Bivia wants to replace long AI responses with immediate and verifiable decisions

Articul8 launches Bivia, a decision-making multimodal AI capable of analyzing documents, images, and videos without generating text.

An AI agent can make 99 correct decisions and still fail its mission because of a single mistake. That is the problem Articul8 AI is targeting with Bivia, a new multimodal AI model designed not to hold conversations, but to make precise decisions within complex workflows.

Introduced on October 28, 2026, Bivia can analyze text, documents, images, video, audio, and industrial signals. Instead of generating several sentences of reasoning before reaching a conclusion, the model directly returns a structured decision, a confidence score, and, where applicable, the evidence supporting its choice.

Articul8 says the approach builds on more than three years of experience deploying AI in semiconductors, energy, manufacturing, and other industrial environments. Bivia is now available through an API and integrations with existing development tools. Why AI agents fail despite making mostly correct decisions

Autonomous agents are typically built around large language models that can reason, write code, and call external tools.

However, complex workflows often require hundreds of smaller decisions: choosing the right tool, checking whether a document meets requirements, deciding whether a task is complete, identifying an anomaly, or determining when to involve a human.

Calling a large generative model for every decision can produce unnecessary text and increase inference costs.

Reliability is another challenge.

Even an agent that makes the vast majority of its decisions correctly can fail when one mistake propagates through subsequent steps. Choosing the wrong tool or prematurely approving an action can compromise an entire workflow.

Articul8 is positioning Bivia as a specialized decision layer that operates alongside large language models rather than replacing them completely.

The idea is to reserve generative models for planning, reasoning, and open-ended work while delegating bounded decisions to a system optimized for structured judgments. Structured answers without generating output tokens

Bivia's defining feature is its decision-oriented interface.

Consider a coding agent that has just modified a file but has not yet run its tests.

A conventional language model might generate a paragraph explaining the available options before recommending that the tests be executed.

Bivia instead receives the project's current state, a question, and a list of possible actions. It directly selects an option such as "Run the tests," accompanied by probabilities for the other available choices.

Responses can include the selected value, a confidence estimate, and a status indicating whether the system has sufficient evidence to answer.

Articul8 emphasizes that Bivia makes these decisions without autoregressively generating output tokens.

That does not mean the model performs no computation. Rather, it avoids producing words sequentially when an application only needs a category, score, or action.

One example on the company's website shows an API decision completed in 55.8 milliseconds.

That is the latency of a particular demonstration, not a guaranteed response time for every request, especially those involving large documents or videos. Nine decision types in one model

Bivia offers different operations depending on the output an application needs.

Decision type · Purpose Decide · Select the next action Choice · Choose among alternatives Boolean · Return yes or no Score · Assign a numerical rating Set · Identify multiple applicable items Span · Locate supporting text Ground · Identify regions in images or video Extract · Retrieve structured fields Data science · Make judgments from data or signals

This structure turns a broad range of problems into outputs that software can process directly.

In an insurance application, Bivia could identify relevant policy clauses, evaluate a condition, and determine the next processing step.

In an industrial setting, the same model could examine an equipment photograph and decide whether an anomaly requires inspection.

Outputs can include image coordinates, exact document excerpts, and video timestamps.

Articul8 therefore presents Bivia not only as a system that makes decisions, but also as one capable of showing what evidence supports each decision. Direct analysis of images, video, and industrial data

Bivia is multimodal, supporting text, code, documents, images, audio, video, and signals.

That matters in industrial environments, where relevant information is rarely confined to one document or database.

A maintenance decision might require an equipment photograph, sensor readings, operating procedures, and previous incident reports.

In one demonstration, Articul8 asks Bivia to locate damage on a car. The model identifies the damaged region and returns a bounding box, while TypeSafe's Jev model is shown still processing the same request.

The company also demonstrates Bivia analyzing a 52-minute, 4K documentary filmed from the International Space Station.

According to Articul8, the model processed every frame and created a searchable index capable of identifying brief events, including an aurora and city lights visible from orbit.

The company contrasts this approach with systems that analyze only selected frames.

However, the demonstration does not include independent measurements of processing cost, latency, or event-detection accuracy. It illustrates an advertised capability rather than establishing a universal advantage over competing systems. A reported 73% workflow completion rate versus 37%

Articul8 highlights an internal evaluation focused on the completion of entire workflows.

In its launch posts, the company reports 73% successful workflow completion for Bivia, compared with 37% for a comparison model.

This metric is particularly relevant because it measures the outcome of a chain of decisions rather than the accuracy of isolated answers.

In a separate internal tool-routing evaluation, Articul8 reports a score of 0.91 for Bivia, compared with 0.84 for TypeSafe's Jev.

The company also tested whether reordering