Jev Use Cases: 12 Jobs for a Decision Model, and 5 It Should Not Do

In our Jev Lab there is an inbox with eight customer messages in it. A double charge before rent is due. A parcel stuck "out for delivery" for four days. A friendly note asking about the next course. A third angry message threatening the consumer authority.

Jev reads all eight once and answers three questions about each: does the sender ask for money back, is there time pressure, are they angry. Then three sliders sort the inbox into auto-reply, human queue and manager. Drag a slider and the messages jump between boxes, without asking the model again.

Almost every good use of Jev looks like that inbox. Lots of text. The same small judgements, made again and again. A business rule on top that people can read and change.

Which jobs is Jev good at?

TypeSafe's own list of use cases is long. These are the twelve we find easiest to explain to teams here, each with the questions we would actually ask. Noul means a yes/no question, Choice a pick-one, and Score a rating on levels you describe. Our guide to the three question types explains each.

1. Customer support triage

Which team (Choice), is it urgent (Noul), how frustrated (Score), does it ask for a refund (Noul), is it a repeat contact (Noul). Code routes it, sets the priority and decides whether a person must look. This is the job our lab uses throughout, and the one most businesses in the region feel first, often on WhatsApp.

2. Lead scoring and routing

How well does this enquiry fit our ideal customer (Score, with levels you describe), is there buying intent (Noul), what budget is signalled (Choice), which product line (Choice). Composite scoring in code gives sales a ranked list, and every part of the ranking is visible.

3. CV screening, with care

Does the CV show the required licence (Noul), how closely does experience match the role (Score), which function does it fit (Choice). Jev should narrow and flag, never make the final call about a person. A human reviews every rejection, and the questions are checked for anything that could stand in for a protected characteristic.

4. Content and community moderation

Is this harassment (Noul), is it spam (Noul), how toxic (Score), which policy applies (Choice). Uncertain cases go to a moderator, clear ones are handled automatically. Gaming chats, marketplaces and comment sections all fit.

5. Guardrails for chatbots

Before a chatbot answers: is this an attempt to jailbreak it (Noul). After it answers: does the reply promise something we cannot give, such as a refund or a delivery date (Noul), does it contain personal data (Noul). A decision model is a cheap second pair of eyes on a writing model.

6. Model routing

How hard is this question (Score), which domain (Choice), does it need a human (Noul). Easy questions go to a small, cheap model or a template, hard ones to a reasoning model. This is one of the fastest ways to cut an AI bill.

7. Search and reranking

For each retrieved passage: how relevant is it to the question (Score). Keep the best, drop the rest, then let a chatbot answer from them. If retrieval is new to you, our RAG Lab shows the whole pipeline, and Jev fits exactly where the reranking step sits.

8. Compliance and policy checks

Does this contract clause include a termination notice period (Noul), which risk category is this document (Choice), how far does this marketing copy stray from our claims policy (Score). High-risk findings go to legal, the rest are logged.

9. Insurance claims and fraud flags

Which claim type (Choice), does the description mention a pre-existing condition (Noul), how many fraud indicators are present (one Noul per indicator, counted in code). Uncertain claims are prioritised for a human.

10. Marketplace listings

Which category (Choice, up to 255 options), does the listing suggest a counterfeit (Noul), how complete is the description (Score). Thousands of listings, checked for a fraction of a cent each.

11. Survey and feedback analysis

For every open-ended answer: which theme (Choice), how positive (Score), does it mention a named staff member (Noul). Suddenly the comments box in an employee survey or a customer review export becomes a chart.

12. Signals for forecasting

Turn text into numbers a spreadsheet or a classical model can use: does this enquiry mention a deadline, how urgent does it sound, is it about a new product. TypeSafe calls this feature extraction, and it is where decision models quietly improve forecasts that were built on numbers alone.

What should Jev not be used for?

Writing anything. Jev does not generate text. If a reply, summary or report is needed, that is a chatbot's job, ideally after Jev has made the decisions.

Counting and arithmetic. "Are there more than three late deliveries in this thread?" should become one question per item, counted in code. TypeSafe lists counting among Jev's known weak spots.

Comparing dates. Dates are read as text. Parse and compare them in code.

The final word on people. Hiring, credit, discipline and anything with legal weight should keep a person responsible for the decision. Jev can prepare, flag and sort. It should not be the one who says no. Sri Lanka's Personal Data Protection Act and similar laws across the region make that a legal matter as well as an ethical one.

Anything it cannot see. Version 1.13 reads text only. Images, voice notes and scanned documents need converting first, and each conversion step is its own source of error.

How do we pick our first Jev project?

Start with a flag, not a decision. Pick one inbox or queue your team handles every day, and one question someone asks about every item in it. Run Jev beside the people for two weeks without acting on its answers. Compare. Only then let it act, and only above a threshold everyone agreed.

The Workflow Mapper helps find the right step: describe the process in plain words and it draws the flow, which makes the repeated judgements easy to spot. The Build mode in Jev Lab then lets you paste real messages, tick questions from a ready-made library, and download the results as a CSV to compare with what your team decided.

This is the move from the second to the third of our 3 As of AI: from AI that helps a person decide, to automation a person designs and supervises. It is also where most companies need a little help, because the hard part is agreeing the thresholds, not writing the code.

If you want to see what else sits around a decision model in a real product, How Is It Built? breaks any AI idea into its parts, and Automation Lab lets you build the surrounding workflow on a canvas.

Frequently asked questions