When compliance becomes computable, what remains a human decision?

From July 8 to 10, I attended the first Automated Design Compliance Checking Global Summit at TUM in Munich, organized by Ghang Lee with André Borrmann and Robert Amor. Researchers and practitioners presented the past, present, and possible future of automated compliance checking: what computers can check, what they still cannot, and what it will take to move these systems into practice. I was honored to present our prototype for automated drawing review.

One question stayed with me. Computers may eventually outperform people at tracing explicit requirements across drawings, models, and specifications. But detecting noncompliance is not the same as understanding the risk behind it. Some failures create life-safety risks. Others create financial, schedule, quality, or administrative risks. They should not all be treated as if they have the same consequence.

Many compliance systems still express requirements as fixed instructions: something must be done in one particular way. But real projects constantly involve trade-offs. If one issue could delay an entire project, it may make sense to evaluate its probability, consequence, and cost of correction—without turning essential life-safety protections into negotiable items.

This creates a difficult question: how should we price risk? If risk is never measured, decisions can become dominated by bureaucracy. But if everything is given a price, we may start treating safety and public responsibility as ordinary commercial trade-offs.

AI makes this question more urgent. Once machines can apply rules at scale, we can no longer treat every rule as unquestionable doctrine. We need to be clearer about why each requirement exists, what risk it controls, and which decisions must remain human.

That’s this week’s brief. If something here helped or something is missing, just hit reply. I read everything.

See you next week, Shicong