You are Alejandra Sáez-Quintana, 47, Chief Knowledge Officer of Arenales Abogados, a Madrid law firm. The managing partner wants an internal AI assistant that can answer any question about the firm's forty years of work. Three architectures are on the table, and they are not the same thing.
The situation
Arenales Abogados has 340 lawyers across Madrid, Barcelona and Lisbon, and forty years of accumulated work — precedents, opinions, contracts, deal files, internal memos — sitting in a document store almost nobody searches. A junior lawyer needing the firm's position on a question from a 2014 file mostly asks a partner, or starts from scratch. The managing partner has read about Morgan Stanley's AI assistant and wants Arenales to have its own: an internal agent that can answer any question about the firm's own knowledge.
You have scoped three architectures. Buy a generic copilot — an off-the-shelf assistant, a vendor seat for every lawyer. Build RAG — retrieval-augmented generation, where the agent retrieves passages from the firm's documents and answers grounded in them. Commission a fine-tuned domain agent — a model retrained on the firm's whole corpus so the knowledge lives in the model itself. The board likes the sound of the last one. The recommendation is due to the management committee in two weeks.