A white name tag reading My Name is Jev floats among grainy pink clouds against pale blue sky.
Jev is one of the decision models explored in this article.Visual reference: TypeSafe.

Breakdown / Decision models · AI explained

Not every AI needs to chat. Some just need to choose.

Sometimes you need AI to sort an email or pick an answer. Here's how decision models work, and what the different options offer.

Think “hot dog or not.” You describe what counts as a hot dog, give the AI something to check, and ask it to choose. You could do the same with your inbox: tell it what belongs in each folder, then let it sort the messages.

That's the idea behind decision models. They handle small questions like whether an email is a bill, which folder it belongs in or how urgently it needs your attention. Sorting things into categories is called classification. They can also give ratings or assess yes/no questions.

You describe what the answers mean

For our inbox example, let's give it three folders:

  • Bills: invoices and requests to pay for a service.
  • Invitations: someone is asking you to attend something.
  • Check myself: the email is unclear, fits both folders or belongs in neither.

“Your electricity bill is ready” should go into Bills. “Come over for dinner on Saturday” should go into Invitations. We're making up these examples to explain the task; we haven't run them through the models.

You give the model an email and those descriptions. It picks a folder, and an email tool connected to it could move the message there. Picking the folder and moving the email are separate jobs.

Then an email says, “You're invited to our workshop. Please pay the registration fee by Friday.” It's both an invitation and a request to pay, so our rules put it in Check myself. If you'd rather put it in Invitations, change the rule. A chatbot can help write these descriptions, but it can't decide your preferences for you.

What changes when the model only needs to choose

A regular chatbot can already sort messages if you tell it how. You can also make it answer in a fixed format that other software can read. This ability didn't start with Jev.

The difference is the focus. Models like Jev, Strands Decider and Laya return choices and scores instead of writing a reply. That narrower job can mean less work for the model on each request. Whether it's faster, cheaper or more accurate still depends on what you're asking it to do.

You can describe a task and try it without first collecting examples to train your own model. Someone already trained the model you're using. Your descriptions tell it what you want; they don't guarantee it'll get the answers right.

The options

Jev, Strands Decider and Laya work with text. They'd read a menu description for “hot dog or not.” OpenAI says its Decisions API also supports images. The other big difference is where they run: some are services you send questions to, while others can run on your own computer.

Jev from TypeSafe

Jev runs on TypeSafe's service. Software sends it the material to check, along with questions and descriptions of the possible answers. It can pick a category, give a rating or estimate the probability that the answer to a question is yes.

TypeSafe says everyone uses the same underlying model. You customize the instructions, rather than train a separate Jev for your inbox. The current version, Jev 1.13, accepts text and text records, but not photos, audio or video. Jev models, decision types.

Strands Decider from Amazon

Amazon's Strands Labs released Strands Decider 2B on October 1. It's open source: you can download the trained model and run it on your own computer. Amazon also released the code, training data and scripts used to build it.

It handles choices, ratings and yes/no questions. Amazon suggests using it to decide where requests should go or check what an AI agent is about to do. The team also says it's weaker at complex reasoning and isn't suited to writing code, summaries or chat replies. Announcement, repository.

Decisions API from OpenAI

OpenAI showed Decisions API in preview at DevDay. An API is a way for software to request something from another service. Tibo Sottiaux says this one uses Luna to make choices from a restricted set of answers and supports images too.

It belongs here as an announced option. We haven't verified its pricing, who can use it or exactly what it returns. Covering the same kind of task doesn't tell us whether it works the same way as Jev or how their results compare. Those are questions for documentation and testing. Sottiaux's announcement.

Laya

Laya is another open-source option, released under the Apache license. You can download its models, run them yourself and train them further for your task. It also supports choices, ratings and yes/no decisions.

Its maintainers are clear about a catch. In their tests of choices and scores, the models did poorly before extra training for those tasks. Better results came after that training. They also document limits on confidence and how much text the models can handle. You can try your question straight away, but useful answers may take more work. Laya's documentation and limits.

The answer still needs checking

Returning one of your three folder names makes the answer easy for software to use. It doesn't stop the model putting a bill in Invitations.

Some tools include probabilities or confidence scores with their answers. Before trusting a high score, check whether it actually goes with correct answers on your messages.

We haven't tested these options against one another. A fair comparison would use the same messages and rules, then check the answers, mistakes, response time and total cost. A model you run yourself avoids a service's per-request bill, but still needs computing resources and upkeep.

For that first inbox test, include more than the easy bills and dinner invitations. Add the workshop email, a vague message and something that doesn't belong in either folder. See what happens when the answer isn't obvious before letting it organize the rest.

Sources