Most companies do not need an AI that can write a poem or hold a conversation. They need something that can look at a support ticket, an invoice, or a security alert a thousand times a day and make a quick, consistent call: escalate or ignore, approve or flag, urgent or not. That gap is what a new startup called TypeSafe AI is trying to fill.
The company came out of hiding this week with a model called Jev, built by Diogo Almeida, a former OpenAI researcher who worked on the technology that made ChatGPT good at following instructions. Jev does not generate text. It takes in a description of a situation and returns a structured answer, such as a category or a yes or no, along with a confidence score, so the software calling it knows how much to trust the result.
The business case is cost and speed. TypeSafe says Jev answers in well under a second, while a typical chatbot-style model can take several seconds and sometimes minutes for the same kind of question. On price, TypeSafe charges a small fraction of a cent per unit of text, roughly a hundred times cheaper than standard chatbot pricing, and does not charge at all for the answer it sends back. If your business runs thousands of small AI checks a day, whether that is scanning claims for fraud, sorting customer emails, or reviewing invoices, that price difference adds up fast.
The company raised 40 million dollars from investors led by DCVC to build this, which is a large amount for a company just leaving stealth mode. That kind of funding suggests investors think there is real demand for AI that plugs into existing software rather than AI that chats with people.
Here is the part worth slowing down on. Independent reviewers who checked TypeSafe's own published test results found that the speed and price numbers hold up, but the quality does not always match bigger, slower models. On one of the company's own benchmark sets, covering tasks like invoice processing and security alerts, Jev scored noticeably lower on accuracy than the larger AI models it was measured against, particularly on invoice processing. TypeSafe's guarantee is that Jev will never return a badly formatted answer, not that the answer will always be correct.
That distinction matters for anyone thinking about using a tool like this. A model that always answers in the right format but sometimes reaches the wrong conclusion is still useful for high-volume, low-stakes sorting, like deciding which support ticket needs a human first. It is a different story for decisions with real financial or legal consequences, where a wrong call quietly buried in a fast, cheap answer can cost more than the savings.
The bigger pattern here is that AI is starting to split into two separate products: one built to talk to people, and one built to sit inside software and make small calls all day. If that split holds, the cost of automating the boring, repetitive judgment calls that every company already makes could keep dropping fast, which is good news for anyone drowning in tickets, claims, or invoices, as long as they check the work before trusting it fully.