
New API enables AI agents to make evidence-based decisions across text, images, videos, documents, and other data formats while reducing reliance on generative models for routine tasks.
SANTA CLARA, Calif., Oct. 28, 2026 — Articul8, a provider of domain-specific generative AI solutions for enterprises, today announced the launch of Bivia, a multimodal decision model designed to help enterprise AI agents make, locate, and verify decisions using organizational knowledge, policies, and operating procedures.
Built on billions of decisions made through Articul8’s enterprise AI platform, Bivia is now available through an API. Developers can use the model to build AI agents that make decisions across text, images, videos, tables, and charts without relying on a generative model for every step.
As enterprises expand the use of AI agents, the cost of using powerful generative models for thousands of routine decisions is becoming a growing concern. According to McKinsey’s State of AI in 2026 report, 40% of respondents from large enterprises said their organizations were scaling AI agents, compared with 27% in the previous year. The same research found that approximately one in five respondents reported that AI-related operating costs, including token costs, were limiting their organizations’ use of AI.
Unlike generic decision models, Bivia can be customized using an organization’s industry knowledge, policies, procedures, and historical decisions. This allows the model to make judgments that reflect specific business requirements. It can also evaluate multiple data types together, including text, engineering drawings, equipment images, and charts.
“We’ve been using decision models within Articul8 for more than three years because we saw how much unnecessary work enterprises were asking generative models to do,” said Dr. Arun Subramaniyan, Founder and CEO of Articul8. “When an AI agent is working through a complex task, it doesn’t need a powerful generative model to answer every question along the way. And businesses need those decisions to reflect their own knowledge and operating requirements. That’s why we’re making this technology available to everyone.”
From Large Volumes of Content to Evidence-Based Decisions
Bivia can analyze an image, document, audio clip, or full-length video and respond to questions in multiple languages. Its responses are linked to the specific pixels, words, or timestamps that support each answer and include a confidence score. When sufficient evidence is unavailable, the model can abstain rather than provide an unsupported response.
During a live demonstration, Bivia processed every frame of a 52-minute, 4K documentary filmed from the International Space Station (ISS) in a single request. The model converted the footage into a searchable index, creating an overview of the content displayed throughout the film. Users could then ask questions and move directly to the relevant moments.
The footage included brief events lasting only a few seconds, such as a view of city lights on Earth’s night side and an appearance of the aurora. Bivia identified these events by analyzing every frame rather than relying on selected image samples, which could have missed them.
“Everyone’s models can caption a thumbnail. Bivia watches the entire thing and tells you the three seconds that matter, with the receipts,” said Dr. Subramaniyan. “It’s the difference between a model that talks about your data and one that actually reads all of it.”
Bivia’s key capabilities include:
- Unified multimodal processing: Text, documents, images, video, audio, and raw sensor signals are processed through a shared model architecture instead of a series of separate models.
- Structured decision outputs: Eight native decision primitives, including decide, yes/no, choose, score, select all, quote the source, find the region, and extract the field, provide structured, machine-readable answers in a single pass.
- Evidence-backed responses: Answers can identify supporting image regions, exact text passages, character offsets in documents, and timestamps in videos. This allows decisions to be traced to their original sources.
- Confidence scoring: Each response includes a confidence score. The model can also abstain when the available evidence is insufficient.
- Full-content analysis: Bivia can process entire videos, documents, and audio recordings in parallel, creating searchable indexes without relying on selected samples.
In industrial applications, Bivia can compare technical requirements with supporting documents and images. For example, it can identify differences between maintenance records and equipment photographs. In Articul8’s internal testing, Bivia achieved 100% accuracy on selected energy and hardware design (Verilog) tasks after being trained on domain-specific data. These in-distribution results demonstrate the model’s ability to be specialized for specific industrial tasks and do not represent performance on previously unseen data.
Designed for Enterprise Decisions at Scale
Bivia supports several decision types, including Choice, Boolean, Score, Set, and Span. Developers can use these capabilities to build applications for evidence verification, tool routing, document processing, and task-completion checks.
The model provides three operating modes, allowing applications to make immediate decisions or apply additional computation and verification when required. It can also abstain when evidence is insufficient, enabling an application to request additional information or escalate a decision.
In Articul8’s internal workflow completion tests involving tool routing, Bivia achieved a score of 0.91, compared with 0.84 for Jev. The company also reports that Bivia delivers consistent decisions regardless of how answer options are presented. While Jev and other leading decision models changed their answers in approximately one in eight cases, Bivia returned the same answer in every test case.
Because Bivia makes decisions without generating output tokens, developers can reduce their reliance on more expensive generative models for routine tasks. This can help lower the operating costs associated with running AI agents at scale.
Bivia is available through an API as a standalone offering or as part of Articul8’s broader enterprise AI platform. The company reports that the technology is already generating substantial savings for customers, including Hitachi Energy, CTC, AZing, and Megazone.
Built on AI Research and Industrial Expertise
Founded in January 2024 as an Intel spinout, Articul8 is led by Dr. Arun Subramaniyan, an aerospace engineer with a PhD from Purdue University and previous leadership experience at Intel, AWS, and GE. The company brings together a global team of AI specialists from industry and academia and reports having filed 68 patent applications.
In June 2026, Articul8 completed its Series B financing at a pre-money valuation of USD 500 million, alongside a strategic partnership with a global industrial software company. At the time, the company reported more than USD 100 million in cumulative contract value. Its platform included more than two dozen domain-specific AI models and integrations with over 25 enterprise and operational systems.
About Articul8
Articul8 AI provides domain-specific generative AI solutions for mission-critical and regulated industries, including energy, manufacturing, aerospace and defense, financial services, and semiconductors. Founded in January 2024, the company helps enterprises convert complex operational and engineering data into reliable, actionable intelligence.
Its platform combines multimodal data understanding, domain-specific AI models, and intelligent agent orchestration to deliver accurate and explainable AI applications that work with existing enterprise systems. Articul8 supports deployments across public, private, and sovereign clouds, as well as air-gapped environments.
The company is SOC 2 Type II certified, with solutions available through the AWS, Microsoft Azure, Google Cloud, and Databricks marketplaces. For more information, visit articul8.ai.
Media Contact

