AI Decision Models: Distinguishing Them from LLMs

Decision models have recently emerged as a distinct approach to AI, with Jev by TypeSafe AI being a notable example. Rather than generating free-form text, Jev delivers structured options, scores, and confidence estimates.

Although conventional AI models have long performed predictions and classifications, Jev is specifically tailored to make structured decisions that software can consume directly. This method offers an alternative pathway for building AI applications, particularly when a system must evaluate data and determine the subsequent action.

How does Jev function?

In a standard LLM workflow, you submit a prompt and receive generated text. Even when JSON is requested, the application typically needs to parse and validate the response before it can utilize the data.

Jev operates differently. You supply context, define specific questions, and specify the expected output type. Jev then returns structured results that your application can process immediately.

Jev supports three primary question types:

• Choice: Selects an option from a defined list, such as billing, technical support, or sales.
• Score: Rates an item on an ordered scale, such as low, medium, or high priority.
• Yes or no: Estimates the probability of a statement's truth, such as whether a support ticket requires human intervention.

You can pose multiple questions about the same input within a single request. For instance, a support application could classify a ticket, assess its urgency, and estimate if human review is necessary. Jev returns these results in a structured format, enabling the application to determine the next steps.

Decision models vs. LLMs

LLMs and decision models serve distinct purposes. An LLM is capable of explaining decisions, generating content, writing code, or reasoning through complex tasks. Jev, on the other hand, is designed for focused decisions with predefined outputs.

This distinction does not imply that LLMs are incapable of making decisions. They can indeed classify information and return structured data. However, Jev is built specifically around this task, eliminating the need for an application to interpret a free-form response.

This makes it particularly useful for repetitive, well-defined decisions. For tasks requiring detailed explanations, creative output, or complex multi-step reasoning, an LLM remains the more appropriate tool.

What are confidence scores?

A key feature of Jev is that its outputs include probabilities and confidence information. These values assist developers in assessing the strength of the model's preference for a particular answer.

For example, a customer support application might ask if a ticket should be escalated. If the model indicates a high probability of escalation being necessary, the application can forward the ticket to a senior support team. If the result is uncertain, it can route the ticket for human review.

However, confidence does not guarantee correctness. A model can be highly confident yet still make an incorrect decision. Developers should test results against real-world examples and select thresholds that align with the consequences of each decision.

It is also crucial to distinguish the probability of an answer from a separate confidence score. The exact interpretation depends on the output type. Applications should adhere to the model's documentation rather than assuming all returned numbers represent the same metric.

Where can decision models be useful?

Classification and routing

Many applications require sorting incoming information before processing it. A support platform might classify requests by topic and direct them to the appropriate team. A business application could categorize documents, emails, or customer feedback.

Jev can return a category from a predefined list, allowing the application to route the item without parsing a generated paragraph.

Scoring and evaluation

Decision models can also evaluate information against a defined scale. For instance, an application could rate a support ticket by urgency, score a document for relevance, or assess if a response meets specific criteria.

These results can be used to rank items, prioritize work, or identify cases requiring further review. The quality of these decisions still relies on the model and how accurately the criteria describe the task.

AI agents and tool selection

AI agents utilize models to determine which actions to take. An agent might select a tool, classify a request, or decide if a task needs additional information.

While a general-purpose LLM can make these decisions as part of its reasoning, a decision model can handle narrower choices before the agent acts.

For example, a custom workflow using Hermes Agent could employ Jev to classify an incoming task before passing it to the agent. It could also assist in determining which workflow should handle a request. This would require a custom integration rather than relying on assumed native support.

Decision models should not be treated as standalone security controls. If an agent can execute commands or modify files, permissions and deterministic rules must enforce which actions are allowed. A model's classification can inform this process, but it should not replace it.

Alternatives to Jev

Jev is part of a broader category of models designed to produce structured decisions. OpenJev and Laya are two other projects worth exploring when comparing approaches.

OpenJev

OpenJev is an open project related to the Jev approach. It is worth examining if you wish to explore an implementation outside a hosted service. Before using it, check its current availability, hardware requirements, supported features, and license.

Laya

Laya is an open-source decision model from Convai Innovations. It focuses on tasks such as choosing from predefined options, rating information, and answering yes-or-no questions. It may be relevant if you want to evaluate a model that can run in your own environment.

These projects should be compared based on actual performance, not just their stated capabilities. Use the same test data and measure accuracy, latency, resource requirements, and how well their confidence estimates align with real outcomes.

Can you run decision models locally?

Running a model locally provides greater control over the software environment and data processing location. However, available deployment options vary by project. Jev is offered through an API, while open-source alternatives may provide options for running inference on your own hardware.

Hardware requirements depend on model size, memory use, and throughput. Some models can run on a CPU, while others benefit from a GPU, especially when processing many requests.

If you want to experiment with local AI models, DaDesktop provides cloud desktops with dedicated GPU options. You can use them for development and testing without purchasing a high-end GPU for your local computer.

Learn more about running local LLMs or explore AI agent environments.

When should you use a decision model?

Decision models are useful when an application needs a defined answer that can trigger a specific next step. They can complement LLMs by handling classification, scoring, and routing while a generative model manages language and more complex reasoning.

The best choice depends on your application. Test the model against your own data, compare it with conventional classifiers and LLMs, and measure accuracy, latency, and cost. For important decisions, include human review and safeguards where necessary.

Want to talk about your use case? Contact us and we'll point you in the right direction.