When I add an AI step to a building workflow, I treat it like an untrusted dependency. It can produce useful candidates quickly, but every output needs a source, a scope, and an owner. That approach fits the current US policy environment: federal agencies are publishing governance and procurement guidance, while permitting and professional responsibility remain controlled by state and local systems. The goal of this pipeline is not to ban generated images. It is to prevent a concept render from quietly becoming an engineering fact. 1. Declare the Output Type I start with a small metadata block in the project log: output_type: concept | effect | scheme intended_use: discussion | option_screening | technical_reference review_owner: named person A concept image supports a conversation about massing. An effect image communicates a possible experience. A scheme sketch tests a more specific design idea. The labels keep a detailed-looking image from being mistaken for a coordinated permit sheet. 2. Capture Site Context Before Generating Form I list the data used for the early study: parcel source, terrain date, building footprints, street geometry, zoning notes, weather file, and any client reference material. I note what is missing. A smooth AI scene built on a wrong grade is still wrong. Shapezo can provide a fast context layer. I select a boundary on a map, and its AI generates an initial 3D model of that area. I use the result to inspect adjacent massing, street connections, open space, and broad relationships between a proposed building and its setting. I do not promote it to survey control, BIM, or permit geometry. 3. Separate Generative Assistance From Professional Checks AI is reasonable for early massing alternatives, facade studies, daylight questions, site summaries, and visual issue lists. I keep licensed review for dimensions, occupancy, egress, fire compartments, structural load paths, lateral systems, mechanical clearances, energy code inputs, stormwater, and utility coordination. The handoff is explicit. The architect owns the design intent and code coordination. The structural engineer owns the structural assumptions. The mechanical and electrical engineers own their systems. A code consultant or accessibility specialist can resolve local amendments and inclusive-design questions. The model creates candidates; the named professional accepts or rejects them. 4. Test for Common Failure Modes I run a predictable failure checklist: Geometry Are stairs continuous? Do floor-to-floor heights fit the program? Are doors, ramps, corridors, and turning spaces plausible at the claimed scale? Structure Is there a believable load path from roof to foundation? Are columns, shear walls, transfer beams, and lateral bracing present where they need to be? Life Safety Can occupants reach a compliant exit from every occupied area? Are travel distances, fire separations, rated openings, smoke control, and emergency access still questions for a qualified reviewer? Fairness and Access Does the scheme provide usable accessible routes, adaptable units, clear alarms, and equitable amenity access? Does the image quietly assume one kind of body or household? 5. Keep a Copyright and Provenance Log For each retained output, I record model name, version if available, prompt, reference images, permissions, date, editor, and final destination. I do not put confidential drawings or personal data into a public service without authorization. The US Copyright Office is still examining training and output questions, so a provenance log is more useful than a confident claim about ownership. 6. Add Disclosure at the Handoff My note says what AI did, what it did not do, who reviewed the result, and which values are provisional. The note travels with the image into a client deck, internal issue set, or permit narrative when relevant. It is short enough that people will actually read it. I also keep the disclosure close to the file, not only in a meeting slide. A later reviewer should be able to open the image, see its source and status, and understand whether it was used for discussion or for a measured decision. That small bit of metadata prevents an old concept from being copied into a new package without its limitations. 7. Stop Before the Signature Before an architect or engineer signs, I remove unverified generated geometry from the authoritative set or replace it with checked information. A building official should be able to tell which drawings are controlled documents and which are exploratory references. No image generator can take the professional oath, answer a correction notice, or carry a license. The pipeline is intentionally ordinary: classify, document, generate, test, review, disclose, and sign. It lets AI save time without pretending that speed changes the legal boundary around a US building permit.