Articul8 推出决策模型 Bivia，帮企业 AI 智能体做结构化决策

Articul8 has introduced Bivia, a decision-making model designed to help enterprise AI agents evaluate information and make structured decisions across text, images, documents and other data formats.

Available through an application programming interface (API), Bivia is designed to handle the repeated judgments involved in completing tasks without relying on a generative model for every step. The launch addresses the computational requirements of AI-agent workflows, where even routine decisions can require repeated calls to large language models (LLMs).

“Enterprise AI is moving beyond generating information to taking action, and that shift makes decision-making just as important as generation,” Ananthanarayanan S, Managing Director, India at Articul8, said in a statement.

Unlike generative models designed to create content, Bivia focuses on decisions such as selecting tools, verifying evidence, identifying inconsistencies and determining whether a task is complete. It can be specialised using an organisation’s data, industry knowledge, policies, standard operating procedures and historical decisions.

The model supports five decision types: Choice, Boolean, Score, Set and Span. These allow AI agents to select options, determine whether a statement is true, assign scores, identify applicable categories and locate supporting evidence.

Bivia can analyse text, images, documents, engineering drawings, tables, charts, audio and video. It can identify relevant image regions, text passages and video timestamps to ground its decisions in source evidence. The model also supports multilingual applications, including Hindi, Gujarati, Tamil and Sanskrit.

Articul8 offers three operating modes: Fast for quick decisions, Fast+think for decisions requiring additional computation, and Deep for decisions requiring further verification. Bivia can also abstain when the available evidence is insufficient to support a reliable decision, rather than returning an unsupported answer.

In its internal evaluations, Articul8 reported near-perfect accuracy on selected energy and hardware design tasks following domain-specific training. The company also reported that Bivia completed approximately 70% of workflows in an internal evaluation, compared with approximately 30% for the alternative evaluated.

The company demonstrated Bivia’s multimodal capabilities by processing every frame of a 52-minute, 4K documentary filmed from the International Space Station in a single request. Available through an API as a standalone offering or as part of Articul8’s enterprise AI platform, Bivia is designed to work alongside generative models, handling decisions throughout a workflow while those models manage tasks such as writing, explaining and planning.