Developer / builder · where opportunities sit

Opportunity shapes

Start with the decision the process repeats—not the model name. Strong shapes: Choice routing with a review destination; Score against a written rubric; Noul to detect a defined condition. Keep arithmetic and authorisation in ordinary code. Escalate drafting and explanation to Copilot or a general LLM after policy.

High-volume routing, agent-trace verification and exception triage are natural first opportunities—see Use cases.

Division of labour

Rules that already work stay. Copilot remains the conversational and reasoning layer; test whether an explicit decision tool improves a particular step. Do not treat a confidence number as proof that a decision is safe.

Reference flow

Receive the event; verify caller and record permissions; retrieve the minimum evidence; reject missing or stale required fields; call Jev via the adapter; validate the response; apply deterministic policy; create a route or review task; record the later human outcome. Keep prediction and final action as separate records.

Adapter owns the TypeSafe key

An Azure Function or equivalent (optionally behind API Management) holds the credential, pins the model version, validates input and normalises typed output. Power Automate, Logic Apps, Copilot Studio and Foundry-style OpenAPI tools can reuse the same business endpoint. Confirm contractual data handling and customer licensing during qualification.

Policy and logging

Confidence is a signal for the policy gate, not permission to act. Write the decision record to OneLake so thresholds and review labels can improve the next cycle.

Microsoft hook points

• Power Automate — custom connector or HTTP to the adapter
• Logic Apps — HTTP to an Entra-protected adapter
• Azure Functions — the adapter itself
• Copilot Studio — custom connector tool or agent flow (recommendation subject to policy)