphs of the documents you can see (public ones and your workspace's own, never another workspace's). A passage is found by its words and by its meaning, so a search for "stormwater utility fee" also finds a passage about "the surface water charge". Each result opens the document at the passage.