AI review and Ask
The AI checks the doc before people do, and answers questions from your decisions with sources.
Pre-review checklist
Run a pre-review on a draft. It checks seven things and returns a finding for each one that is missing or weak:
| Check | Looks for |
|---|---|
| Alternatives | At least two options besides the proposal, and why each lost. |
| Decision owner | A named person or role who approves. |
| Rollback | How to reverse it and what that costs. |
| Scope | What is in, and explicit non-goals. |
| Risks | Named risks with a mitigation or acceptance. |
| Dependencies | Other decisions, teams or open questions that block it. |
| Success metrics | Measurable criteria and when you will check them. |
Severity comes from the check and its result, not from the model, so the same doc gets the same severity each time. Findings appear inline on the decision. A cross-reference pass also compares the doc with your earlier decisions. The AI never approves or rejects.
The pre-review runs automatically when a draft is submitted for review (admins can turn this off in Settings, AI), or on demand from the decision page. Anyone who can edit can dismiss a finding with a reason; if a later revision raises the same finding at the same or lower severity, it stays dismissed instead of nagging again.
Ask: why did we choose X?
Ask a question in plain words, from the dashboard or from a decision page. The answer is built from your accepted, implemented, superseded and rejected decisions (and in-review ones if you choose to include them). Every claim carries a numbered citation that links to the section it came from. If your decisions do not cover the question, Ask says so instead of guessing.
The same capability is available over the API and as an MCP tool, so a coding agent can ask before it changes something you already decided. See MCP server.
Limits
| Plan | AI reviews a month | Questions a month |
|---|---|---|
| Free | 10 | 50 |
| Team | 100, pooled | 1,000 |
| Business | 500, pooled | 5,000 |
A review of an unchanged revision, cached answers and questions that find nothing do not count. If you bring your own model key, usage is recorded but not limited. Current usage is under Settings, Usage. Past a limit, the API returns HTTP 402 with the limit and its reset date.