Jev: the AI that doesn’t talk. It decides.

Walk into any planning office on a Monday morning and count the decisions. Is this order urgent? Is that supplier delay a problem or just noise? Should this item be stocked at every hub, or only centrally? Does this customer complaint need escalation?

None of these are hard on their own. A good planner answers each in seconds. The problem is that there are ten thousand of them, every week, and the planner is one person with a coffee and a full inbox.

For two years we have thrown large language models at this problem. They are brilliant, but they are also slow, expensive at scale, and occasionally creative in ways a supply chain can’t afford. Ask a chatbot to classify 50,000 item-locations and you’ll get a large bill, a long wait, and a handful of answers that don’t match any option you gave it.

On 15 September, a small San Francisco company called TypeSafe AI opened early access to Jev. It doesn’t talk. It decides. And I think supply chain people should pay attention.

Meet Jev: fast thinking, bottled

Jev comes from TypeSafe AI, founded by Diogo Almeida, one of the people behind ChatGPT and the RLHF training technique at OpenAI. TypeSafe calls Jev a System One model, a nod to Daniel Kahneman’s fast, intuitive thinking as opposed to slow, deliberate reasoning.

The idea is simple and a little radical. You don’t ask Jev to write anything. You hand it a block of information (text, JSON, an email, a record) and a set of questions whose possible answers you define in advance. Jev answers all of them at once, in roughly 70 to 500 milliseconds, and tells you how sure it is.

It knows three kinds of question:

• Choice: pick one option from a list of up to 255, with a probability for each.
• Score: place the input on an ordered scale, like low, medium, high.
• Yes/no: the probability that a statement is true.

That’s it. And that constraint is the superpower. Because Jev can only return answers you allowed, it can’t invent a category, misspell a status or hand you a paragraph when you wanted a code. TypeSafe trains it on synthetic data with a method it calls Reinforcement Learning for Calibrated Decisions. The goal is that a 90% confidence really means right about nine times out of ten.

The company claims Jev is 40 to 200 times faster and 40 to 400 times cheaper than frontier language models on these tasks. Those are TypeSafe’s own benchmarks, so treat them as a promise to test, not a fact. But even a fraction of that changes what is economically possible.

Decide, calculate, explain

Here is the mental model I’ve been using. Almost every piece of supply chain work is one of three things:

For the last two years, many of us have pushed all three jobs into a chatbot. That works in a demo. At scale, it’s like hiring a brilliant consultant to sort your mail.

Jev fills the gap in the first row. It takes the thousands of small, repetitive judgement calls off the planner’s desk, and it does so with a confidence score. The confident calls flow straight through. The uncertain ones go to a human, who now spends time only where judgement is really needed.

That’s the real promise: not replacing planners, but giving them back their Mondays.

Where Jev fits in supply chain

The sweet spot is any decision that is high volume, needs a fast answer, and has a known set of possible outcomes. Once you start looking, they are everywhere:

• Order and email triage. Classify every incoming message by intent, urgency and site, then route it to the right queue or agent.
• Exception management. Sort thousands of MRP or transport exception messages into “act now”, “watch” and “ignore”.
• Supplier risk monitoring. Feed in fresh news about each supplier and ask: has a material event happened? Get a probability, not a summary to read.
• Master data quality. Score millions of records for plausibility at a cost that makes a full scan realistic.
• Escalation and complaints. Score severity against your escalation ladder and flag customer-impacting misses automatically.
• Guarding your AI agents. Route requests to the right agent, and check that a generated answer meets your rules before it leaves the building.

But the example that got me most excited is one every supply chain leader has wrestled with: defining the stock strategy.

Worked example: a stock strategy for Upshift

Meet Upshift, a fictional bicycle maker. It builds only road and gravel bikes, each as a regular and an e-bike version, in two plants, and ships through five distribution hubs. Add frame sizes, colours and spare parts, and you quickly have thousands of item-locations.

For every one of them, someone has to decide:

• Do we make it to stock, assemble it to order, make it to order, or not stock it at all?
• Where do we hold it: at the plant, at one central hub, or at every hub?
• Which service level does it deserve?
• Which replenishment method fits?

Most companies answer this with an ABC/XYZ grid and a lot of planner judgement. The grid is fast but blind. It can’t see that a new colour launches in February, that e-bike batteries are dangerous goods with a limited shelf life, or that dealers were promised 48-hour delivery on the best sellers. The planner sees all of that, but can’t review 8,000 lines every quarter.

Step 1: give Jev the full picture

For each item-location, Python first calculates the numbers: average demand, variability, how often demand is zero, lead time, cost and margin. Then we add the soft context a planner would know. Jev receives it all as one block of state:

{
  "item": "Gravel E, frame M, colour Sand",
  "location": "Hub North",
  "weekly_demand_avg": 6.4,
  "demand_cv": 0.9,
  "zero_demand_weeks_pct": 31,
  "lead_time_weeks": 5,
  "unit_cost_eur": 2150,
  "lifecycle": "mature",
  "customer_promise": "dealer delivery 5 working days",
  "notes": "Colour Sand replaced by Olive in the 2027 range, launch February. Battery is dangerous goods, max 6 months in storage."
}

Step 2: ask the strategy questions, all at once

Step 3: read the answers and the confidence

For our Sand gravel e-bike, an illustrative answer could look like this:

Notice what happened. Jev picked up the colour change and the battery shelf life from the notes, which no ABC/XYZ grid would ever see. It is very sure about the obsolescence risk and less sure about the service level. That doubt is useful information.

Step 4: let confidence do the routing

1. Answers above your confidence threshold, say 0.85, are applied automatically.
2. Python then calculates the numbers that follow from the policy: safety stock, reorder point, lot size. Jev never does the maths.
3. Answers below the threshold land in a planner review queue. A language model writes a two-line note for each: what is ambiguous and why.
4. Every month the run repeats, and only the items whose strategy changed are flagged.

The result: the planner stops reviewing 8,000 lines and starts reviewing the few hundred that really need a human eye. The stock strategy stops being a yearly project and becomes a living, monthly process.

What Jev can’t do, and the traps to avoid

Every new tool deserves healthy scepticism. A few things to keep in mind:

• It doesn’t calculate. No forecasts, no safety stock, no optimisation. Keep your maths in code and your solvers where they are.
• It doesn’t write. No reports, no emails, no code. Pair it with a language model for the explaining.
• Stale data gives confident wrong answers. This is the big one. Jev can’t tell when its input is out of date. Feed it last year’s demand and it will be very sure, and very wrong. Log the age of your data next to every decision.
• Test the calibration yourself. Label a few hundred cases by hand, run Jev, and check whether 90% confidence really means 90% right. That’s one afternoon of work and it tells you where to set your threshold.
• It’s early days. Jev is in limited early access, pricing isn’t settled, the model is proprietary and the benchmarks are self-reported. Build the pattern so that you could swap the model later.

Start small, start this week

You don’t need a transformation programme to try this. Pick one decision your team makes hundreds of times a month. Pull 200 past examples where you know the right answer. Run them through a decision model, and compare the result with your current rules on four things: accuracy, calibration, cost and speed.

If it works, you’ve just found a way to give your planners back hours every week. If it doesn’t, you’ve learned something real in a day, instead of believing a benchmark.

For years, AI in supply chain meant forecasting or chatbots. Jev hints at a third category: small, fast, honest decisions, at a scale no team could ever handle by hand. The best supply chains are not the ones with the smartest people. They’re the ones where smart people spend their time on the decisions that matter.