Jurin AI Inc., a company building AI agents for high performing enterprises, has introduced JJ-1, a new family of decision models designed for structured decision-making inside enterprise workflows. The announcement signals a shift from AI that generates answers to AI agents that take action. JJ-1 focuses on decisions such as routing customer requests, assessing urgency, and determining whether escalation is required.
A Decision Layer for Enterprise Workflows
Rather than generating an answer token by token, JJ-1 scores predefined options directly and returns probabilities that software can use to determine the next action. The model is engineered around three priorities: speed, purpose-built design, and structured output. Typical requests are processed in approximately 8 to 30 milliseconds on a single NVIDIA B200.
Structured outputs include choice probabilities, yes/no probabilities, and scores that applications can use directly. For developers building AI agents and SaaS products, these decisions can occur many times within a single workflow. A system may classify a request, assess urgency, choose a route, decide whether to escalate, and select the next action before any work is completed.
Benchmark Results and Cost Efficiency
Across more than 88,000 questions spanning six evaluation areas, including public English- and Japanese-language datasets and Japanese enterprise workflow data, JJ-1 achieved an aggregate accuracy of 79.2%. That result compares with 76.5% for Jev 1.13 and 71.35% for GPT-6 Luna Decision. The company also reports that JJ-1 delivered up to 99% lower inference cost than generative large language model approaches.
On the Japanese enterprise workflow benchmark, JJ-1 reached 97.43% accuracy, while GPT-6 Luna Decision recorded 91.99% and Jev 1.13 recorded 91.75%. The company reports this performance alongside up to 99% lower inference cost than generative large language model approaches. These advantages matter for high-volume agent deployments where multiple decisions occur inside a single workflow.
Handling Japanese Business Language
Japanese business language can be highly context-dependent, and the company highlighted the phrase 'kekkō desu' as an example. Depending on the conversation, the expression can signal either acceptance or refusal. In recorded examples, JJ-1 correctly distinguished between these uses and returned the decision in milliseconds.
Leadership Commentary
Rise Ooi, Founder and CEO of Jurin AI, said that once AI agents start doing real work in Japan, they make decisions constantly. He noted that one user request can involve several decisions before anything gets done, and at scale every extra model call adds time and cost. Ooi added that the future is not one giant model doing everything, but rather the best generative models for deep reasoning or generation and a decision model for routing, urgency, and escalation.
Huck Yang, Ph.D., an IEEE Technical Committee Member and IEEE SPS member, commented on the infrastructure significance of the JJ-1 release. He said the model demonstrates that a world-class text model can be built and trained entirely on Japan-based cloud infrastructure. Yang added that leveraging ASIC and GPU technology helps democratize the Japanese open-source ecosystem for the broader community.
Availability and Technical Report
Jurin AI is publishing the full benchmark methodology, testing conditions, and cost comparisons in the JJ-1 Technical Report, which is available on the company's website. Developers can access the model to evaluate its decision-making capabilities in their own workflows. JJ-1 is available to try today at inference.jurin.ai.
The launch of JJ-1 reflects a broader industry movement toward specialized models that handle high-frequency, structured decisions inside enterprise software. By optimizing for speed, cost, and Japanese business language, Jurin AI aims to address a practical bottleneck in agent deployments. The company's benchmark claims and public availability provide an early look at how purpose-built decision models may complement generative AI systems.