Eight cents. That is what one early user reported paying to classify 1,018 research papers with Jev, in the first 48 hours after launch. Writing summaries of the same papers with a language model cost $3.99.
That gap is the whole story. Jev isn’t built to write. It’s built to decide, and a supply chain runs on small decisions.
In my first post on Jev I explained the idea and walked through one worked example. This time there are no exercises. Just ten jobs where I think it earns its keep, explained plainly enough that you can spot the eleventh in your own team.
Jev in one minute
Jev comes from TypeSafe AI, a company founded by former OpenAI researcher Diogo Almeida. It was announced on 15 September 2026 and is in early access through a waitlist. TypeSafe calls it a System One model, after Daniel Kahneman’s fast, intuitive thinking.
You give Jev a block of information, such as an email, a record or a status message. You also give it questions, and for each question you define the possible answers. Jev answers all the questions in one pass and says how sure it is. There are three kinds of question:
- Choice: pick one option from a list of up to 255.
- Score: place the input on a scale, such as low, medium, high.
- Yes/no: the probability that a statement is true.
Because it can only return answers you allowed, it can’t invent a category or hand you a paragraph instead of a code. TypeSafe reports answers in 70 to 500 milliseconds and input at $0.042 per million tokens, with output free. On its own benchmark, Jev scores 67.8% against 67.9% to 74.1% for frontier language models. So roughly on par, at a fraction of the time and cost. All of these figures are self-reported.
Jev doesn’t write, doesn’t calculate and doesn’t read images yet. That makes the recipe for a good Jev job simple. The decision is made many times a day, the answers come from a fixed list, and a doubtful answer can wait for a person. All ten jobs below pass those three tests.
1. Reading supplier order confirmations
Buyers send hundreds of purchase orders a week. The confirmations come back as emails, portal messages and PDFs, in every format and language. Someone reads each one and updates the ERP by hand.
Jev reads the purchase order line next to the supplier’s reply and answers one question: what did the supplier confirm? The options are “as requested”, “new date”, “new quantity”, “partial”, “rejected” and “unclear”. When the answer is “new date”, a parser pulls out the date and code compares it with the date you need. Jev is weak at date arithmetic, so don’t ask it to do that part.
Confirmations that match the order go straight into the ERP. The buyer only sees the changes, the rejections and the unclear replies.
2. Sorting order change requests
Customers change orders all day. Move the delivery, add two units, change the address, cancel the lot. Each request lands in a shared inbox and waits for the order desk to read it.
Jev classifies the kind of change, scores the urgency and answers a few yes/no questions: is this a cancellation, does it mention a new delivery address, is the customer unhappy? Code then checks the facts Jev can’t see, such as whether the order has already been picked.
Simple changes on orders that haven’t moved yet can be handled automatically, with a confirmation back to the customer. Everything else reaches the order desk already sorted and ranked.
3. Weighing market intelligence for the forecast
Sales teams send a steady stream of notes. A dealer plans a spring promotion. A competitor cut its prices. A big customer mentions a new store. Demand planners have to decide which of these should move the forecast, and most notes get skimmed or lost.
For each note, Jev answers: does this describe a change in demand for one of our product families? If yes, in which direction, how large (small, medium, large) and when (this month, next quarter, later or not stated)? That “not stated” option matters. Without it, Jev has to guess.
The planner gets a short, ranked list of notes worth discussing in the demand review. Jev never touches the forecast number. That stays with the planner and the statistical model.
4. Judging which shipment delays matter
Carriers send thousands of status updates: delayed, held at customs, missed pickup, damaged. Most of them don’t affect a customer. A few of them will cost you one if nobody acts.
Code first adds the facts to each event: the promised date, the customer and whether stock at the destination covers the gap. Jev then answers three questions. What kind of event is this? Is a customer delivery at risk? And what should happen: nothing, inform the customer, expedite, rebook or escalate?
Here the confidence score earns its place. Informing a customer is cheap, so it can go out automatically at a modest confidence. Expediting costs money, so it needs a higher bar or a person to confirm.
5. Deciding what happens to a return
Every returned item needs a disposition. Restock as new, restock as second grade, refurbish, send back to the supplier under warranty, recycle or scrap. The decision usually waits on a person reading the inspection notes, and the item waits with it.
Jev reads the customer’s return reason and the inspector’s notes and picks the disposition. It also answers two useful yes/no questions. Does the inspection confirm the customer’s reason? Does the note describe a possible product defect?
Items move out of the returns area faster. And the defect flags, counted by code over a month, give engineering an early signal that nobody had time to collect before.
