Jev is the hottest discussion in town right now - a new class of AI model that combines the large, unstructured input of a LLM with the short, sharp output of a if-then-else rule. It has a simple purpose: to make decisions based on a single question given a mass of textual data. Taking just hundreds of miliseconds and costing a fraction of a cent to execute, it moves away from the heavyweight and costly approach of using LLMs to decide on true or false type questions.
My colleague explaimed while doing some tests, “This changes everything. Decisions for free!” I think he was right (largely). TypeSafe, the start-up behind Jev, calls Jev a “System One” model, after Daniel Kahneman’s term for fast, intuitive judgement, a hunch, and my colleague’s verdict was itself a perfectly good System One response to a System One product!
What this series tries to add is the first attempt at a System Two pass: the slower, deliberate look at exactly where a fast, radically cheap way of getting a neural network to make a structured judgement call actually lands inside the applications we build for a living, so that the instinct that something has changed turns into a map of what, and where.
I am going to explore this issue in three parts:
- What is Jev?
- How might Jev be best used in Platform - Case and Workflow?
- How might Jev be used in Customer Decision Hub?
Large language models behave in a “reasoning” way to a question: ask it a yes-or-no question and it gives you a paragraph. A chat model reasons in public, token by token, and someone pays for every token on the way to the answer. For an enterprise that wants one thing only — approve or refer, urgent or not, team A or team B — that is a lot of expensively purchased language standing between the question and the actual decision. And it takes a long time to get to that answer.
TypeSafe AI, a young San Francisco startup founded by Diogo Almeida, who previously worked on ChatGPT at OpenAI, has spent the last few weeks arguing this is the wrong shape of model for most business decisions. Its new model, Jev, opened in early access on 15 September and has since become the most talked-about launch of the autumn.
Jev’s pitch rests on a distinction familiar to anyone who has read Kahneman’s “Thinking, Fast and Slow”: fast, intuitive judgement versus slow, deliberate reasoning. TypeSafe calls Jev a “System One” model — built for the fast lane. Rather than reasoning its way to an answer in visible text, it takes a block of state (a support request, a transcript, a customer record) and a set of typed questions, and returns typed answers in a single pass. It is fast - a fraction of a second and many hundred times cheaper than a LLM call.
There are three question types on offer. “Choice” picks the best-fitting option from an unordered set and returns a probability for each candidate. “Score” rates the input against ordered levels — low, medium, high, say — and returns a distribution alongside the headline number. “Noul”, TypeSafe’s own term for its yes/no primitive, returns a single probability that a statement is true, with no separate confidence figure, because with a yes/no question the number already carries the certainty: 0.5 means genuinely undecided.
Put in terms a Pega consultant will recognise, the three primitives translate easily into questions that already sit inside case types and decision strategies.
- “Could this insurance claim be fraudulent?” is a Noul — a single probability, nothing more.
- “Is this complaint likely to be resolved quickly?” could be asked the same way, though phrasing it as yes/no throws away information a Score question would keep: same-day, this week, or escalate, each with its own likelihood.
- “Is this a high-value or high-impact case?” is a Noul too when used as a first-contact webchat with no history behind it.
- "What is the most likely customer intent behind this email, from this list of categories” is a clean Choice,
- “Which model handset should we offer this customer on renewal” returning a probability across the whole list rather than a single verdict — which, not coincidentally, is the example that matters most once we reach arbitration and the Customer Decision Hub.
The appeal is speed and cost, and TypeSafe’s own numbers on both are bold — one widely circulated benchmark this month put Jev at roughly 193 times the speed and 445 times cheaper than an equivalent chat model asked to make the same call. Whether those multiples survive contact with a real production workload is a separate question, and one worth revisiting once Jev has a few months of live traffic behind it rather than a launch demo. But the underlying logic is sound enough: if the only output required is “yes, 0.83,” there is no commercial reason to pay a model to write an essay first.
That is a genuinely useful primitive, and the use cases surfacing in the first fortnight look sensible rather than gimmicky: routing support tickets by urgency and team, scoring inbound leads ahead of a human sales pass, acting as a fast ranker of priorities. None of that is a new problem. What is new is a model built to answer it directly, in milliseconds, without composing a sentence it will never need to write.
How to integrate Jev into a Pega process - Kajal Kaspate
Where this gets interesting for Pega practitioners is the word “decision” itself, which TypeSafe uses freely, and which already sits at the heart of a great deal of Pega’s own case management and decisioning work. The next two blog posts will look at if, when and how Jev might be used in our applicaitons and how we can talk to our clients about it.
So, I will look at where a Jev primitive like this lands inside case management and workflow — routing, triage, and when a Jev-style hunch earns a place alongside Process AI’s own predictions. The appeal is obvious: feed in the Case summary, documents and state and get fast decisions based on it. - without lots of NLP processing, historical analysis and analytical complexity.
The post after that turns to the harder case: the Customer Decision Hub, arbitration, and whether “decisions for free” has any real claim on next-best-action at all.