A new kind of AI called decision models is spreading fast in the systems businesses run. While conversational assistants are still the dominant use case, decision models are gaining real momentum by giving an answer in a fixed, reliable form, something chatbots cannot do. These models are designed to sit inside larger systems as a reliable component, short answers instead of a paragraph of free text, which makes them genuinely useful. Though a reliable answer to one question is not the same as a decision you can explain or defend, and for anyone automating real work in a regulated setting, that gap is where the difficulty and the value lie. They are genuinely useful, and also easy to misread, because a reliable answer to one question is not the same thing as a decision you can stand behind.

Having a decision come out the same way every time, carry the rule behind it, and be acted on without hesitation is the type of challenge we work on at UnlikelyAI. Which is why a new class of models built for reliable answers is worth a closer look. In this blog we explore where decision models come from, and why they are surfacing now.

Where they came from

The lineage is worth tracing. BERT, released by Google in 2018, set the pattern: a base model that others could specialise to narrow tasks. BERT was hugely popular, but the community moved to framing all problems as text generation, e.g. with the GPT-3 paper ‘Language Models are Few-Shot Learners’ and others.

So when an independent researcher trained a BERT-style architecture on simple classification and extraction tasks, using new techniques to improve its calibration, they were really breathing fresh life into an existing and well-trodden path. A new Silicon Valley startup doing the same, disguised as a revolution, is a far bolder attempt to eclipse such recent history. This is the context behind Typesafe AI’s release of Jev, a closed model positioned explicitly as an alternative to general-purpose LLMs, with little discussion of the fact it was functionally a reboot of previous-generation technology. Amid all this noise, that independent researcher made the apparently reasonable complaint that there was nothing novel about it and released their work openly as Laya. OpenAI has since announced a decisions API, now in beta, covering much the same ground, and several of these tools now take images as well as text. A landslide of such models are now landing from a variety of the big names, all with a shared vision of getting a model to return a single constrained answer rather than a paragraph of plausible text.

What a decision model is

A decision model takes a question with a fixed set of allowed outcomes and returns exactly one of them, along with a probability you can actually use. You might ask which team an email should go to, finance, accounts or tech support, and you are guaranteed one of those three back, not a paragraph that begins "Well, I can see that... so I would recommend... unless...". The same shape covers classifying something into a set of categories, or scoring it on a scale, such as the severity of an incident.

An LLM asked the same question gives free text: unstructured output, no real confidence number, and a slower, more expensive call. A decision model gives a single prediction in a fixed format, fast and cheaply, which is exactly what you need when the output feeds the next step of an automated system rather than a human reading it.

None of this is new in spirit. It descends from earlier natural-language work: classification, sentiment analysis, retrieval. In fact, it’s even possible with constrained decoding to guarantee an LLM’s output will comply with a formal specification (although this does not change the underlying difference in the models and is really just an inference-time stand-in for what decision models were designed and trained for). What has changed, however, is the zeitgeist: the true achievement of generative AI has been to convince people that AI has arrived and is smart enough to mean business. Where previously you’d need a team of researchers to train your own model, which you had to fight to convince your boss would ever be worth it, now these capabilities (and belief in them) has been democratised and productised. Laya, built from independent research on calibrating a BERT-style architecture, is open source and open weight. Jev, from a new startup, is closed and gives little away.

Where this is going

After years of pushing models to generate more, a serious part of the field is now chasing reliability in the small components that do one thing dependably enough to run unattended.

The argument the new entrants make is a reliability-first one: that you only build software on a dependency you can trust to behave, and real automation needs parts you can inspect, test and constrain. It is a case for AI as infrastructure rather than AI as a conversation, and it is gaining ground.

What they do well, and what they leave to you

A decision model is very good at giving a fast, cheap, reliable answer to a single well-defined question, with a confidence you can act on in code. The limitation is scope, an atomic question is not a whole decision: "Is this clause satisfied?" is answerable in one shot. "Is this insurance claim covered under this policy?" is not; it is many such questions, combined under a structure that decides how their answers add up.

These models are reliable but are not able to provide explainability. They are statistical models that run once and return a prediction, with no convincing insight into their working or causal justification for their result. For a single atomic question that is perfectly fine, a yes or no does not need an essay. A decision model makes one small decision well, but what about the bigger, more consequential decisions, the ones a business actually worries about?

Those are a different kind of problem. They are compound, built from many conditions that have to be combined under rules rather than settled in one pass. When it comes to a regulated decision, it needs to be explained and defended: a lone classification, however confident, cannot do that by itself. The model is a black box, which is acceptable only because each one is doing something tiny.

Where this fits in what we do

Our work is closely related to decision models. In our case it’s integrated in the whole structure around it. We turn rules, policies and legislation into actionable decisions for firms that need a clear verdict, not a long, LLMy answer. The difference is where explainability is, and in our case it is built into the structure around the model.

The structure is what lets us answer "is this claim covered?" and show the clauses relied on and the leaf answers that led to the verdict. A reliable leaf is necessary; an explainable whole decision is the point.

We turn a policy or set of rules into a structure, then evaluate its simple, atomic tasks and propagate the answers through the system to produce a single overall verdict. Those atomic tasks are the same shape as what a decision model does: small models, atomic questions, and they can run on anything from small language models to architectures like BERT. Laya looks similar to models we have experimented with for exactly this, and we plan to benchmark a variety of open-weight models this quarter.

These tools are shipping as a product that we treat as one ingredient, and the industry moving toward reliable, constrained components is the direction we’ve been pushing the industry in for years. It also suggests the right question to ask when you evaluate a decisioning tool for regulated work. Not only "how accurate is the classifier?", but "what structure turns its answers into a decision I can explain?" For anything that has to hold up to a regulator, that second question is the one that matters.

Want more from the UnlikelyAI Lab? Follow us on LinkedIn, or join our newsletter for research on trustworthy, explainable AI.