Jev Is a Step in the Right Direction, but It’s Incomplete for Enterprise Operations

When TypeSafe AI released Jev on September 15, it gave operations teams what general-purpose language models have struggled to provide: decisions in well under a second, in a format software can use directly.

What it cannot provide, however, is the evidence behind each answer, and without that evidence, every decision that moves money or affects a customer must either go to a person for review or go unexplained.

What Jev gets right: answers software can act on

Jev solves a problem most automation teams know well: language models answer in prose, so every decision has to be parsed and checked before software can act on it.

Jev returns one of three typed results instead, a choice from a list you supply, a score on a rubric you define, or the probability that a statement is true, and it scores all the possible answers at once rather than writing one word at a time, which is where its speed comes from.

Within days developers had it scoring leads and triaging email, and open-source copies of its interface appeared. In our own test, we gave Jev one claim and business rules fitted tightly to it.

It returned the same coverage answer every time, but it could not say which rule decided it or what in the file supported it: the appearance of adjudication without the reasoning behind it.

For classification and triage, the sorting at the front end of claims, lending, KYC and invoice processes, Jev is a clear advance, and OpenAI's new Decisions API, built on GPT-6 Luna, brings a similar approach to image inputs as well as text.

What Jev cannot supply: the evidence behind each decision

The name TypeSafe chose points to the gap.

‍The company says it named the model class after Daniel Kahneman’s fast, intuitive System 1 thinking, and Kahneman described System 1 as “implicit (not available to introspection),” while slower System 2 is “potentially rule governed.”

Jev is built in System 1’s image: fast, and unable to account for how it reached an answer.

In operational terms, a regulated decision needs three properties, and a well-built typed classifier can supply only the first two.

1. Consistent means the same input produces the same output. No regulated operation can do without it, and it can be engineered: researchers at Thinking Machines Lab ran one prompt 1,000 times at temperature 0, the setting meant to remove randomness, and got 80 different completions, then made all 1,000 identical by changing the server arithmetic. Consistency is a property you can switch on in the servers, and it says nothing about whether the answer is right.
2. Calibrated means the system knows how sure it is: Jev returns a confidence score, and TypeSafe advises that when the score is low, your software should not act on the answer and should hand the case to a person. The score tells you when to escalate, but not why.
3. Justified means every condition is tied to the evidence that decided it and to the version of the rule that applied: e.g., coverage applies because the schedule on page 17 shows the policy was in force on the date of loss. A probability, however well calibrated, tells you how likely an answer is to be right; a justified decision shows why it is right, which is what an auditor, a regulator or a customer contesting the outcome will ask to see.

Why the usual workarounds leave the gap open

Teams that like Jev’s speed will try to close the gap around the model, and each of the four obvious ways runs into a standard that regulators and courts have already set.

1. Trust the repetition. Repetition proves only that an answer repeats: Jev’s clean record in our test covered only the cases our rules anticipated. Where a fully automated decision significantly affects a person, the GDPR gives them a right to “meaningful information about the logic involved,” which a repeated answer does not carry.
2. Trust the confidence score. Credit regulators stopped accepting a score as a reason long ago: since amendments in 1976, the US Equal Credit Opportunity Act has required lenders to give the specific reasons for turning an applicant down, and Regulation B calls “failed to achieve a qualifying score” an insufficient one.
3. Explain it afterward. An explanation written after the fact is a plausible story in place of a record: when Anthropic researchers planted hints that changed a model’s answer, its written reasoning mentioned the hint only 25% of the time for Claude 3.7 Sonnet and 39% for DeepSeek R1. The EU Court of Justice, in a 2025 ruling on a credit check by Dun & Bradstreet Austria, requires an explanation of “the procedure and principles actually applied,” which an after-the-fact story cannot give.
4. Put a person on every decision. This fallback quietly cancels the gain. Europe’s data protection authorities say a review only counts if it is “meaningful, rather than just a token gesture,” so a reviewer with no evidence to check has to rebuild each case from the documents, and capacity grows only as fast as the review team.

The record therefore has to come from the mechanism that makes the decision, which raises the practical question of which decisions need one.

How to tell which of your decisions a fast classifier can own

To tell which decisions a fast classifier can own, list the decisions in one process and label each one either a route or a determination.

A route is internal and reversible, such as which team picks up a claim or which queue an invoice joins.

A determination moves money or changes someone’s rights, such as paying or denying a claim, approving a loan, clearing a customer through KYC or releasing a supplier payment. If it is wrong, someone will ask you to show why it was made.

Routes are where a fast classifier pays off. They need only one safeguard, that uncertain cases reach a person before anything happens, and a confidence score like Jev’s is built for exactly that.

Determinations need a record, and five questions, which we call the decision record test, show whether a determination has one.

‍A gap on any line means the determination is not ready to run without a person.

The fifth question is less demanding than it sounds: the EU Court of Justice has said that showing how a change in the data would have changed the result can be enough.

The sort also gives you a low-risk place to start. Hand the routes to a fast classifier now, where its speed pays off at once. Run each determination in shadow mode first: the system decides every case alongside your team, but only your team’s decision takes effect, so you can compare the two without risk.

Let the system take a determination over only once its decisions hold up against your team’s and its record answers all five questions.

How Otera closes the gap: decisions that carry their own evidence

Answering five questions by hand for every case does not scale, so the system has to write the record as it decides.

Otera has worked this way from the start, because no probabilistic system, human or AI, can run a regulated operation: you need judgment enforced as deterministic code. AI models read your policies, procedures and contracts and formalize them into encoded mathematical logic for the machine to execute, so the business can check its own logic before go-live.

Each decision then runs as code against those rules and records the evidence behind every condition, where it was found and which rule version applied, while uncertain cases go to a person.

Because the model formalizes but does not decide, its variability stays in rules your team reviews, and the decision itself is deterministic, auditable down to the clause that produced it.

Allianz Partners runs claims decisions this way in production, and HFS Research’s case study of the deployment draws the lesson that “it is essential to create visibility into AI decision-making, ensuring every action can be explained, traced, and reversed if required.”

Jev is good news: it shows the market now expects decisions it can rely on, and for routing and triage it may be all you need.

For decisions that move money or affect a customer, evidence is what turns a fast answer into one an enterprise can let run on its own.

About Otera‍

Otera brings autonomous operations to the enterprises that the world depends on. Transforming highly regulated work that once required humans at every step into governed Autonomous Agents that decide and act end-to-end. Trusted by Allianz, Siemens, and SAAB.