AI decides, a person vouches: FlowSharp integrates Jev

Most business decisions are closed answers

In a business process, free text is the exception. Much more often what you need is a closed answer: is this request urgent? Which category does it belong to? Is the contract signed? These are small decisions, repeated dozens of times a day, and they determine who gets the work and with what priority.

Handing them to a general-purpose LLM works, but it is like asking a copywriter to apply a stamp: it is slow, it costs more than it should, and it is hard to control because the answer comes back as text that someone has to interpret. Since release 0.89, FlowSharp, among the first corporate applications to integrate Jev, TypeSafe's System 1 model, offers an alternative: an AI task can simply decide.

The name echoes Kahneman's distinction in "Thinking, Fast and Slow": one system that reasons carefully and one that answers instinctively. For us it is a product metaphor, not a scientific claim. The LLM is System 2: it reasons and writes. Jev is System 1: it decides closed questions in a fraction of the time and cost. In the Designer's AI task, choosing between the two is a single button.

What Jev does inside a process

Jev answers typed questions. There are three types: true/false, classification into defined categories, and a score from 1 to N. Each task can hold up to 20 questions, and every answer comes with a calibrated probability: the model states how confident it is instead of just answering.

The answers are written into the case data, and one of them can become the outcome that routes the flow. The most common uses are triaging support requests, lead scoring, screening CVs and checking documents: is it signed? Is it the expected type? Is it complete?

A concrete example: a support request comes in. A single Jev task with three questions establishes whether it is a complaint, which category it belongs to and how urgent it is on a scale of 1 to 5. The three answers land in the case, and the urgency decides which queue the work enters, with nobody having to read and sort it by hand.

When the model isn't sure, a person decides

Every task has a confidence threshold. If the model stays below it, the answer is not a forced guess but "undetermined". This is the point where an automatic decision stops being a risk and becomes information: a person is needed here.

Since release 0.91, the human fallback lives inside the same task. You define who decides, with what instructions and among which outcomes, without building a separate parallel step. The person finds the work where they already look for it: in the "To do" queue, the dashboard, the calendar and the Kanban board.

The administrator can also hide the model's suggestion so that whoever decides is not influenced by it. That can be useful, for example, when screening CVs, where you want an independent judgment. The result is a clear division of labour: AI decides where it is confident, and a person vouches where it is not.

Measure first, then move the threshold: gradual adoption

True to the principle "measure first, then automate", FlowSharp keeps the two paths distinct. The KPIs separate automatic decisions from human ones and calculate the fallback rate. Every AI call is logged with its estimated cost. This way the threshold moves on the basis of data, not gut feeling: if the fallback rate is high and people almost always confirm the model, the threshold may be too strict; if they often correct it, it may be too loose.

Getting started doesn't require redesigning anything. An existing manual task can be converted into a Jev decision while keeping the same outcomes, so the rest of the flow stays as it is. The "Try" button runs the decision on a real case without changing anything, so you can see what the model would have answered before you trust it with the work.

If you prefer a ready-made starting point, the System One Presets offer four preconfigured models: CV screening, lead scoring, triage and documents.

For integrators: providers, security and channels

On the technical side, the providers are ready to use: TypeSafe, with the jev-latest model, and OpenRouter, with typesafe/jev-1.13. You can pick whichever access route suits your infrastructure and your company's purchasing policies.

Data Shield masks reserved fields for Jev as well, so sensitive case data never reaches the model. The same rules apply across every channel: REST API, MCP and browser extension. Starts requested by external AI assistants must be confirmed by a person, because an assistant can propose an action but cannot launch it on its own.

One practical tip: Jev performs better with instructions and criteria written in English, even when the data stays in Italian. It pays to write the task's questions and criteria in English and leave the cases, field names and content in the company's own language.

See it live: the "Request triage" scenario

The simplest way to understand how Jev works is to watch it at work. The "Request triage" demo scenario routes six sample requests with a single Jev task and three questions, and it is ready for a live trial.

You can see where the model decides on its own, where it answers "undetermined" and who picks up that work, and then the KPIs show how the decisions were split. You measure first, then decide how much to automate.