AI in building design: match automation to the project stage

Start with a deliverable the team needs

A project team can use AI to organize requirements, compare options, interrogate supported drawing data and prepare issue registers. Each operation needs its own inputs and acceptance check. “Design a building” combines too many decisions to evaluate as one task.

Name the next deliverable: a comparison of three massing options, a reconciled room schedule, or a list of unresolved references. Record the current stage and revision. This prevents a concept image or preliminary estimate from acquiring the authority of an issued drawing.

The relevant question is how much checked work the operation saves. Measure preparation, execution, review and correction separately. A quick first answer can still create a slow review.

Compare options under the same assumptions

An option study needs fixed inputs: site boundary, retained features, orientation, programme and constraints. Define the metrics and their calculation method before generating alternatives. A high area total may reflect a changed measurement rule rather than a better layout.

Autodesk's Generative Design guidance describes studies driven by Dynamo logic, inputs, goals and constraints. This is a useful example of a structured option workflow. It does not mean a generated option has resolved every architectural or engineering requirement.

Ask for a feasible baseline, rejected alternatives and reasons for rejection. Keep a proposed relaxation visible: reducing a service zone to improve an area score changes the brief. The team should approve that change before the study continues.

Environmental results also need their method, model extent, weather assumptions and level of detail recorded. Treat an early screening result according to its documented purpose. The fact that a tool produces a numerical score does not establish suitability for a submission or detailed design decision.

Use schedules to test information quality

Schedules make a practical first trial because missing IDs, duplicate rows and totals are easy to inspect. Supply a current export and explain whether it includes linked models, alternatives, demolished objects and placeholders.

A language assistant can help classify exceptions and explain differences. Use reproducible calculations for totals, and keep the calculation method available. For geometry, require an object-reading tool, checked script or configured CAD connector. A visual description of a plan is a different output.

For a Revit project, prepare the relevant DWG views, PDF sheets and schedule export. Check units, visibility and revisions across the package. This provides a supported route for a drawing task in ARAI without claiming native RVT inspection. See AI for Revit for the distinction between model-side tools and exported evidence.

Worked example: compare a room schedule with a plan

Suppose a concept package contains 24 room identifiers, while its schedule has 23 unique rows. The immediate job is to explain the mismatch. It is too early to infer a missing room function or adjust a floor-area target.

Compare plan A-101 revision B with the supplied room schedule from the same revision. Preserve room IDs exactly. Return matched IDs, missing rows, duplicates and name conflicts with source locations. Keep unreadable tags separate. If you compare areas, use only the declared measurement rule and identify each source value. Do not add rooms or revise the design.

The reviewer checks every exception and confirms that the matched and unresolved counts account for all 24 identifiers. If area values disagree, first establish whether the boundaries and conventions match. A schedule measured to internal finishes and one based on another boundary can differ without an arithmetic error.

After the author corrects the source, verify the reissued pair. Updating only the spreadsheet leaves the drawing coordination unresolved.

Keep requirements and measurements connected

A technical finding needs both project evidence and an applicable requirement. Keep source title, edition, clause, relevant conditions and reviewer status alongside the measurement. A code citation with no accessible passage is an unresolved research item.

OpenAI's accuracy guidance recommends checking technical facts and references. Search and document tools help find evidence; the project team still establishes applicability. Code-research checks give a practical record format.

Trial one operation in ARAI

Choose a drawing task such as an object inventory, specified distance or bounded area check. Set the input revision, units, scope and expected evidence before sending it. Compare the result with the applicable project requirements from separately verified sources.

Prepare a building-design task. Review the draft and attach the current files. Accept the returned material only after checking the selected objects and opening any output file.