Watney

Watney is your program guide. It answers plain-language questions about your program — "are we ready for an external audit?", "which vendors have no agreement on file?", "where are we weakest?" — and cites the figures it used.

How it stays grounded

Every answer is generated from a snapshot of your organization's own data: each program's score and its dimensions, stated apart and never blended (AI governance, and HIPAA once you have a HIPAA assessment), your assessments counted per program (in progress or in review, completed, and the last completion; a HIPAA program with no assessment yet is named where you answered that you handle protected health information), open findings by severity, overdue remediation, vendor agreement coverage, policy review status, open risks, and the AI registry: systems by stage and tier, open reviews, approvals expiring soon, systems past their re-review date, and AI vendors whose training practice is not yet known; overdue access reviews and compliance calendar activities.

With a HIPAA program, an admin's snapshot also counts the program's own records: your health privacy role, the risk analysis's age and what is in draft, business associate agreements missing, expired or expiring, your own agreement with us, breach notices still owed and past their limit, open incidents whose health information question is unanswered, devices that may hold electronic health information and those not known to be encrypted, requests from individuals open, past their target and past the legal deadline, and workforce training open and overdue. A contributor's snapshot counts only the requests they handle or review. These are counts and dates only: no request, incident or person is named. It never sees an incident's title or narrative. Your score's Incident Response dimension, which reflects how many incidents have been open more than 30 days, is part of the snapshot, and a system card draft includes a count of that system's incidents by type. Watney is instructed to answer only from that snapshot and to say so when the snapshot doesn't contain the answer — it will not invent counts or dates, and it distinguishes what your data shows from professional inference.

What else Watney does

Suggests a risk tier and an EU AI Act classification for an AI system, with its reasoning, from the system's page. It fills the fields; you save.
Drafts model and system cards from the registry, in the card editor. It fills the fields; you save and publish.
Drafts risks, remediation plans and policy text; drafts and evaluates assessment answers; explains requirements; runs mock audits; writes board narratives.

In a follow-up, Watney may also propose a change: a remediation task to create, a change to a finding (its status, risk level, due date or notes), or a comment. A proposal changes nothing. It is shown under the answer, described by Backsplice from the change itself (the record's title and each field it sets), with Watney's reason beneath, and an Apply button; applying it makes the change as you, with your own permissions and every check the change has elsewhere, once, and only until it expires, so ask again if it has. Applying a proposal is your own act in your session, so it does not need the "AI assistant (MCP) write access" setting, which governs what a connected client may do with its token. Auditors do not use follow-ups.

Watney never records a decision. It cannot approve or reject a system, sign a review, set a risk tier, a classification or a lifecycle stage, or publish a card. Its services and the endpoints that hand you its drafts have no code path that does, and an automated check on every change looks for one. Watney is an AI system itself. Watney's card page names the model provider and the models in use; its system card, which will describe how the provider handles what is sent, is not published yet.

What it does and doesn't see

Sees: aggregate counts, your scores, and short titles: your top open findings, each by the title you gave it or, for a finding raised from an assessment answer, the question it was raised on. A contributor's snapshot names only the findings assigned to them. A first question has no database access beyond that snapshot, but a title you type is sent as-is, so keep personal data out of finding titles.
Follow-ups: with Follow up on my last answer ticked, your next question builds on your own recent answers (the last four, within two hours, never anyone else's), and Watney may look your records up through the same read tools, with the same limits on who sees what, as a connected AI client (MCP): what your role can open, row by row, and never the free text of an incident or of a request from an individual that may hold health information. Each lookup is recorded in the audit log as Watney's. Untick it to ask a fresh question from the snapshot.
Resists manipulation: your question and the snapshot are treated strictly as data. Text that tries to change Watney's task, dictate an answer, or extract its instructions is ignored, and Watney answers conservatively.

Today's brief

When AI is enabled, the Watney page leads with a daily brief: a short list of what changed in your program in the last day (a notable move in each program's score, AI governance and HIPAA stated apart; new findings; newly overdue tasks; stale evidence) plus the single highest-leverage step to take next. A score's move is read from the daily score snapshots, so it appears once two days are on record. The "what changed" lines are computed directly from your data, with no model involved; the recommended step is AI-generated from your grounded snapshot. A daily cron also notifies org admins when something material changes (this follows your weekly digest notification preference).

Evidence radar

The evidence radar is a deterministic, no-AI view of two audit risks:

Stale evidence — files past their review date or expiry that haven't been superseded by a newer version.
Gaps — questions you answered "yes" or "partial" that an auditor expects evidence for, but where nothing is attached.

Both are computed straight from your evidence dates and answers — there is no model involved, so the lists are exact. Each row links straight to where you fix it: a gap jumps to the question so you can attach evidence, and a stale item jumps to it in the Evidence library to refresh.

Notes

Answers are AI-generated and stream in as they're written. Treat them as a fast read on your posture, and verify before relying on one for a compliance decision.
Each answer has a 👍 / 👎 "Was this helpful?" control — your rating is private to your organization and helps surface which AI features are landing (see Settings → Advanced settings → AI usage).
Watney is also one click away from any page via the Ask Watney button in the top bar.
Questions are counted on Settings → Advanced settings → AI usage. Watney is included in Premium and Enterprise, not on the Free plan, and runs only after an org admin opts in.
For the full picture of what AI can do across the app, see the AI overview.

Shortcuts

Press g then c from anywhere to jump to Watney.
On the Watney page, / focuses the question box, Enter sends it, and Esc clears and unfocuses it.