Several application domains require data to be enriched prior to its use. Data enrichment is often performed using expensive machine learning models to interpret low-level data (e.g., models for face detection) into semantically meaningful observation. Collecting and enriching data offline before loading it to a database is infeasible if one desires online analysis on data as it arrives. Enriching data on the fly at insertion could result in redundant work (if applications require only a fraction of the data to be enriched) and could result in a bottleneck (if enrichment functions are expensive). Any scalable solution requires enrichment during query processing. This paper explores two different architectures for integrating enrichment into query processing - a loosely coupled approach wherein enrichment is performed outside of the DBMS and a tightly coupled approach wherein it is performed within the DBMS. The paper addresses the challenges of increased query latency due to query time enrichment.