Roche cuts GenAI build time from months to days and saves $100K–$250K annually in patent attorney hours

Roche’s patent process is high-stakes and labor-intensive: European patents often trigger oppositions and appeals that require exhaustive research across roughly 50,000 EPO Board of Appeals cases. Attorneys spent many hours on reference books and keyword searches, a slow, costly process that often produced incomplete insights. Rising case volumes and pressure to respond faster left the Basel patent team with limited scalability and rising costs.

In 2021 Roche adopted Dataiku to bring GenAI into patent law analysis, starting with pilots for semantic search, full-text analysis, and deep search of EPO appeal cases (AskWhitebook). These efforts evolved into Themis PatAI, an agentic AI workspace that unifies specialized GenAI projects—including AskWhitebook and TheLake (natural-language Q&A over department Google Drive documents)—via Dataiku Agent Connect. The agent routes queries to the right sub-agent and was initially built for 80 European patent attorneys and paralegals.

Time to build new GenAI projects fell from months to days. Roche reports $100K–$250K in annual savings primarily from reduced attorney hours per case, plus $375K–$475K saved by avoiding consultancy costs through in-house citizen development. Insight quality improved for more complete case law analysis, and TheLake is used by 30 attorneys every month.

• • Citizen development on a governed GenAI platform can let IP attorneys ship complex legal research tools without IT bottlenecks.
• • Unifying fragmented RAG/search projects behind an agentic orchestrator reduces switching cost and improves attorney adoption.
• • LLM Mesh-style prompt/model comparison with token-cost visibility supports responsible experimentation on sensitive legal data.