6. Filling in new item master data
A new item needs dozens of attributes chosen from lists. Product group, storage class, dangerous goods or not, temperature control, handling unit, planning family. The information is usually there, buried in the description and the supplier’s datasheet.
This is almost a textbook Jev job. Most attributes are a choice from a fixed list, and Jev can answer all of them in one call. Adding the tenth question costs a little more input but almost no extra time.
The master data team then reviews only the fields where Jev was unsure. Clean data at the start is far cheaper than cleaning it later.
7. Coding spend for your carbon footprint
Many companies estimate the carbon footprint of what they buy from spend. Each purchase line is mapped to a category, and each category has an emission factor. The hard part is the mapping: thousands of free-text lines, written by hundreds of people.
Jev picks the category for each line, and code multiplies the spend by the factor. If your list has more than 255 categories, split the question in two: first the family, then the category within it. Lines with low confidence go to the sustainability analyst.
The same mapping gives procurement a spend cube they can trust. Two reports from one pass over the data.
8. Pre-sorting products for customs review
Tariff classification needs an expert. There are thousands of codes, the rules are legal and the company carries the responsibility. Jev should not have the last word here, and its 255-option limit means it couldn’t pick among all the codes anyway.
It can do the pre-sorting. Has this product’s description or composition changed materially since it was last classified? Which chapter of the tariff does it most likely belong to, so it reaches the right specialist? Does the description mention properties that call for an export control check?
The customs specialist still makes every decision. But their time goes to the items that changed or carry a risk, not to reading unchanged descriptions.
9. Reading warehouse incident reports
Shift leaders write short reports every day: a near miss at the dock, damaged pallets, a forklift fault, a stock discrepancy. The reports are filed, and the patterns in them stay invisible.
For each report, Jev picks the incident type, scores the severity and names the most likely cause area: people, process, equipment, packaging, supplier or unknown. One yes/no question matters most: is this safety-related?
Safety-related reports with a high severity go to the safety manager at once, even at modest confidence, because a false alarm is cheap. Code counts the rest by month. For the first time, the warehouse manager sees trends instead of a pile of forms.
10. Preparing the S&OP agenda
Before every S&OP meeting, dozens of issues come in from demand review, supply review, finance and product management. The team preparing the meeting has to decide which ones need an executive decision. Many are duplicates of each other.
For each issue, Jev answers: does this need a decision at executive level? Which area is it (demand, supply, finance, new products)? How large is the impact? And, choosing from the list of open issue titles, is it already covered by another issue?
Code orders the result into a draft agenda, and a language model writes the one-page brief. The team spends its time on the trade-offs instead of on sorting.
The ten at a glance
| Job | Process | What Jev decides | Who handles the doubt |
|---|---|---|---|
| 1 | Procurement | What the supplier confirmed | Buyer |
| 2 | Customer service | Type and urgency of a change | Order desk |
| 3 | Demand planning | Whether a note should move the forecast | Demand planner |
| 4 | Transport | Whether a delay puts a customer at risk | Transport planner |
| 5 | Returns | Disposition of a returned item | Returns inspector |
| 6 | Master data | Attributes of a new item | Master data team |
| 7 | Sustainability | Spend category of a purchase line | Sustainability analyst |
| 8 | Trade compliance | Which items need a customs review | Customs specialist |
| 9 | Warehousing | Type, severity and cause of an incident | Safety manager |
| 10 | S&OP | Which issues need an executive decision | S&OP lead |
Notice the pattern. Jev decides, code does the maths and the lookups, and a person handles the doubt. The confidence score is what connects them.
What to watch out for
- It reads instructions literally. Negations and vague options confuse it. Write the options as you would for a new colleague on their first day, and always offer an “unclear” or “not stated” answer.
- No maths and no dates. Counting, totals and comparing dates stay in code. Ask Jev one clear question per item instead.
- Emails are untrusted input. Someone can write text designed to push a decision. Keep actions that move money or goods behind a higher threshold or a person.
- The benchmarks are TypeSafe’s own. Its evaluation used answers from two frontier models as the reference, not labels checked by people. Test on a few hundred of your own cases before you believe any number.
- It’s early days. Jev is proprietary, in limited access, and prices may change. Once you tune your thresholds, pin the model version, because a silent update can shift the answers.
Which one first?
Pick the job where your team reads the most text to choose from a fixed list. For many teams that will be supplier confirmations or order changes. Pull 200 past cases where you know the right answer, run them through Jev and compare the result with what your team decided.
Most of a planner’s week isn’t planning. It’s sorting. A tool that does the sorting well, and admits when it isn’t sure, gives that week back to the planning.
Which of these ten would you try first, and what is the eleventh on your list?
