Shionogi improved AI accuracy from 50% to 90% using proprietary data

Shionogi Pharmaceutical has improved a generative AI system's accuracy from 50% to 90% by optimizing how the system handles the company's extensive proprietary data. The pharmaceutical company manages vast quantities of confidential information—research findings, trial data, chemical structures—and the challenge has been to harness AI's capability to understand and generate text without exposing or misusing that sensitive material. The title of the article (translated as "Shionogi Pharmaceutical raised generative AI accuracy from 50% to 90%: how did they optimize vast confidential data?") frames the core question: how can a regulated company deploy AI reliably while protecting trade secrets and regulated information? Shionogi achieved this improvement but has not yet disclosed the specific technical methods—whether through fine-tuning the AI model on proprietary data, implementing retrieval-augmented generation (feeding the AI curated internal documents), applying domain-specific rules, or another optimization technique. The article does not specify when the system will move from development to production use, whether it will be shared with partner organizations, or whether Shionogi intends to offer the capability as a commercial product. The milestone is significant because it suggests that pharmaceutical and other regulated industries can move beyond treating generative AI as an exploratory tool and toward building systems with sufficient accuracy and safety for operational use on mission-critical tasks—provided the organization invests in careful data preparation and optimization.