Checking drawings with AI is not one AI doing everything. It is a team
🧩 Checking drawings with AI is not one AI doing everything. It is a team: one lead that dispatches, five expert checkers that each check one thing.
This week's office AI tip is about AI design, the agent architecture itself:
・The lead agent never reads a drawing. It dispatches and merges
・Five expert checkers: GFA; plot ratio, site coverage and height; fire safety and means of escape; barrier-free access; drawing standard
・Each gets one written checklist with clauses copied word for word from the source (B(P)R, the Fire Safety Code 2011, 2024 edition, the Barrier Free Access Design Manual 2008, 2025 edition, APP-151 and others), verified by code
・Every row names the sheet, the clause and the exact words on the drawing
・Then no more AI: code finds every quote on the sheets and recomputes every number
・The Authorized Person reviews every row and decides
I suggest Hong Kong practices keep one anyway, and finish it before submitting to BD.
Real run, using Claude Code: a fictional 16-sheet GBP set with 6 errors planted on purpose.
・5 experts in parallel: 4 min 45 s, 67 rows, all 6 errors caught
・Fair comparison, same model and same checklists, 1 agent doing everything: 8 min 24 s, also all 6
・So splitting did not catch more this time. It was faster, at roughly 60% more token cost, and the other issues raised were a mix of real and false (real 6 vs 5, false alarms 10 vs 7)
・Code recomputed 36 sums; 4 rows did not match: 3 were real errors on the sheet, 1 was the checker misreading a table
A compliance register is UK practice; Hong Kong law does not require one. I suggest Hong Kong practices keep one anyway, and finish it before submitting to BD. AI does the checking. It does not sign for you.
If you were splitting the work, what would you let AI check? Tell me what you think in the comments 👇
