AI That Decides Without Writing: What Are Decision Models?
AI That Decides Without Writing: What Are Decision Models?
Three weeks. Three launches. All pointing in the same direction.
On September 15, TypeSafe AI released Jev. Three days later, Convai Innovations published Laya, an open-source model. On September 29, OpenAI unveiled the Decisions API at DevDay, and on October 6, opened beta access to all developers.
Three different companies, three different business models, and three products that do the same thing: they decide, without writing a single word.
Essay Questions vs. Multiple Choice
The simplest way to understand the difference is to think about a test.
An LLM answers essay questions. It's the type of model behind the most well-known AI assistants. It writes responses word by word and can write virtually anything: an email, a contract summary, an analysis. That flexibility is what makes it so useful, and also what makes it slower and more expensive, because every word takes time and costs money. Even when the answer is just one word, the model still goes through the full process of composing text.
A decision model answers multiple-choice questions. It reads the input, reads the options, and picks one. It doesn't write anything. Instead, it delivers something a traditional LLM can't reliably provide: a confidence score for each option. Something like "88% chance it's A, 10% chance it's B, 2% chance it's none of the above."
All three models handle the same types of questions: selecting an option from a list, assigning a score on a scale, or determining whether a statement is true. And all three charge, when they charge at all, only for the text they read, not for anything they generate.
Why the Percentage Matters
A well-prompted LLM will also pick the right option most of the time. What it can't reliably tell you is when it's uncertain. Ask it how confident it is, and it will write a number, but that number is just more generated text, not an actual measurement.
With a decision model, confidence comes bundled with every response. That changes what you can safely automate: above a certain confidence threshold, the AI proceeds on its own; below it, the decision gets routed to a human. And that threshold can be set higher in situations where mistakes are more costly.
A large share of the day-to-day decisions inside any company are effectively multiple choice. What kind of request is this? Which team should handle it? Is it urgent or can it wait? For questions like these, a decision model responds faster, costs less, and flags when it's not sure.
Three Models, Three Profiles
Jev (TypeSafe AI). The first of the three to launch. Offered as a cloud service and priced at $0.042 per million input tokens, less than half the cost of the Decisions API. Handles text only.
Decisions API (OpenAI). OpenAI's entry into the category. Runs on a single model, GPT-6 Luna, priced at $0.10 per million input tokens, and is the only one of the three that also processes images. Still in beta.
Laya (Convai Innovations). Open source, released under the Apache 2.0 license: runs on your own infrastructure with no per-use cost. One version covers more than a hundred languages, including Portuguese. The trade-off is that it was designed as a base model meant to be fine-tuned for specific tasks, according to its own authors, without that fine-tuning it barely outperforms random guessing. It also handles significantly shorter inputs than the other two. This is the option we're evaluating for cases where keeping everything in-house matters.
None of This Retires the LLM
Decision models don't replace LLMs. They replace LLMs in the specific situations where an LLM was being used just to make a choice. When a response actually needs to be written, a reply to a customer, a summary, a recommendation, the LLM is still the right tool.
Think of it like a front desk. Someone at reception looks at each envelope and decides which desk it goes to; at the desk, someone else reads the contents and drafts a reply. The decision model is the person at reception. The LLM is the one at the desk.
Innovation isn't always about building the most powerful model. Often, it's about using the right model in the right place.
What Comes Next
On a current project with a property management company, we ran Jev and the Decisions API side by side with the LLM we were already using, same decisions, same day. In the next post: the numbers. Speed, cost, and accuracy.
If you're thinking through where AI fits into your company's processes, and where it doesn't — we'd love to talk.
Tags: #ArtificialIntelligence #DecisionModels #LLM #OpenAI #DecisionsAPI #ProcessAutomation #EnterpriseAI #MachineLearning #GenerativeAI #Classification #NLP #TextProcessing #AIForBusiness #Automation #TechInnovation