Decision Models and Jev – Definition, How They Work, and Costs
Decision models output structured decisions. This entry explains how they work, their costs and latency, as well as branchly's assessment of Jev.
What are decision models?
Decision models are AI models that output structured decisions rather than open-ended responses. Typical results are a category or a score; some systems also output a probability. Applications use the results, for example, to classify content or route inquiries. Software can process them directly instead of having to extract them from longer response texts.
For clearly defined tasks, decision models can be faster, more cost-effective, and more reliable than general language models. This makes them compelling for internal workflows where context needs to be evaluated and rigid if-then rules are insufficient. An outputted probability can help contextualize the result and forward uncertain cases for review.
The term is not used consistently. Some providers use it to describe a specifically trained model, while others mean an API or a language model that outputs responses in a predefined format. Therefore, the training approach, permitted outputs, and results for the specific use case are more meaningful than the product name alone.
An example: A system categorizes a support inquiry as "product consulting" or "customer service" and routes it to the appropriate process. A language model can accomplish the same task using structured outputs; a separate decision model is not strictly necessary for this.
How do decision models work?
The process usually involves four steps:
-
Define task and options: The team defines which categories, actions, or values are permitted.
-
Provide inquiry and context: The model receives the information it needs to make the selection.
-
Output the decision: The result is returned in a defined format, for example as a category or score.
-
Catch error cases: The team tests the decision with realistic examples and defines when a human should take over.
A probability can serve as an indicator for downstream systems of how confident the model is in its decision. This allows clear cases to be handled automatically and uncertain cases to be routed for review. This only works reliably when the probabilities are calibrated for the specific use case. A fixed output type alone does not guarantee a correct decision.
Structured outputs and function calling can also deliver machine-readable results. Whether a specialized decision model performs better depends on the model, the interface, and test results—not solely on the output format. Therefore, compare systems using the same tasks and data.
Costs and latency: what can be compared
For clearly defined tasks, decision models can be more cost-effective and faster than general language models. Their structured outputs are often shorter and avoid additional tokens for explanations. Total costs still depend on the model pricing, input length, hosting, and usage volume. For self-hosted models, infrastructure and operations must also be factored in.
Response time depends on the model, hardware, and specific system architecture. Manufacturer metrics and demo speeds therefore do not automatically reflect how quickly a model responds with your own inquiries and under real-world load.
TypeSafe AI quotes $0.042 per million input tokens for Jev and a response time of 70 to 500 milliseconds (Source: TypeSafe AI, 2026; Pricing details, 2026). These are manufacturer figures. Actual total costs and performance depend on the specific workflow, so models should be compared using identical tasks and quality criteria.
Open-source models can be hosted locally and customized. This does not make them "free": compute power, integration, monitoring, and maintenance must be included in the calculation. The advantage may lie in greater control over the model and operations, not necessarily in lower total costs.
At branchly, platform pricing is based on active sessions, not the number of individual model calls. Starter costs €1,000 per month and includes 2,000 active sessions; Professional costs €2,000 and includes 6,000. Session limits are calculated on an annual basis to account for seasonal fluctuations. A platform plan therefore cannot be directly compared to per-token API billing (Source: branchly, 2026).
Comparing decision models and language models
| Feature | Generative language model | Decision model |
|---|---|---|
| Typical output | Natural language response | Structured selection, category, or score |
| Best suited for | Dialogues, explanations, and open-ended questions | Clearly defined selection and routing tasks |
| Speed and costs | Increase with response length and scope, among other factors | Can be lower for a specific decision task |
| Probability | Not necessarily part of the response | Can attach a probability to the decision |
| Reliability | Good for diverse language tasks, but can generate incorrect responses | Can be more reliable for a clearly defined task; must be validated for this |
| Limitation | A plausible-sounding response is not automatically correct | Even a formally valid and highly rated decision can be incorrect |
The approaches are not mutually exclusive. A language model can understand or explain an inquiry; a decision model can then output a defined decision. This allows open-ended communication and structured decisions to be combined within the same workflow.
Jev by TypeSafe AI: what the hype is about
TypeSafe AI describes Jev as a model for structured decisions and cites RLCD (Reinforcement Learning on Constructive Decisions) as its training approach. According to the provider, Jev outputs a selection or a score along with an associated probability, for example. The published pricing and latency data also originate from TypeSafe AI. They do not yet indicate how well Jev solves a specific task in enterprise use.
