A year ago, we started building an AI agent to support counterfeit detection on a large online marketplace. It is a difficult space. Listings offer incomplete evidence, fraud tactics evolve as defences improve, and organised sellers actively try to blend in. We set out to explore whether a multimodal agent could help turn scattered signals into structured, explainable assessments.

In this talk, we reflect on that first year: the decisions we made, the assumptions we had to revisit, and what we learned along the way. We’ll touch on combining LLM reasoning with image inspection, domain knowledge, and specialised tools, as well as evaluation, observability, uncertainty, and human oversight. We’ll also share our perspective on where agentic AI has been useful, where it has struggled, and what operating against adaptive adversaries asks of such a system.