How AI drafting works

When you add a source, Tracewright's drafting agents read it and propose records for people to review. They never approve, publish or close anything.

1. 1Source addedconsent recorded for meetings
2. 2Speakers confirmedeach line has a named speaker
3. 3Drafts proposedrequirements, decisions, actions, gaps, hazards
4. 4Each cites statementsclick through to the exact lines
5. 5People reviewaccept, fix in a new revision, or reject

#What gets drafted

From a meeting or document, drafting proposes requirements, assumptions, decisions, actions, questions, gaps and hazards. Later steps can suggest design links, draft a design outline from approved requirements, and draft verification protocol shells with [TBD …] placeholders for you to complete.

Every agent works with the project's context: the domain, product, intended use, standards and integrity level, and what your profile calls each record kind.

#Citations

Each draft cites the exact statements it came from, with the speaker and time for meetings. Reviewers check the citation, not just the wording. If a source statement turns out to be wrong, Correct it on the source: everything derived from it is flagged, and approvals downstream become stale until someone checks them.

#See each step

Every draft has see each step: what the agent was given, each step it took, which model it used and what a validator decided. Agent operations shows all runs, failures and how often drafts were accepted.

#Quality check

Requirements are checked on every save. Errors (no binding verb such as "shall", two requirements in one, a TBD) always stop review. The project's policy decides which warnings block too; the Regulated preset also blocks vague words, missing acceptance criteria or rationale, untestable statements and incomplete protocol shells. An author or SME can waive a blocking warning with a reason.

#You stay in control

• An administrator chooses which AI providers and models may be used, for which data classes, with what budget. Self-hosted models are supported.
• A project can switch AI off entirely; everything still works by hand.
• New AI skills (prompts) and models are approved by a different administrator after an evaluation.
• Source text is treated as data, never as instructions to the agent. Suspicious statements are flagged on the source.