ChatGPT for architects: Revit exports and automation

Start with one answerable BIM question

A useful first ChatGPT task is “Find duplicate room numbers on Level 02 in the proposed phase,” rather than “Check my BIM model.” The first has a defined population, observable exceptions and a result an architect can reopen in Revit. It also works without giving a chat service access to the live model.

Separate three workflows: analysis of an exported schedule, code that you review and run inside Revit, and a configured connection exposing model tools. They require different inputs. An RVT filename in a conversation is not proof that the assistant opened the model, and a plan image contains neither the full parameter set nor every phase and design option.

Export enough information to locate each finding

Duplicate the schedule for the review so its filters and fields can be inspected without disturbing an issued view. Record the Revit release, model revision, selected levels, phase, design option and whether linked models are included. Turn off grouping that would conceal individual instances when the task requires one row per room.

Autodesk documents schedule export to delimited text. Keep headers and a declared delimiter; confirm the resulting rows match the review schedule. Add room number, name, level, phase and area. Include a stable element identifier through an appropriate export or read-only routine if the standard schedule cannot supply it. Room numbers alone may repeat across buildings or phases.

Provide an approved naming or finish-code table separately. Its revision matters just as much as the model revision. A changed office standard can produce apparent errors in otherwise consistent model data.

OpenAI’s PDF visual retrieval FAQ describes an Enterprise-only feature for PDFs attached in conversations; GPT Knowledge and Project Files use text-only retrieval. If visual retrieval is unavailable, attach separate readable image crops of the plan.

Prompt: audit a room schedule without inventing rules

Audit rooms_L02_revD.csv against naming_rules_v3.csv. These are exports from Revit 2026, proposed phase, primary design option, host model only. Area is in square metres. Check duplicate composite keys of building, level, phase and room number; blank names; and names absent from the supplied approved list. Preserve original values. Return the element identifier where supplied, source row, failed rule and proposed action. Do not normalize spelling or fill blanks automatically. Report missing columns and ambiguous matches before continuing.

This produces a worklist, not a revised model. Verify the imported row count and inspect several exceptions in Revit. Check a room that should pass as well as one that should fail: a filter error can otherwise make every reported exception look convincing.

Handle units and rounding explicitly

For an area comparison, state the schedule’s displayed precision and the comparison tolerance agreed for that job. Do not interpret an integer display as an exact underlying value. Preserve decimal separators during table import, especially when the delimiter is also a comma. Distinguish a blank area from zero and an unplaced room from a placed room with a valid boundary.

Ask for a read-only script before a write script

ChatGPT can help draft code for a chosen Revit automation environment. Supply the exact Revit API release, Dynamo or pyRevit environment, interpreter and parameter identifiers. Request a read-only report first. The code should state what it selected, what it excluded and which elements could not be read.

Before allowing writes, define whether each parameter belongs to the instance or type. A type change can affect many placements. Require a proposed old/new value table, explicit handling of read-only parameters and a transaction strategy that reports failure. Do not accept “ignore errors and continue” as a completion policy.

Test reviewed code on a disposable model copy using your BIM team’s worksharing procedure. Compare the selected identifiers before the write, then reopen the saved model and read the values back. Keep an output log with the model revision, script version and skipped elements. A screenshot of a completion dialog does not show that the intended elements changed.

Connected Revit access is a separate setup

OpenAI supports configured MCP tools, subject to account and workspace controls. A Revit integration must provide compatible transport, authentication and concrete model operations. Do not assume that installing an arbitrary local Revit server makes it available inside ChatGPT.

Autodesk’s current Revit MCP setup specifies Revit 2027.2, a separate Technical Preview add-on and local STDIO clients such as Claude Desktop and Cursor. It is not a plug-and-play ChatGPT connection: ChatGPT requires a supported remote transport and account permissions, so this local server needs a vetted remote adapter or bridge. For an ordinary review, use drawing and schedule exports or a supported local CAD client.

Check an exported Revit plan in ARAI

Export the relevant Revit plan to DWG and a matching PDF. Record the view, revision, export units and coordinate basis; open the DWG in CAD and check a known dimension. Add the room schedule, selected room scope, length and area units, and agreed boundary rules for measuring each room. Ask ARAI to compare the schedule with areas from those drawing contours. Request object identifiers and sheet locations so you can review each finding in CAD.

Check room areas in ARAI. Attach your files to the draft before sending. Check the reported discrepancies against the drawing and schedule before making changes. This workflow uses drawing exports; native RVT editing is not included. See the DWG preparation guide.