Giving an agent a knowledge graph of the whole firm cut its time to answer five firm-wide questions by 73%, its input tokens by 92% and its model calls by 66%, with no loss of accuracy.