How can customs agencies inspect vast volumes of cargo without slowing legitimate trade or handing consequential decisions entirely to an algorithm?
In this episode of AI at Work, I speak with David Smason, founder of CargoSeer and Head of Customs AI at BigBear.ai, about a form of workplace AI that operates far from the usual office productivity conversation. CargoSeer was built to help customs teams combine cargo documents, x-ray images, routes, commodity codes, tariffs, and historical patterns before an experienced officer makes the final decision.
David explains why traditional customs work can overwhelm even highly trained people. A single shipment may involve dozens of documents, different languages, handwritten information, revenue questions, and national security risks. His vision is one operator overseeing hundreds or eventually thousands of analyses while concentrating attention where human experience matters. He compares the result to a James Bond command center, although the real work involves careful data processing, explainable recommendations, and a clear role for the officer.
We discuss CargoSeer’s internal findings that an operator using its supporting technology can assess a shipment in under five seconds, compared with several minutes in some existing processes. That speed comes with an important qualification. False positives may delay legitimate cargo, while false negatives may allow serious risks through. David says each customs agency must define its own risk appetite, operating procedures, success measures, and thresholds rather than accept a universal model.
Human feedback remains central to the system. David compares an AI model to a baby that must be taught continuously. When an officer corrects a false alert, the correction can improve later recommendations. He argues that operators value this influence because their knowledge becomes part of the technology instead of being discarded.
We also discuss localization and proof. David estimates that CargoSeer’s models are currently about 85 percent standardized, with the remaining work shaped by local data, language, culture, tariffs, agreements, and operating practices. His advice to public agencies is direct: demand evidence that the system works on your data, in your environment, and within your workflow. Test it, try to break it, and define performance before live deployment.
Could this model of AI-supported judgment work in other high-consequence roles where speed, evidence, and accountability must coexist? Listen to the conversation and share your thoughts with me.
