When AI should decide, not talk: decision models and more deterministic workflows

Most AI workflows start with a large language model (LLM), the kind of model behind ChatGPT or Claude that reads and writes text. That is a sensible start. The trouble shows up later, when the model sits in the middle of a process and its job is no longer to write. Its job is to choose.

Route this ticket to billing or support. Approve this refund or escalate it. Is this intake form complete? Which category does this invoice belong to? These are decisions, and a business needs them to be consistent, auditable and cheap. An LLM answers them by generating text, and free text in the middle of a pipeline is where things break: a label that is not on your list, an answer phrased differently than last time, a sentence your code cannot parse.

Workflows need decisions, not paragraphs

A step that decides has three requirements that a chat reply does not.

• Consistency. Similar inputs should land in the same place. If two nearly identical tickets go to different teams, people stop trusting the automation.
• Auditability. Months later, someone will ask why a claim was escalated or a lead was dropped. You need the input, the options, the choice and how sure the system was.
• Cost and speed. Decisions run at volume. A step that fires on every email, form and document has to be fast and inexpensive per call.

Generating a paragraph to answer a multiple-choice question works against all three.

A new category: decision models, with Jev as the current example

On September 15, 2026, TypeSafe AI launched Jev, which it calls a "System One" model. The name borrows psychologist Daniel Kahneman's term for fast, intuitive judgment, as opposed to slow, deliberate reasoning. TechTarget reports that the company was founded by Diogo Almeida, a former OpenAI researcher.

Jev does not write. According to TypeSafe's documentation, it answers three kinds of questions: a choice (pick one option from a list you define), a score (rate something against levels you define) and a "noul" (the probability that a statement is true). Choices and scores come back with probabilities and a confidence value. You set it up with a prompt, not with training data.

The vendor's claims are ambitious: end-to-end responses in 70 to 500 milliseconds, confidence that is calibrated (higher confidence should mean higher accuracy), and pricing that charges only for input tokens, the chunks of text models bill by. These are claims. TypeSafe's own launch post acknowledges that its headline benchmark workflows were written by its own team and scored against answers from other models.

Analysts quoted by TechTarget see a clear fit and a clear boundary. Ned Bellavance of Ned in the Cloud compares Jev calls to "a function you'd add to a program or automation routine," and adds that "for human-in-the-loop workflows, you're still going to be using an LLM." Laurie Voss of Arize expects the major labs to "release decision models very, very quickly."

Separate talking from deciding

The pattern matters more than the product. We split AI steps into two kinds.

Talking steps read messy input and produce language: pulling fields out of an email, summarizing a contract, drafting a reply. LLMs are good at this.

Deciding steps pick from options you defined in advance. They should never return free text. There are three ways to build them, in this order of preference:

1. Plain rules when the logic is exact: an amount over a limit, a required field missing, a date past a deadline. Code is faster, cheaper and easier to verify than any model.
2. A decision model like Jev when the input is messy language but the answer still comes from a known list.
3. An LLM with structured outputs, where the model must answer inside a fixed schema. OpenAI, for example, states that its Structured Outputs feature ensures responses adhere to the schema you supply, so the model cannot invent a label that is not on your list. The confidence, though, is usually a number the model writes into its answer, and it needs the same checks.

Every deciding step gets a confidence threshold. Above it, the workflow acts. Below it, a person decides. Set the threshold per action: auto-tagging a ticket can tolerate more risk than issuing a refund.

A reference architecture in six parts

1. Input. The email, form, document or event that starts the workflow.
2. Extraction. An LLM or a parser pulls out only the fields the decision needs; irrelevant text can make a decision worse.
3. Decision. A rule, a decision model or a schema-bound LLM picks from defined options, including an explicit "none of these" option, and returns a confidence value.
4. Action. Above the threshold, the system routes, approves or updates the record.
5. Human review queue. Below the threshold, or for high-impact actions, the case goes to a person with the input, the options and the model's output attached.
6. Log and evaluation. Every decision is stored with its input, options, model version, threshold, output, confidence and final outcome. That log is your audit trail and next quarter's test set.

1. 01InputThe email, form, document or event
2. 02ExtractionOnly the fields the decision needs
3. 03DecisionDefined options and a confidence value
   Human reviewBelow the threshold or high impact, a person decides
4. 04ActionAbove the threshold: route, approve, update

Log and evaluationEvery decision stored; next quarter's test set

What "deterministic" does and does not mean

A deterministic format is not a deterministic answer. As Bellavance told TechTarget, Jev's format is deterministic but its content is not, and there is no guarantee of the same answer or confidence on a later run. A valid label can still be the wrong label.

Vendor numbers are not independent benchmarks. Jev launched in mid-September 2026, and early independent tests are mixed. In one test published in Towards Data Science on a bank-support classification task, Jev was more accurate than the open model it was compared with, but its answers in the 0.7 to 0.9 confidence range were right only about half the time. Sending its low-confidence cases to a fallback LLM made results worse, not better. The same article describes a separate test in which messages that fit none of the categories, with no "none of these" option offered, were still assigned one with confidence of 0.99 or higher.

That last finding is a design lesson: if an input can fall outside your options, give the model a way to say so.

Calibration is local. A model that looks calibrated on someone else's benchmark can be overconfident on your invoices.

Lock-in is a real risk. According to MarkTechPost, TypeSafe has not published model weights or a self-hosting option. Jev is also listed on Cloudflare Workers AI, which gives you a second place to call it, but not a second model. Keep the decision step behind your own interface so you can swap the engine.

Sometimes rules win. If you can write the logic as an if-statement, do. Counting, arithmetic, date comparisons and record lookups belong in code; Towards Data Science reports that TypeSafe itself cautions Jev can struggle with the first three.

How to evaluate a decision step before you trust it

We use the same plan whether the engine is Jev, an LLM with a schema or a set of rules.

1. Build a test set from real past cases. Use a few hundred decisions your team already made, including awkward cases and ones that fit no category.
2. Measure agreement with human decisions. Overall, per category and on the cases that matter most. A model that is right on routine tickets and wrong on refunds is not ready.
3. Check calibration. Group answers by confidence. When the model says 90%, is it right about 90% of the time? Set the threshold on one slice of data and confirm it on another.
4. Measure cost and latency on your inputs. Long documents cost more and take longer, whatever the launch post says.
5. Run in shadow mode. The decision step runs next to the current process, logging what it would have done, while people keep deciding. Compare, adjust the threshold, then switch.
6. Pin the model version and rerun the test set whenever it changes.

What this means for your business

Most AI workflows contain a few steps that should decide instead of talk. Finding them is often the fastest way to make automation more reliable and cheaper to run, whether the answer is a decision model, a fixed schema or an if-statement that should have been there all along.

Decision models like Jev are a promising tool for those steps, and an early one. Treat vendor claims as hypotheses, test on your own history, keep a person on the uncertain cases and log every decision.

If you want to know which of your AI steps should decide instead of talk, book a meeting with us.