A passed check is not a checked drawing
Most architects will recognise this: an issued PDF opened after the fact, and a problem no checker flagged. A tag sitting on top of a dimension. A hatch that reads as the wrong material. The right view on the wrong sheet. The register said "issued", the title block fields were complete, and the sheet was still wrong.
I see the same pattern in the AI agents I build for my own projects. Before an agent hands over a drawing, a slide or a video, it runs automated checks: does the text fit, are all the files there, are any banned words present. These checks are fast and useful. Yet in one case a QA pass reported that everything was fine, and the layout was visibly broken the moment the file was opened at full size.
I see the same pattern in the AI agents I build for my own projects.
The lesson is a simple split:
1. Automate what can be counted: title block fields, sheet counts, revision sequence, file names.
2. Make the agent open its own output and look at it. Current models can read images, so this is a step it can do, not a step to skip.
3. Keep the final look, and the responsibility, with the professional.
As AI produces more drawings and documents, the question is not whether the checks passed. It is whether anyone looked.
Which checks would you automate first, and which would you never hand over?
