Agents Lookup

Learn / Decision models

Decision models: the cheap, fast step inside an agent

Most steps in the recipes here are not writing. They are choices: is this email a refund request, is this lead a fit, can this order ship. A decision model answers those closed questions in milliseconds for a fraction of a cent, and leaves the writing to a language model.

Example modelJev, from TypeSafe AI
ReturnsA choice, a level or a probability
Does notWrite text

What it is

Jev is a hosted model that takes some text plus a typed question (yes or no, pick one of these, score from 1 to 5) and returns an answer with a confidence score. It never produces prose, which is why it is fast and cheap. The community directory Made with Jev catalogues more than 800 public builds, and most of them use it for the same few jobs.

The pattern those builds share fits the way every recipe here is built: a language model writes, the decision model decides, code acts, and a person signs off on anything customer-facing, financial or hard to undo.

Where it fits in these recipes

The jobA published buildAuthor's figureRecipe here
Triage an inbox into buckets500 emails classified (Riley Brown)3.5 cents totalSupport ticket triage
Score leads and outreach before sending700 leads scored (Romàn)40 s, $0.09Inbound lead qualification
Screen for fraud, escalate the unsure ones100 emails, low-confidence cases sent to a larger model (Hassan)96/100 correct, about $0.07Order-exception triage
Classify a stack of documents26-sheet plan set (Trinay Hari)2.9 s, $0.0052Receipt categorization
Break down competitor ads724 ads from 37 brands (Matthew Berman)about 40 s, $0.09Content repurposing
Pick internal links for SEOInternal linking tool (Ian Nuttall)Up to 500 pagesNo recipe yet
Ask a question of every row in a databasejev() for PostgreSQL (Zachi)129 rows in about 1 sWeekly KPI digest
Check an agent's action before it runsAgent safety monitor (Nick Ma)Caught most test attacks, few false blocksAny recipe with a send or pay step

Every figure is the one its author published on the linked page. We have not re-run them.

New: decisions about images

Jev 1.14, released on 14 October 2026, added image input (PNG or JPEG) and a "Box" answer that points at one region of an image, such as the signature line on a scanned form. Images are billed as ordinary input tokens. TypeSafe AI labels both features as preview, not production-ready, so test them on your own sample images first.

That opens up checks that used to need a person or a large model to look at every photo: sorting receipt photos, checking product photos against marketplace image rules, flagging damage in return photos, confirming proof-of-delivery photos, and checking that a form is signed.

See all decision use cases

Is your step a good fit?

  1. The answer is a fixed set of options

    If you can list the possible answers before the call (refund, exchange, question, spam), it fits. If the step needs a written reply, keep it on a language model.

  2. The same decision repeats

    The savings are per item. Hundreds of emails, orders or leads a week add up; one decision a day does not.

  3. A wrong answer can be caught

    Act on the confident answers and send the rest to a person or a larger model. Start with actions you can reverse.

Caveats

Want a decision step added to a recipe?

Setup help covers wiring a triage or scoring step into any recipe in the directory, with the confidence threshold and fallback set up for you.

Request a quote Browse recipes