Jev use cases: where a decision model fits in a business, and where it fails

A decision model like Jev fits a business task when six things hold: the possible answers are known in advance, a person could decide in a few seconds, the input is text, the volume is high, a mistake is cheap or can be caught, and no arithmetic or date logic hides inside the question. That covers routing tickets and leads, guarding AI agents, filtering search results, tagging large catalogues and turning text into features for a classical model. It does not cover anything that needs calculation, writing, or an answer that is not in the text. Used inside those lines, it decides for a fraction of a cent; used outside them, it is confidently wrong.

What is the six-point fit test?

Before any new use, I check a task against six conditions. If one fails, the decision model is the wrong tool, or at most one step next to a language model:

1. The answers are known in advance. A fixed list of options, a scale or a yes or no.
2. It is a quick judgement. Something a knowledgeable person decides in a few seconds with the right context, not a chain of reasoning.
3. The input is text, or code can turn it into text or named categories first.
4. Volume or speed matters. Many decisions a day, or decisions needed in a fraction of a second.
5. A mistake is cheap or catchable. Unsure cases can go to a person or a larger model, or the decision feeds a system that tolerates some noise.
6. No arithmetic, counting or date comparison is needed inside the question. TypeSafe's own list of weaknesses says it plainly: Jev is not a calculator, and it reads dates as text (TypeSafe documentation).

Which business tasks fit best?

• Routing tickets, emails and leads. In an independent study on synthetic support tickets, choosing the right one of four queues scored 89.0% (jev-ood-calibration). Which queue, which department, how urgent: this is the natural first job, and it pairs with the routing described in AI customer support for e-shops.
• Guarding AI agents and checking AI output. Is this tool call destructive, is this message an injection attempt, does this source support this claim. In one of TypeSafe's own worked examples, eight citations were checked and all four planted errors were caught. Where these checks sit in a system is covered in AI guardrails.
• Filtering search results before an answer is written. Asking "does this passage answer this question" for every retrieved passage costs almost nothing: in a TypeSafe legal example, 1,200 scoring calls cost 0.0645 US dollars. Other ways to improve retrieval are in advanced RAG.
• Features for a classical model. This is where the evidence is strongest. On 2,000 wine reviews, predicting the score from the average gave an error (RMSE) of 3.09, word counts 2.47, asking Jev for the score directly 2.15, and 38 small Jev questions fed as columns to a gradient-boosted model 1.77. Many narrow probabilities, combined by a model trained on real outcomes, beat one direct answer.
• Tagging at volume. By my arithmetic, tagging 50,000 e-shop products at about 300 tokens each is 15 million tokens, roughly 63 US cents on Jev, which makes tagging a whole catalogue again after every change a non-event.
• Classification with a few examples. In one independent run, given category definitions and labelled examples, Jev reached 92.40% on a standard 77-category banking set, against 93.66% for the fine-tuned model in the original paper (jev-banking77-experiment). How classifiers compare on your own email is in AI text classification.

Which tasks should stay in code or go to a larger model?

• Anything with arithmetic, counting or prices. Compute it in code and pass the result as a named category, such as "over budget", rather than a raw number.
• Date logic. Extract the parts, compare them in code.
• Questions whose answer is not in the text. Without an "unknown" option, a decision model still picks an answer, often with high confidence, as described in Jev and structured outputs.
• Writing replies, summaries or documents. A decision model cannot write. Pair it with a language model or with templates.
• Forecasting from raw numbers. Numeric data is its documented weak spot.
• Hostile input without other defences. Instructions hidden in the text can move its answers, as with any model; the defences are in prompt injection.

Which patterns make it safe in production?

1. The model judges, code acts. Code computes, filters and executes; the model only answers typed questions.
2. Ask every question in one call. TypeSafe's parallel-questions example measured one call with 13 questions at about 12 times cheaper and 10 times faster than 13 separate calls.
3. Combine narrow answers in code, with fixed weights at first and a fitted model once you have labels.
4. Route by calibrated probability and by stakes. Act, ask for confirmation or escalate, with thresholds set per question on your own labelled sample, as explained in Jev vs LLMs.
5. Offer an escape option everywhere: "unknown", "none of these", "not stated".
6. Run it in the shadow first, beside the current method, on 1,000 to 2,000 real cases. The independent studies above cost well under one dollar each: the calibration study about 6 US cents for 4,621 items.
7. Log the input, the questions, the answers, the model version and the outcome. That log becomes your test set, as covered in LLM evals.
8. Pin the model version once thresholds are tuned. An alias such as "latest" moves when a new release ships, and answers can shift with no change on your side.

What do the terms and the service mean for a business?

Three points matter before you build on it. The terms forbid using its outputs to train or distil another model (TypeSafe Master Customer Agreement, section 2.3(b)), so you cannot use Jev to label data for a classifier of your own. It is a hosted service with no version you can run yourself, and TypeSafe paused new signups on 22 September 2026 because of demand, so plan for the provider changing terms, limits or availability. And it is English-first: TypeSafe says other languages work with lower accuracy, so a Greek business tests on its own Greek examples first. How licence terms shape what you may build is covered in open model licences.

When is a model of your own better?

When you have labels and ask the same questions thousands of times. A follow-up in the phishing benchmark's repository trained a 4-billion-parameter open model on 1,000 labelled emails in 18 minutes on a gaming graphics card, and it scored 97.4% on 500 held-out emails, where Jev, asked one question, had scored 62.6% on the same dataset (jev-phishing-bench). Supervised against zero-shot is not a fair race, but it shows what labels are worth. Open alternatives exist too: an Apache 2.0 replica built on a 9-billion-parameter model scored 90.12% against Jev's 93.21% on its own test set (Bespoke Nimble).

Frequently asked questions

Can Jev run on my own server?

No. Jev is available only as a hosted service, directly or through gateways such as OpenRouter. If the data may not leave your servers, an open model you run yourself is the route, as described in running an LLM on your own server.

Can a decision model answer my customers?

Not by itself. It can decide what the message is about, how urgent it is and whether a human must see it. The reply itself comes from templates, a language model or a person.

Where do I start?

Pick one decision your team makes hundreds of times a week, write the options, label 200 real examples, and run the model in the shadow beside the current process for two weeks. The numbers from those two weeks decide the rest.

Decisions like these sit inside most of the AI agents for business that I build. If you have one in mind, send me the decision and twenty real examples, and I will tell you whether it passes the fit test.