The claim of "no hallucinations" also requires context. If a model only outputs values within a predefined format, it cannot formulate an open-ended response outside of that format. However, it can still make the wrong choice. A restricted response format does not prevent factual errors.
Speed and pricing also depend on what you compare them against. Before deployment, companies should use their own test cases to evaluate how often the model decides incorrectly and what total costs will arise based on expected usage.
What interests us most about Jev
Traditional machine learning classifiers and semantic classification models are not new. What we find exciting about Jev is its more general approach: a decision model that can be applied across different domains without having to be retrained for every workflow. Added to this is the probability with which the model outputs its decision. Combined with low pricing, this makes Jev compelling for internal decision-making processes.
At branchly, we tested Jev ourselves. Due to data protection considerations, we are not yet using it in production. We see great potential and are planning to use decision models across multiple workflows.
Open-source approaches and the OpenAI Decisions API
Following the introduction of Jev, open-source projects such as Laya and OpenJev were discussed as alternatives. They follow the principle of structured outputs, but differ in model foundation, training data, licensing, and maturity. Community benchmarks can provide orientation, but a single benchmark does not prove performance in a different use case.
Before production deployment, licensing terms, maintenance status, security requirements, and the reproducibility of test results must be reviewed. Self-hosting provides greater operational control, but brings infrastructure overhead and ongoing responsibility.
OpenAI announced a Decisions API in preview in September 2026. The announcement focused on outputs from a predefined response range rather than open-ended explanations. Availability, performance metrics, and pricing of a preview are subject to change. Procurement decisions should therefore rely on current product documentation and hands-on testing, rather than a demo alone.
Decision models on websites
A decision model can handle a clearly defined task, such as categorizing or routing an inquiry. For open-ended product questions, however, a system requires access to relevant content and the ability to answer follow-up questions in dialogue. The right technology depends on the task at hand on the website.
branchly is an AI layer for websites. Chat and AI Search help visitors find content and clarify questions; additional modules guide them to relevant next steps. Agentic RAG, hybrid search, NLP, and conversational AI form the technical foundation. These interactions give marketing teams insight into what visitors are searching for.
A decision model could be one potential building block for a clearly defined selection task. branchly stands for a more comprehensive website experience: multiple AI modules through which visitors find content and interact with what you offer.
Related terms
Frequently asked questions
What are decision models in simple terms?
A decision model outputs a selection or evaluation in a predefined format, for example as a category or score. Other software can process this result directly. The format does not guarantee that the decision is correct.
What is Jev by TypeSafe AI?
Jev is a model by TypeSafe AI for structured decisions. TypeSafe cites RLCD as its training approach and states that Jev outputs selection values with associated probabilities. The published performance and pricing figures are manufacturer specifications; whether they hold true for a specific use case must be evaluated separately.
Are decision models more accurate than language models?
For clearly defined tasks, decision models can make decisions more reliably than general language models. An included probability helps contextualize the decision and identify uncertain cases. To be meaningful, it must be calibrated for the specific use case.
Are decision models cheaper and faster?
For clearly defined tasks, they can be faster and more cost-effective than general language models. TypeSafe AI quotes $0.042 per million input tokens for Jev and a response time of 70 to 500 milliseconds (Source: TypeSafe AI, 2026). Actual costs and latency depend on the specific workflow.
Can decision models hallucinate?
A system with a restricted output format cannot generate arbitrary free-text responses outside of that format. However, it can still output an incorrect decision. A valid structure is no proof of factual correctness.
What open-source alternatives to Jev exist?
Projects such as Laya and OpenJev are being discussed as open-source approaches for structured decisions. They differ in their model foundations, licensing, and maturity. Check the current project status and test them with your own tasks before drawing conclusions for production use.
Is Jev already being used in production at branchly?
No. We tested Jev at branchly, but due to data protection considerations, we are not currently using it in production. We see great potential and are planning to use decision models across multiple workflows.
Is branchly suitable for medium-sized businesses and how much does it cost?
branchly is designed for medium-sized businesses with a marketing team responsible for the website experience. Starter costs €1,000 per month and includes 2,000 active sessions; Professional costs €2,000 and includes 6,000 active sessions. The platform is priced based on active sessions and plan tiers, rather than a separate token price per decision (Source: branchly, 2026).
Are decision models GDPR compliant?
The model concept alone says nothing about data protection compliance. What matters is the provider, hosting, data flows, retention, and human oversight in the specific deployment. branchly operates its AI infrastructure in Microsoft Azure data centers in the EU and does not use customer data for training (Source: branchly, 2026).
