AI can design.
It can suggest boundaries, interfaces, tests and architecture options. It can critique coupling, propose alternatives and generate implementation paths faster than most teams can review them.
So the question is no longer whether AI can help with design.
The question is who has authority to approve it.
This is why Craig Larman’s Applying UML and Patterns still has value. Not because teams need to return to heavy UML, but because responsibility-driven design still asks the right question: who owns this decision, and under what constraints?
For early experiments, AI should have room to move fast. Generate options. Compare approaches. Learn quickly. Throw away what does not work.
But once something touches users, data, workflows, risk or long-term product direction, the rules need to tighten.
Approval is the new design question
- who approves the responsibility boundary;
- which rule sits in which layer;
- what AI is allowed to change;
- what evidence is needed before the design becomes permanent;
- where human judgement remains non-negotiable;
- what must remain easy to test, replace or govern.
The unit of responsibility has changed
In modern AI systems, the responsibility might not be a class. It could be a service, workflow, policy, agent, tool or human approval point. But the question is the same: who owns the decision, and under what constraints?
How this maps to Drexing delivery
Early ideas can move fast. Real clinical workflows need clear rules, evidence, oversight and human approval.
AI may one day redesign parts of the system itself, as long as the rules still hold. We are not there yet for clinical workflows, but it shows why governance matters more than any single architecture diagram.
The future is not human design versus AI-generated code.
It is dynamically allocated design authority: deciding which decisions AI may make, which constraints it must respect, what evidence is required and which decisions remain human-authorised.