---
title: "What Is a Decision Model? How It Differs from an LLM"
description: "A decision model reads a case and a fixed set of options and returns a calibrated probability for each in one pass, instead of writing text. How decision models work, when to use one over an LLM, and how they are measured."
canonical: "https://www.nace.ai/blog/what-is-a-decision-model"
updated: 2026-11-12
---

# What Is a Decision Model? How It Differs from an LLM

> A decision model reads a case and a fixed set of options and returns a calibrated probability for each in one pass, instead of writing text. How decision models work, when to use one over an LLM, and how they are measured.

Blog · Nov 12, 2026 · Nace.AI

**Drex, a small model that decides.** Generally available. Ahead of Jev on the public Decision Index 0.3.1, with about 3B parameters active per token.

- [Explore](https://www.nace.ai/drex)
- [Get started](https://console.nace.ai/?utm_campaign=drex-launch&utm_content=announcement&utm_medium=website&utm_source=nace.ai)

Much of the work AI does in production is choosing, not writing. Which queue gets this ticket. Whether this transaction looks like fraud. Which button an agent clicks next. Whether a passage answers the question. The usual way to make those calls is to ask a large language model, let it think, and parse an answer out of the text it writes. That is slow, it costs output tokens, and the same input can come back with different answers.

A decision model is built for these calls and nothing else.

## What is a decision model?

A **decision model** is an AI model that reads a case and a fixed set of options and returns a probability for every option in a single forward pass. It does not generate text. It can only answer with the options you gave it, and how sure it is comes back as a number you can act on.

You send three things:

- **The case**: the facts the decision depends on, such as a ticket, a transaction, a document or a screenshot.
- **The question**: what is being decided.
- **The options**: the answers you will accept.

You get back one probability per option.

## Decision model vs. LLM

In the language of dual-process psychology, a chat model is System 2: it writes its way to a decision one token at a time and names an answer at the end. A decision model is System 1: it has seen the shape of the decision before and answers in one pass.

|  | Language model (LLM) | Decision model |
| --- | --- | --- |
| How it answers | Writes text token by token, then names an answer | Reads the case once and scores every option |
| What comes back | A sentence, or JSON to parse | A probability for each option, and nothing else |
| Answers outside the options | Possible | Not possible |
| Billing | Input and output tokens | Input tokens only |
| Confidence | Not measured directly | Measured, and can be thresholded |
| Failure modes | Truncation, malformed JSON, invented options | Wrong option, with the confidence to show it |

Several questions about the same case can share one pass over it.

## What a decision model can answer

Decision models answer three kinds of question:

- **Choice**: pick one of several named options, such as which team should handle a ticket. You get a probability for every option.
- **Yes or no**: for example, does this invoice match the purchase order. You get the probability of yes.
- **Score**: place the case on a scale you define, such as risk from low to critical. You get the expected level and the spread across levels.

## Why use a decision model

**Speed.** Without a token stream there is no waiting for an answer to be written. In our [technical report](https://www.nace.ai/research/towards-foundation-decision-models), Drex Max answers a typical request in 183 ms. GPT-6 Sol takes about 2 seconds and Claude Opus 5.5 about 3.5, on Decision Bench v4, where Drex Max is statistically tied with GPT-6 Sol and 1.6 points behind Opus.

**Calibrated confidence.** A well-calibrated decision model is right about as often as it says it is. When Drex Max is at least 93% sure, it is right 97% of the time. That is what lets you automate the confident calls and send the rest to a person.

**Consistency.** Ask the same question twice and you get the same answer. When the options in a prompt are shuffled, only 4.3% of Drex Max's answers change, and most of those are ones it had already marked as unsure.

**Nothing to parse.** There is no free-form output to validate. Truncated responses, broken JSON and answers that are not one of the options cannot happen.

