Decision models and Jev: quicker, cheaper answers from AI

The short answer: a decision model returns a typed judgment — a yes/no probability, a label or a score — instead of writing text. Putting one such as Jev in front of a large language model lets it rule out irrelevant material and route routine cases in bulk, so the expensive model reads only what matters. The result is faster answers and lower AI bills.

Two kinds of thinking

Psychologists describe two modes of thought: fast, intuitive System One judgments and slow, deliberate System Two reasoning. AI systems benefit from the same split. A large language model is a superb System Two — it can reason, plan and write — but using it for every quick decision is like asking your most senior expert to sort the mail.

Jev, from TypeSafe, is built as a System One model. It cannot write text or reason in steps. What it does is answer precisely framed questions — “Does this email ask for a refund?”, “Which department does this ticket belong to?”, “How relevant is this document to the question?” — in about a second each, many in parallel.

Three kinds of decision

• Yes / no

  A probability that a statement is true: is this an invoice, is this line an error, is this lead in our service area?

• Choose a label

  Pick one category from a list: bug, feature request or question; sales, support or billing.

• Score

  Rate on a defined scale: urgency, relevance, sentiment, fit.

The funnel: narrow first, think second

1. Collect everything. Hundreds of emails, files, tickets, log lines or search results.
2. Decide in bulk. The decision model answers one sharp question for each item and drops what doesn’t qualify.
3. Check the borderline. Items with uncertain scores are flagged for closer review instead of being guessed.
4. Reason on what survives. The large language model reads only the shortlist, with a smaller, cleaner context window.

Every item ruled out in step two is text the large model never has to read — and never has to be paid for.

Why it is quicker and cheaper

• Smaller jobs, smaller model. A typed judgment needs far less computation than generating paragraphs of text.
• Parallel by design. Hundreds of independent decisions run at the same time rather than one after another.
• Less for the large model to read. Fewer tokens in context means lower cost, faster responses and often better focus.
• Measurable confidence. A probability tells you when to trust the answer and when to send it to a person.

Where decision models shine — and where they don’t

Clear, literal questions get the best results: one judgment per question, with boundary cases spelled out. Subjective or expert-defined categories are better treated as a shortlist for a person or a larger model to confirm.

Putting it to work

In the agentic systems SEGULA LABS builds, decision models handle the high-volume triage and large models handle the thinking. We measure both through the AI engineering loop, so each step uses the fastest, most affordable model that meets the quality bar. Want to see where that could cut time and cost in your workflows? Talk to us about AI automation.