On October 28, Articul8 announced Bivia, a multimodal decision model offered through an API. The company positions it for specialisation around business rules and procedures. This is a product announcement with internal evaluations, not independent confirmation of accuracy across every industry.
The development is interesting for agents that must choose a next step without necessarily writing an explanatory paragraph. An application might need to select the team handling a request rather than compose a customer response. That is an example of possible use, not a Bivia trial conducted by SIExpo.
What the developer receives
The product page describes typed outputs including choices, Boolean values, scores and extracted spans. An application should check status before value. For abstained responses, value is null and an abstention reason is provided. The allow_abstain option enables that behaviour; disabling it makes the model choose an answer.
This distinction matters immediately during integration. If code treats every returned package as a successful decision, an abstention may disappear. A separate path for missing information, further review or human handling would be more useful. This is an editorial proposal, not a claim that a particular customer has already implemented that process.
Articul8 describes confidence as calibrated. A number in a response, however, is not our verification of the probability of a correct decision on somebody else's data. The product page includes both illustrative comparisons and replays of captured responses. These show the interface and concept, but are not an independently reproducible evaluation of a real organisation.
An answer and an action are separate steps
One possible initial experiment is routing internal requests in observation mode. A person continues to determine the correct team while the system records its suggestion. The evaluation could then compare correct choices, abstentions and cases where a confident wrong choice would create additional work.
We would retain ambiguous examples in such an evaluation. A test containing only obvious cases would barely examine the value of abstention. An abstention is also not necessarily a failure: it may be the appropriate outcome when the input lacks sufficient information. Reporting only the proportion of answered requests would miss that distinction.
A team should also specify what may follow a choice. Recommending a folder differs from an action with material consequences. An API response does not itself grant permission to execute. The application should retain an explicit definition of that permission and its boundaries.
Those boundaries matter when comparing systems, too. A comparison that lets one system defer but requires another to answer every time would measure different tasks. Our suggested assessment would keep the permitted outcomes consistent and inspect the consequences of errors, not merely the attractiveness of the returned format.
What zero output tokens means
The site advertises decisions without generated output tokens. This does not imply an absence of input data, computation or service charges. Likewise, we infer no universal acceleration from the demonstrations and promise no specific savings.
SIExpo has not sent private data to the API, activated a subscription or benchmarked the model. The verifiable development is the product offering and its described response contract. Whether the approach improves a particular agent remains a question for a specific, predefined evaluation.
*Cover: AI illustration of inputs, decisions and abstention, not an actual product screen.*
*Editorial verification: 31.10.2026, 09:12:27 +0200 (Europe/Sofia).*