## Where decision models fit

Decision models work best on high-volume calls where the options are known ahead of time:

- **Triage and routing**: sending tickets, emails and requests to the right queue.
- **Risk scoring**: [fraud alerts, disputes and credit risk bands in banking](https://www.nace.ai/industries/banking), or claims severity in insurance.
- **Document checks**: whether a value matches, whether a clause is present, whether a file is complete.
- **Search and retrieval**: deciding which pages and tables hold the answer, so an agent reads only those. See [Drex for Search](https://www.nace.ai/drex/search).
- **Computer-use agents**: choosing the next click on a screen. See [Drex Max for browser use](https://www.nace.ai/drex/browser-use).
- **Judging agent runs**: scoring whether an agent finished its task.

## When an LLM is the better tool

A decision model is not a replacement for a language model:

- **Open-ended output.** Writing, summarizing and code need a model that generates text.
- **Options you cannot list.** If the answer could be anything, there is nothing to score.
- **Closed-book knowledge.** A decision model is strongest when the facts are in the case you send. Answering from what a model memorized is still where large language models lead.

The two often work together. An agent built on an LLM plans and writes, and hands the many small choices along the way to a decision model.

## How decision models are measured

Decision models are scored on whether they pick the right option and on how well their confidence matches how often they are right.

- **The Artificial Analysis Decision Index** is an independent benchmark that runs every model itself on six evaluations, covering documents, tool calls and consistency. Its scoring costs a model the most for confident mistakes. See [how the top decision models rank on the Decision Index](https://www.nace.ai/drex/decision-index).
- **Task benchmarks** such as DecideBench and JevBench test narrower skills, and OSWorld-Verified is used to test how well a model judges computer-use runs.

## Try a decision model

[Drex](https://www.nace.ai/drex) is our family of four open decision models. Weights and inference code are on [GitHub](https://github.com/nace-ai/drex-decision-models) and [Hugging Face](https://huggingface.co/nace-ai), and Drex Max, which also reads screens, images and video, is served through the Drex API. The [technical report](https://www.nace.ai/research/towards-foundation-decision-models) covers how the models are built, trained and evaluated.

## About the product

**Drex is generally available.** Facts in, a probability for every option out. Benchmarks, board games, the playground and the API are on the product page; an API key takes a minute.

- [Get started](https://console.nace.ai/?utm_campaign=drex-launch&utm_content=announcement&utm_medium=website&utm_source=nace.ai)
- [Explore](https://www.nace.ai/drex)
- [Talk to sales](https://www.nace.ai/contact?product=drex&source=blog)

## For agents

- [Drex API guide](https://www.nace.ai/drex.md): the decision model: request shape, question types, benchmarks and deployment, as markdown
- [Decision endpoint](https://console.nace.ai/v1/systemone): POST with model "drex-latest", a state string and typed questions (choice, noul, score); https://console.nace.ai/v1/models lists checkpoints
- [Document endpoints](https://console.nace.ai/v1/documents/{parse|split|classify|extract|ground}): POST a source (a public URL, an uploaded workspace file or a finished parse job) and poll GET https://console.nace.ai/v1/documents/jobs/{id}; the same key and credit as decisions
- [Console](https://console.nace.ai): API keys, the playground, usage and billing; send `Authorization: Bearer $NACE_API_KEY` on every request
- [Console docs](https://console.nace.ai/docs): quickstart, question types, the document routes, errors, limits and pricing
- [Nace Document Processing](https://www.nace.ai/perception.md): the Perception series: parse, split, classify, extract and ground any file type with a location on every value
- [Drex for Search](https://www.nace.ai/drex/search.md): Drex for Search: decision-powered retrieval for agents, Drex finds the page, cited search through the CLI, SDKs, MCP and REST, and the public benchmark ranks
- [Industries](https://www.nace.ai/industries.md): Nace by industry: the files Document Processing reads and the decisions Drex makes for accounting, banking, mortgage, insurance, legal, healthcare, government, construction, logistics and professional services
- [Developers](https://www.nace.ai/developers.md): The execution layer for agents: Perception, Decision and Search behind one key, the Drex Max frontier, install steps, a scripted agent session and the four deployment boundaries
- [Site index](https://www.nace.ai/llms.txt): every page on nace.ai; append `.md` to any page URL or send `Accept: text/markdown` for the markdown version