JevMade Sign in
← Back to guides

JevMade field notes / Architecture proposal

Choose when a workflow should act or ask for help

Kai Waehner explains how a workflow can use Jev's answers to route cases toward automation, a language model or a person.

Original by Kai WaehnerAgent workflows

Listen to this guide

JevMade’s plain-English explanation

0:00 /

AI narration

Credits

“Jev by TypeSafe AI in Workflow Orchestration and Process Intelligence” by Kai Waehner. Read the original source.

This expanded guide is an AI-narrated adaptation prepared by JevMade. It expands the source’s essential ideas, examples and caveats in JevMade’s own words and is not a word-for-word reading. The synthetic voice does not imitate the author or imply their endorsement.

Our summary

Some steps in a business process need judgment rather than a fixed rule. Kai Waehner describes asking Jev to assess those cases, then sending them along different paths. The workflow, which coordinates the steps, still decides whether to act automatically, use a language model or ask a person.

Code compares the model's answer with agreed cutoffs and records the path taken. The record also keeps the model version, cutoff and eventual outcome. Before changing the live process, a trial can score cases alongside the existing decisions without letting those scores control what happens.

This is an architecture proposal, not a measured implementation. A probability does not guarantee that an answer is correct, and changing the model can change its behavior. Approved rules remain clearer when an exact explanation is required. Cases sent to the hosted service also leave the organization's own systems.

Key takeaways

  1. Let the workflow own the cutoffs and fallback paths instead of treating the model's probability as permission to act.
  2. Record answers, model versions, review decisions and outcomes so thresholds can be checked against real cases.
  3. Start with a parallel trial that cannot alter live decisions. Keep exact approved rules for cases that must be explained line by line.

Conceptual architecture, not a tested deployment. Integration and savings claims are the author's reporting. A learned decision model is not a security or regulatory boundary, and another model with a compatible interface needs its own evaluation.

kai-waehner.de · Original published

Read the original guide Opens the author’s site in a new tab.

Keep this for later

Sign in to bookmark experiments, guides and videos, and keep notes only you can see.

Continue to sign in

We’ll bring you back to this listing.