I did not begin my career as a product builder. I learned by selling, opening markets, working with customers and building the systems a commercial team needed to operate. That experience changed how I think about agentic workflows.
A workflow is not the diagram created after the work has been cleaned up. It is the real sequence of decisions, exceptions, missing information and judgement calls that people manage every day. Commercial leaders often know this version well because they see where the official process stops matching the customer.
1. Start with the exceptions
The standard path is usually easy to document. The exceptions reveal the real product. They show where information is incomplete, where customer context changes the answer and where an experienced person quietly prevents a mistake. Before automating a workflow, I want to know what breaks it and how the best operator responds when it does.
This does not mean designing for every rare event on day one. It means making the boundary visible. The agent should know when it can continue, when it needs more information and when the work should move to a person.
2. Context is often more valuable than automation
A fast answer is not useful when it ignores the customer, the market or the decision behind the request. At Twimbit X the important shift was not simply generating information. It was shaping intelligence around a particular account and commercial situation. The workflow became valuable when it carried context forward.
Good agent design therefore needs a way to collect, preserve and apply context. Automation without context creates more output. Context turns that output into work someone can use.
3. The person closest to the customer should shape what comes next
Product priorities are often decided far from the moment of use. Commercial and delivery teams can close that distance. They hear the repeated questions, see the workarounds and know which delays matter to the customer. That knowledge is a product input, not just feedback.
The goal is not to let every request become a feature. It is to find the repeated unit of work underneath the requests. When the same judgement is being recreated again and again, there may be a workflow worth building.
4. Human review is a design choice
People sometimes describe human review as something that will disappear when the model improves. I think that is too simple. Some review exists because the technology is still learning. Other review exists because the work carries consequences, ambiguity or taste.
For Curious East the question is not whether a person is present. The question is where human judgement improves the final experience. Agents can support itinerary building and execution. People can review the plan, handle exceptions and stay accountable for what reaches the traveller.
5. Commercial leaders can become product builders
Commercial leaders understand value, urgency and the messy reality of adoption. Agents are making it easier to translate that knowledge into working systems. The opportunity is not to become a traditional engineer overnight. It is to learn enough to define the workflow, test assumptions and build closer to the customer problem.
That is the larger lesson for me. The tools have changed, but the work still begins in the same place. Find the problem. Understand the customer. Build the system. Then watch what happens in reality and improve it again.