Jev in use: where decision models are already making calls today

From experiment to infrastructure in four weeks

Jev is an AI model that does not write text but makes decisions. You give it a situation and a question with fixed answer options, and it replies with a choice and a figure for how confident it is. How exactly this works is explained in our article “What is Jev?”. This one is about how quickly and for what it is now being used.

The pace is unusual. TypeSafe AI introduced Jev on 15 September 2026. On 9 October, the start-up announced a funding round of 870 million US dollars at a valuation of 7.5 billion, led by Andreessen Horowitz. According to TechCrunch, TypeSafe backs this with the claim that a third of Fortune 500 companies already use Jev. According to The New Stack, Andreessen Horowitz itself speaks of 25 percent. TypeSafe does not name the companies, and only a few cases are publicly documented.

Usage via OpenRouter, a platform through which developers access many AI models with a single account, is easier to measure. There, Jev reached around 1.08 billion requests in the week to 14 October. In second place is Perplexity's Decider with 43.2 million. In the week starting 5 October, TypeSafe accounted for 92.4 percent of all decision requests according to OpenRouter. These figures only show what runs through OpenRouter, though. Direct contracts with TypeSafe or self-hosted models do not appear there.

Workday: an AI assistant that builds its own screens

The best-known example so far comes from the software company Workday, whose HR, finance and IT programs run in many large companies. On 13 October, Workday made its AI assistant Sana the new home screen. According to The New Stack, Sana builds each view live while you type, and Jev decides which building blocks appear.

If you ask about your payslip, for example, Sana shows a card on which the amount only becomes visible after a click. For a holiday request, a date picker appears, and Sana fills in the rest of the form. Instead of having a language model generate an interface word by word, Jev picks from ready-made components in Workday's design system. Sana founder Joel Hellermark said, in essence, that a classification model that only chooses from existing components is practically instant, while generating whole interfaces is slow.

This shows an important advantage of such models: they can only choose from permitted options. Developer Andrew King put it this way: a shape nobody anticipated cannot appear on screen. In his tests with sample data, the loading time of an automatically generated dashboard fell from 29 to 3.8 seconds. Anything Jev answers with less than 50 percent confidence is passed on to a conventional language model.

Workday has now listed TypeSafe as a subprocessor, for all product lines, not just Sana. According to TypeSafe, Jev only runs through its own interface and processes data in the US. Enterprise customers can agree that no data is retained.

Jack & Jill: pre-sorting applicants at a fraction of the cost

The only case study TypeSafe has published so far comes from Jack & Jill. The start-up matches applicants and companies via two AI agents: Jack advises professionals, Jill the hiring companies. For every open role, the system pulls around 6,000 candidates from its database every day. A low-cost language model, Google's Gemini 3.1 Flash Lite, previously picked the 2,000 most promising from them.

According to TypeSafe, Jack & Jill switched this step entirely to Jev within ten days. The cost per 1,000 candidates scored fell from 75.5 to 9.2 US cents, a drop of 88 percent. Scoring all candidates for a role took a median of 10.3 instead of 20.3 seconds. Quality stayed the same: Jev kept 94.6 percent of the candidates on the list whom hiring managers later actually wanted to meet, Gemini 93.9 percent. According to the case study, the difference is not statistically significant.

Jack & Jill estimates the savings at 265,000 US dollars a year at current volume and 500,000 US dollars with expected growth. According to TypeSafe, the start-up now uses Jev in more than 15 workflows, for example to show new applicants matching roles straight after sign-up.

One caveat: this is a manufacturer case study, based on a test with 150 real roles. It shows what is possible, not what happens on average. And it raises a question that goes beyond technology: here, an AI helps decide which people even appear in front of an employer.

The typical uses: many small decisions

Most applications have one thing in common: they involve very many small, recurring decisions for which a large language model would be too slow and too expensive. TypeSafe and specialist media mainly name these areas.

Sorting requests: which team is responsible for a support ticket? Is a customer message urgent? Is a prospect a serious sales lead? Such questions have fixed answers and come in large numbers.

Assessing risks: does a payment look fraudulent? How serious is a security alert? It helps here that Jev gives a probability with every answer. Companies can decide that only very confident cases are handled automatically and the rest goes to a person.

Monitoring AI agents: may an agent carry out a certain action? Does its next step make sense? Decision models check this in a fraction of a second before anything happens. Evaluating the answers of other AI models is part of this too.

Controlling interfaces and content: as at Workday, the model picks suitable building blocks. According to a test by MindStudio, hobbyists also use it for a browser extension that labels posts on X as breaking news, golden nuggets or AI slop while they load.

Speed is the decisive argument. In a test on OpenRouter's blog with banking requests, Jev answered after a median of 175 milliseconds and was right 81 percent of the time. Claude Opus 5 was right 84.4 percent of the time but needed 2.3 seconds and cost about 22 times as much.

Limits, open questions and what this means for you

Jev has documented weaknesses. According to TypeSafe's own documentation, it struggles with counting, arithmetic and dates, tends to favour the first answer option listed and does not automatically treat instructions smuggled into data as hostile. For decisions about people, money or security, this means companies should measure carefully how often the model is wrong and what a mistake costs.

It is also unclear how robust the adoption figures are. The Fortune 500 figure comes solely from the manufacturer and its investor, and the two numbers differ. So far, independently documented are mainly the usage via OpenRouter, Workday as a prominent customer and a number of developer tests.

For you as a user in Germany or the EU: you will hardly ever operate Jev directly. It sits inside software that your employer or a service uses, such as an HR portal or customer service. At most, you will notice it because applications respond faster. Because Jev processes data in the US according to TypeSafe, European companies need to take this into account in their data protection assessment. If this matters, there are now open competing models that can run on your own servers. Which ones they are is covered in our overview of the new decision models.

Our assessment: the force of the adoption is real, but it shows less in individual spectacular applications than in billions of invisible mini-decisions. That is exactly why it is worth looking at who controls what these models are allowed to decide.