Let an AI agent draft a document you issue

Before you begin

Understand where the boundary is, because it's the reason this is safe to turn on. An AI agent in the platform is a service principal with its own access profile, and service principals are refused every state transition on every document type — issue, send, approve, and void alike. That refusal lives in the state machine that governs the document, not in the agent's instructions. A prompt that says "never issue an RFI" is a request. A transition the service principal cannot perform is a fact, and it holds for a document type added next year that nobody remembered to write a prompt rule for.

Steps

1. Give the agent its access the same way you'd give a person theirs. Go to Admin & Settings → Governance → Permissions → Project Roles and use the AI agent template, narrowing the domains it needs. If the agent should pick up new jobs automatically, tick Add to new projects on its row.
2. For an RFI, open the RFI and click Draft answer with AI. The draft lands in the answer field as editable text.
3. Read it against the contract documents. Edit it, or delete it and write your own — it's a starting point, not a response.
4. Issue the answer yourself. The RFI moves state because you moved it.
5. For submittals, open the AI submittal log panel and let the agent draft the register from the specification. The panel says plainly that nothing is created until you apply it.
6. Review the drafted rows, tick the ones that are right, and click Apply. The applied rows enter the submittal register as drafts, exactly as though a person had typed them.
7. Work those drafts through your normal submittal process. Nothing skips a step because an agent proposed it.

What happens next

The division of labour settles into something predictable: the agent does the transcription and the first pass, and a person does every act that has a consequence outside the system. An agent can save an hour assembling a submittal register from a spec section. It cannot send that register to the architect, and no configuration mistake makes it able to.

Applied drafts are ordinary drafts. They carry source metadata recording that an agent proposed them, so you can tell later, but they're subject to the same approvals and the same audit trail as anything else. There is no separate agent-approval queue to forget to check.

Where the record lives

The document keeps its own history: what the agent drafted, what the person changed, and who issued it. Because the agent's access is a project role like anyone else's, you can inspect what an agent could reach on a given job through the same Explain view you'd use for a person — which is what makes the answer to "what did the AI have access to" a lookup rather than an argument.

Troubleshooting

The agent won't create the document type you expected: agent drafting is wired for RFI answers and the submittal log today. Other document types are governed by the same draft-only rule, but there's no agent entry point for them yet — you'll be writing those by hand.

The agent's draft is confidently wrong: that's the expected failure mode, and it's why the review step is not optional. Treat a draft as a proposal from someone who has read the documents quickly and has no liability.

You want the agent to send it once it's obviously right: it can't, and there's no setting that changes that. If issuing without review is genuinely what you want, that's a person clicking issue after skimming, and at least the record says who.