A team is building a semantic search system that must return the 10 most similar documents to a query from a corpus of 50 million embeddings (1536 dimensions). They implement exact nearest neighbor (ENN) search using brute-force cosine similarity. Latency is 8 seconds per query. Their manager demands sub-100ms latency. A colleague proposes switching to an Approximate Nearest Neighbor (ANN) algorithm. The developer objects: "ANN gives wrong answers — we need exact results." How should the team evaluate this trade-off?