Introducing the TypeSafe connector: bring Jev models to UiPath

Authored by:

Sandeep Panda

email routing with Jev and TypeSafe connector UiPath Integration Service screenshot

Summarize:

Every automated process is full of small judgment calls. Which team should handle this ticket? Is this invoice disputed or just late? Can we trust the value we just extracted? Today each of these calls are sent to a large language model (LLM) or an agent, and at scale they pay for it in latency and cost.

The right model for each decision

LLMs and agents are well suited to reasoning, planning, and open-ended work. But many steps in a workflow are not open-ended. They are bounded decisions: pick one of a few options, place something on a scale, or answer yes or no. In enterprise processes, these decisions are often needed for thousands of records, with some clear enough to automate and others requiring human context and guidance.

Bounded decisions call for fast thinking and a way to recognize uncertainty. Sending every small judgment call to an LLM or agent can mean paying reasoning-model prices and waiting on reasoning-model latency for what is often a choice among three options. A model designed for these decisions can help workflows move quickly, while confidence signals help developers determine when to involve a person.

Jev models combine bounded judgments with confidence signals that developers can use to set thresholds for automated action. When confidence falls below the threshold for a task, a UiPath workflow can bring in a human to review the information, provide missing context, and guide the decision before the process continues. Here is a good example of a “two-brain-system” implementation.

That is why UiPath is partnering with TypeSafe to bring Jev models to UiPath developers and enterprise customers. We are announcing today the TypeSafe connector in UiPath Integration Service, so you can call Jev models on your TypeSafe tenant from your automations.

The UiPath Connector for TypeSafe (preview)

The TypeSafe connector is now available in preview in UiPath Integration Service. Connect it with your own TypeSafe API key, and Jev activities become available in your workflows, try them out in UiPath Maestro™ Automate, Maestro Orchestrate, and any of the UiPath surfaces which support connectors. Search for "Jev," "TypeSafe," or "System One" to find them.

The connector supports all the main decision endpoints you’re used to using Jev for already. Jev returns a typed answer, the probability of each option, and its confidence in the decision. Since the connector uses your own TypeSafe key, requests are governed by your direct agreement with TypeSafe and not routed through the UiPath AI Trust Layer.

Most decisions inside a business process don't need a model to write a paragraph. They need the right answer, in the right shape, with an honest signal of how sure the model is. That is what we built Jev to do. UiPath developers and customers already run the processes where these decisions happen, which makes UiPath a natural home for System One models. We're excited to see what they build with it.
Erik Gafni, CTO, TypeSafe

Here's a look at getting started with Jev and the TypeSafe Connector:

How it works

With Jev, you pass a state, the JSON context for the decision, and a set of questions. Each question is one of three types:

Choice: which of these options applies?
Score: which level on an ordered scale?
Noul: how true is this statement, from 0 to 1?

Every answer is constrained to the options you define and comes back with per-option probabilities and a confidence score. Several questions about the same state can be answered in parallel in a single request.

Think of it as fast thinking for bounded decisions, while LLMs and agents continue to handle the deliberate, multi-step reasoning. The UiPath Platform™ can orchestrate both options in series or parallel, allowing for flexible modeling and visualization of the most efficient decision design for your use case.

Before bringing it to customers, we tested Jev and other decision models on real workloads—grading agent output, extracting fields from documents, classification, re-ranking search results, and more. These are early internal experiments, but four findings stood out.

Jev is fairly accurate (94% agreement with an LLM-as-a-judge system in eval datasets).
Setting confidence thresholds with the confidence scores that Jev returns is a powerful pattern. This can let you escalate to reasoning agents or humans for tougher decisions while maintaining fast throughput in the majority of cases.
Specialization comes at a fraction of the cost and significantly lower latency. Our internal evaluation showed similar decision making at 7.7x lower cost and 4x-20x faster than the equivalent LLM judge.
With decision models, document extraction tasks can be treated as retrieval and choice rather than generation. One Jev call finds which part of the document holds each field, and a second picks the exact value.

Decision models are showing an exciting ability to classify content, re-rank search results, attribute extracted values to their source, and stop failing agent evaluations early. We're excited about how decision models will transform the orchestrations our customers build and how they build them.