The Challenge
A legal technology company with over five years in market faced displacement as the legal AI boom arrived. Rivals like Harvey AI and Legora raised large rounds and captured headlines, siphoning AI talent away from established players. The client's product was not AI-powered, leaving them unable to attract the engineering talent needed to compete. Without a credible AI product in market quickly, the company risked losing customer trust, market share, and access to future capital.
What They Built
Bonsai Labs embedded as an AI strike team and built a proprietary legal research assistant on Microsoft Azure and OpenAI, trained on the client's private document dataset — reaching beta with paying customers at month three and general availability at month six.
Bonsai Labs embedded with the client as an AI strike team with a three-month mandate to ship a proprietary legal research assistant. The engagement ran in four stages. First, deep domain immersion: the team worked alongside in-house lawyers to understand how legal research actually functions, what questions practitioners ask, and what accuracy requirements would determine trust. Second, infrastructure: a privacy-first build on Microsoft Azure was designed to keep the client's proprietary document dataset secure while enabling OpenAI-powered retrieval. Third, evaluation dataset co-creation: rather than relying on generic benchmarks, legal staff helped build the evaluation set to reflect real practitioner queries. Fourth, tight weekly sprints benchmarked against a single KPI: first-answer correctness. The team discovered that lawyers abandon a tool instantly if the first response is wrong — making that metric the primary engineering focus. Beta launched at month three with paying customers. General availability hit at month six. The resulting traction unlocked a new funding round.
AI Role
Despite operating in a high-security, privacy-sensitive legal environment with millions of documents to index, the team reached beta with paying customers in exactly three months — the original mandate.
Infrastructure
• Microsoft Azure (privacy-first cloud infrastructure) • OpenAI (LLM for legal research and retrieval) • Client's private document dataset (proprietary training/retrieval corpus) • Custom evaluation dataset (co-created with legal staff) • RAG (retrieval-augmented generation) architecture
Integration Points
• Client private document dataset → Azure-hosted vector store → RAG retrieval layer • Practitioner legal queries → OpenAI LLM → first-answer legal research responses • Responses → evaluation dataset benchmarking → weekly accuracy sprint reviews • Validated product → beta (month 3) → general availability (month 6) → funding unlock
