Decision models: practical AI where it used to be too costly

At MarsBased, we like to get our hands on a new technology before we talk about it. This time, it's Jev and the new wave of decision models. Last week, I was talking to Juan Artero Martín, one of our tech leads, and we agreed that these models could represent a shift comparable to the arrival of LLMs or coding agents.

The basic idea

When you can define the possible answers to a question, you give the model those options and the relevant context. It returns a decision, together with probabilities or confidence estimates. Think of classifying a support ticket, assigning its priority or deciding whether a comment needs moderation.

Why it is interesting

For suitable tasks, this can cost a fraction of what you'd spend using a general-purpose LLM. That part isn't new. What is interesting is that you can define a task in natural language, get a structured decision your software can use, and potentially avoid training a separate model for each use case.

At MarsBased, we've typically used LLMs for tasks requiring more complex language understanding, and trained traditional machine learning models for classification, prioritisation and similar operations. Decision models give us another option. In some cases, we can deliver those capabilities without taking on a dedicated training and fine-tuning process.

Beyond Jev

This goes beyond Jev. We've also been exploring models that run on our development machines. Alternatives such as Laya and Clef make local deployment possible. And they're also expanding beyond text. OpenAI's Decisions API supports images, opening up possibilities for visual classification and moderation.

What this could mean for our clients

What excites me most is what this could make viable for our clients: useful capabilities that previously cost too much to run or required too much work to maintain. These new tools could bring AI into products where, until now, it simply didn't pay off.