Start with the problem.
Design the environment.
Then decide what the system needs.
I am interested in what happens when we stop treating the application, the workflow or even the model as the natural starting point — and begin with the outcome a person is actually trying to achieve.
What if the application is no longer the experience?
A customer may ask for a case-management replacement because employees spend too much time finding missing shipments. But the desired outcome is not a better case. It is knowing where the shipment is, what happened and what should happen next.
Case management may still be an important capability underneath — for ownership, audit, escalation or exceptions. That does not mean the employee should have to operate it. CRM, ERP, tracking, knowledge and specialist systems can increasingly become capabilities inside an experience rather than destinations the user has to navigate.
Does the user need another application — or do they need their problem solved?
Design the environment. Connect the capabilities.
The model is only one part of an AI-native solution. Context, data, tools, permissions, policies, feedback and human expertise determine what the environment can actually do. My role is to understand the business and human context, shape that environment, define boundaries and connect the capabilities that need to work together.
Reliable data, state and evidence should not become probabilistic just because the experience becomes more flexible.
CRM, ERP, APIs, knowledge, automation and specialist systems remain valuable — increasingly as composable capabilities.
Intent and context can shape the interaction instead of forcing every user through the same predefined application journey.
Useful agency needs explicit permissions, boundaries, evidence and escalation paths — not merely access to a model.
Models, data, tools, agents and applications become building blocks. The real opportunity is not choosing the perfect component, but combining capabilities into environments that can evolve as needs change.
People, companies and customers may each bring their own AI environments, capabilities and governance. Architecture should let these environments interact through clear contracts and interoperable interfaces without requiring them to become one system.
Standardize the truth. Govern the capabilities. Personalize the experience.
Keep people where judgment matters.
A human approval step is not automatically human control. If every routine decision requires a click, attention becomes another bottleneck and approval can become reflex. I am more interested in designing systems that remove predictable work while making uncertainty, consequences and unresolved questions visible.
Automate the obvious. Verify the uncertain. Escalate the consequential.
The goal is not to remove people from work. It is to make human attention available for the situations in which experience, responsibility, creativity or judgment actually changes the outcome.
Reliable at the core. Curious at the edge.
Enterprise execution needs determinism where evidence, permissions, transactions and safety matter. Exploration needs room for alternatives, prototypes and occasionally unexpected results. A useful environment needs both — and a way to recognize whether a deviation is a defect or a better direction.
Freedom at the edge. Determinism at the core. Evolution in between.
Prototype to learn. Specialists to industrialize.
AI makes it possible to move from an idea to something tangible much faster. That does not make specialist expertise interchangeable. A working prototype can expose assumptions, make an experience testable and shorten the distance between business reality and engineering — before security, integration, reliability and scale are taken forward with the people who specialize in them.
Architecture is also about what we have not settled yet.
If execution becomes cheap and continuous, judgment may become more important than working time as a design constraint.
A grammatically correct message can still miss intent, culture, history and the relationship between sender and recipient.
AI can lower the barrier to creation without making years of domain experience interchangeable.
Reliability needs controlled execution. Learning also needs the ability to recognize a useful deviation when one appears.