Product1 min read
The Decision node: calibrated AI judgments in your workflows
Until now, a workflow that had to make a judgment asked a Claude agent for structured output and branched on what it wrote. That works, but you never knew how sure the agent was. The new Decision node asks TypeSafe Jev instead, and gets back how sure it is with every answer.
Oshadha VimukthiBuilder of Worfilo
What it does
Jev is a model built for judgments rather than text. You send state and typed questions; it returns probabilities. The Decision node supports all three question types (yes/no, choice and score), asks them in one parallel call, and branches on the result four ways: a yes threshold, a choice's options, score bands, or rules.
When the model is unsure, the node says so. Below your minimum confidence it fires an uncertain port, which you can wire to an agent or a person. Clear cases stay cheap; hard cases get attention.
How it did
We ran 40 labelled support tickets three times through the node, each with three questions: team (4 options), bug or not, and priority. That made 120 live requests.
- Team routing: 100% correct, with a mean confidence of 0.97.
- Bug or not: 88% correct. Requests that mention an error but are really questions were the hard part.
- Median latency 0.70s and p95 0.82s. Asking one question took the same time as asking three.
- About 490 input tokens per request. Output tokens are free.
It is a small set we wrote ourselves, so read it as a sanity check. The benchmark script ships with the backend, so you can rerun it on your own tickets.
Try it
Add a TypeSafe credential, drop a Decision node on the canvas and pick the Support triage preset. The AI decisions guide covers every setting, and the lead scoring use case shows a full workflow.
Share this post
Build it on the canvas
Describe a workflow, watch every node run, and ship it as an API. Free to start.
More from the blog
All postsKeep reading
- AI decisions: classify, score and branch with the Decision nodeUse the Decision node and TypeSafe Jev to answer yes/no, choice and score questions with calibrated probabilities, then branch, score and escalate on them.
- AI lead scoring workflowScore inbound leads with the Decision node: TypeSafe Jev rates fit and intent, a weighted scorecard sorts them into hot, warm and cold, and unclear leads go to an agent.
- Decision node: typed judgments with TypeSafe JevThe Decision node asks TypeSafe Jev yes/no, choice and score questions, then branches on calibrated probabilities, a weighted scorecard or rules.