In a proof of concept, Astellas Pharma's multi-agent AI for handling pharmaceutical manufacturing deviations reached 69% coverage of cited information and 81% accuracy, meaning no incorrect information was mixed into answers. Under certain assumptions, the company estimated that deviation-handling effort would fall by 67%, or 98 hours, and that rolling the system out globally to three research laboratories, five production sites and QA would save 980 hours per month. The case was presented at a media session during Microsoft AI Tour in Tokyo, held at Tokyo Big Sight.
A deviation is an event in which production departs from set procedures or standards, such as foreign matter getting into a product or a machine stopping temporarily, and the company must judge whether the affected lot can be shipped. When a deviation was reported, a CMC team made up of quality assurance, research and production staff searched for and organized the relevant information and decided the response after discussion. Astellas's Capability Enhancement Lead described the problem, in a statement translated from Japanese: "Ideally we would search comprehensively based on similar past cases and choose the best-informed response policy, but because a large volume of past deviation cases is registered, it was difficult and time-consuming to search them comprehensively and do a first-pass organization. As a result, we depended on the knowledge of experienced staff, and determining a response policy took a great deal of effort and time."
Astellas judged that general generative AI was not sufficiently comprehensive or accurate, so it adopted a multi-agent design in which specialized agents work under an orchestrator and run feedback loops to verify and improve results. The agents are Data Feeder, Credibility Critic, Summarizer, Report Writer, Reflection Critic and Translator, coordinated by a Manager agent. The environment runs on Microsoft Azure App Service, with information searched through Azure AI Search and Azure OpenAI Service used to solve problems based on that information. According to Astellas, a single agent matched the multi-agent system on accuracy, but the multi-agent system was far better on coverage.
In one demonstration, asked whether a batch could be released or should be discarded after a freeze dryer leak test failure, the system answered that such a failure is in principle treated as a critical deviation and the batch is basically discarded, while also showing conditions under which shipment is possible. Asked for as many past metal foreign matter deviations as possible, it identified 113 cases. Use will start in research laboratories and expand in stages from the next fiscal year, and Astellas plans to apply multi-agent AI next to setting the Quality Target Product Profile.