Blog

Maria Kuroedova

Millions of decisions sit inside enterprises where the same calls are being made again and again. Each one of those decisions eats into someone's day, and most of them are still made by hand. A new category of AI is emerging fast to change that: decision models. Compared to LLMs, which generate fluent free text by predicting the most probable words, decision models return one short, clear answer from a fixed set of options, generally the same one every time.
Are these models enough to automate a decision a business can defend? This piece covers what decision models are, where they are limited, how UnlikelyAI automates decisions that follow written rules, and the use cases where our technology applies.
What are decision models and their limits
A decision model is purpose-built, to deliver structured outputs, unlike a general-purpose LLM that can be pointed at any problem at any level of complexity. A decision model takes a question with a fixed set of allowed answers and returns exactly one of them, an output that can be confidently plugged into a process. It works well on questions such as which team should handle a case, instantly giving one of "finance", "accounts" or "tech support", instead of a paragraph of text.
When questions are granular and well-defined these models work well. The problem arises when decisions carry ambiguity, or complexity, or are made up of many smaller answers. When the answer is not straightforward an explanation is required, and because they are statistical models they cannot help with providing reasoning.
"Is this payment suspicious?" is really many smaller questions, combined under rules that decide how the answers add up. A decision model answers one of them well, but assembling answers into a single verdict is the harder problem, and one weak or unexplained link puts the whole decision at risk.
Why are decision models landing in the market now
The idea of decision models is not new. It descends from a long line of classification models; BERT, released by Google in 2018, is a familiar example, a model class quite unlike an LLM and well established rather than novel. So why are they showing up more and more now?
In the Bank of England and FCA's 2026 survey, 86% of financial services firms reported using AI, yet even counting planned rollouts, no more than 9% expected to let it approve or decline a customer outcome, execute a trade, or move money. Adoption is spreading fast; autonomy over real decisions is very low. The capability is there; the trust is not.
What firms want is a decision they can act on fast, which a long generated answer cannot provide. That demand has brought decision models into the spotlight through a recent wave of releases: Jev, Laya, OpenAI's decisions API and others.
What UnlikelyAI does
What we have built is closely related to decision models: a decision layer that sits beside the AI stack and makes decisions with precision and explainability. Our Platform turns a policy or set of rules into a structure, evaluates its atomic tasks, and propagates the answers through the structure to a single overall verdict. Those tasks are the same shape as what a decision model does: small models answering atomic questions, running on anything from small language models to architectures like BERT.
We also change where the logic comes from. Rather than experts hand-building and maintaining thousands of rules, we turn the policy, ruleset or regulation itself into the decision structure: the document is the logic. The verdict comes with the rule that applied and the facts it rested on shown in full, deterministic and explainable all the way through, including the judgement-heavy parts where everything else falls back to a black box.
The decisions the Platform can automate
Any decision that follows written rules can become automated. A few examples:
Economic crime flagging. Does a conversation, transcript or document meet a defined economic crime risk category, such as money laundering, bribery or fraud? Every case is tested against your conditions the same way every time, each condition is shown as triggered or not, and every flag carries its category decision and the reasoning behind it.
Vulnerable customers. Does an interaction show vulnerability, and was it handled the way Consumer Duty and your policy require? The platform checks every exchange against the signals your policy names, and builds the Consumer Duty record as the conversation happens rather than reconstructing it afterwards.
Financial promotions. Does a piece of marketing copy meet COBS, PERG and your internal standards, and if not, which rule does it break? The platform returns one of three outcomes, approve, reject (with the relevant rule cited), or pass to a compliance reviewer, turning a review bottleneck into a same-day workflow.
And any decision governed by written rules. Onboarding completeness, complaint routing, disclosure checks, lending eligibility. If it is made again and again, governed by policy, and has to be justified afterwards, it fits.
In production
UnCheck, our accountancy and audit solution, is in production with a global accountancy firm. It shows how decisions can be automated in practice: confirming everything a standard requires to be disclosed, completing tick-and-tie, checking that every figure agrees with every other place it appears. It gives accountants their time back for the work that genuinely needs judgement. The same approach fits any decision governed by written rules, which is most of the decisions that run in business workflows.
Have a decision worth automating?
The decisions that cost firms the most are the ones a reliable answer alone cannot settle: the compound, judgement-heavy, defensible ones at the heart of regulated work. Those are exactly the ones we built for.
Bring us one. Book a demo, or tell us about a decision you need to automate.




