Decision models

A decision model receives some content, such as a support ticket or photo caption, and questions with defined answer options. It returns bounded results: a category with probabilities, a score or a yes-or-no likelihood.

TypeSafe AI calls these ‘System One models’, a nod to fast, instinctive human thinking. The comparison is a metaphor, not evidence that a model thinks like a person.

‘Decision model’ has an older and broader life in operations research and decision theory, where it means any formal representation of a decision problem. The modern version is an AI model that returns bounded, typed decisions – a category, a score or a yes-or-no probability – rather than generated prose.

Why it matters for content

Many content operations depend on repeated judgements with known answers: which queue should receive a support message, whether a passage breaches a policy, how relevant a document is to a query or which taxonomy term fits an item. These jobs don’t always need a paragraph from an LLM. They need a value that software can use without first persuading a free-form answer to behave like data.

That makes the content around the decision unusually important. Someone has to define the available options, write questions that draw meaningful distinctions and choose which part of the source material the model sees. A muddled taxonomy or vague quality scale will still produce tidy numbers.

The probabilities also create a practical route for review. A system might route high-confidence decisions automatically, send uncertain ones to a person and record disagreements for evaluation. However, teams still need to test whether the confidence scores match real outcomes; a number with two decimal places can look more settled than it deserves.

How decision models change the picture

Decision models separate judgement from generation. An LLM can still draft a reply or summary, while a decision model handles bounded tasks such as classification, scoring, routing or checking. In an agentic system, that could mean using one model to choose the next permitted action, code to enforce the rules and an LLM to write whatever a person will read.

For content teams, this gives structured content another job. Take a charity case study stored as separate fields: the story, a pull quote, the photo credit and a consent record. A check on whether the quote is cleared for use needs the quote and the consent record, and nothing else. Whether you can send just those two depends on decisions made long before anyone bought an AI tool: how the content model splits things up, and whether consent was recorded as data or left in someone’s inbox.

The fixed output removes one familiar failure: the model cannot invent a category outside the options or return malformed prose where code expects a number. It can still misread the content, choose a perfectly well-formed wrong answer and attach a high probability to it. Format guarantees make systems easier to build; they don’t guarantee to supply truth.