AI can draft your RFI. Start by testing the question.
For project managers and project engineers: test revision accuracy, review effort and answer quality before making AI part of your RFI routine.
The short version
Autodesk and Procore offer AI-assisted RFI drafting. A small, measured pilot can show whether it improves the questions your project actually needs answered.
10 RFIs
Suggested starting sample for a practical pilot, not a statistical benchmark
2 weeks
Suggested trial period before reviewing the workflow
Save briefing
Checking your account save status...
The decision
If your project already uses a platform with AI-assisted RFI drafting, test it on one work package before buying another writing tool. Give the pilot a specific job: help the project engineer turn a documented uncertainty into a clear, answerable question.
This is a workflow guide based on published product documentation reviewed on 10 September 2026. We have not tested these products on a live project.
What the products actually offer
In an announcement updated on 3 June 2026, Autodesk said its Assistant for construction workflows in Forma had left beta. The article also describes AI-generated RFI drafts that users review and refine. Autodesk says access depends on licence type and user permissions. That supports investigating an existing entitlement; it does not establish your project's access or likely savings. Autodesk announcement.
Procore's current RFI tutorial labels its drafting feature open beta. It describes suggested subject, question, cost-impact and schedule-impact fields, plus a search for related information. It requires appropriate permissions and enabled Assist functionality, and tells users to review the generated content for accuracy. Procore RFI tutorial.
Neither document establishes a measured improvement for an Australian project like yours. Our recommendation is to evaluate the complete review process, including the effort needed to correct the draft.
Start with a question you can check
Choose ten closed RFIs from one package, with permission to use their supporting records. Treat ten as a manageable starting sample, not proof of reliability. Include a straightforward clarification, a drawing/specification conflict and an item where a later revision changed the answer.
For each case, assemble the information available when the question arose. Keep the eventual response out of the drafting inputs. Otherwise, the test may reward an answer the engineer could not have known at the time.
Record the drawing number and revision, specification section, relevant photograph or marked-up extract, location, and the decision needed. Let the reviewer identify missing evidence before trying to improve the wording.
For example, a hypothetical services RFI might concern two documents showing different sleeve locations. The useful output identifies both references and asks which location governs. An unsupported statement that the discrepancy will delay completion adds confidence without evidence.
Give the draft a narrow brief
Ask the assistant to state the uncertainty, cite the supplied references and propose a concise question. Require it to list missing information separately. Tell it to leave cost and programme consequences unconfirmed unless the records establish them.
Have the project engineer check every reference against the actual source. Then check whether the question contains one decision, whether that decision is already recorded elsewhere, and whether the right person can answer it.
Keep the usual approval and issue process during the pilot. A technically fluent draft should not become an instruction to the subcontractor merely because it reads well.
Before uploading records, confirm that the selected feature and account meet your organisation's data requirements. Autodesk publishes feature transparency cards covering functionality, data sources and safeguards; these are useful procurement inputs to review alongside your own agreement. Autodesk Trusted AI.
Measure the work that remains
For the closed cases, count unsupported statements, wrong revisions, missing references and questions needing substantial rewriting. Measure drafting and checking time together. A fast draft that requires a lengthy investigation has not improved the process.
If that first test is satisfactory, run a supervised two-week trial on one active package. Record whether responses answer the question without another clarification round. Track time spent preparing and reviewing each RFI separately from time waiting for the designer; they are different constraints.
Agree the expansion decision in advance. Continue when the combined preparation and checking effort improves, source accuracy holds and the receiving team finds the questions useful. Pause when review becomes harder or an unsupported impact statement survives the checks.
Hardhat take
The strongest early use case is a better first draft inside a controlled project process. Start with the software and records you already have, and make the engineer's verification time visible. That gives you a purchasing decision grounded in your project rather than a polished demonstration.