CDF Workflows: Autonomous Actions vs Human Approval?

Question

Hi everyone, I have a question regarding operational workflows in Cognite Data Fusion.

Suppose CDF detects a potential quality degradation in a batch process based on process parameters (e.g., temperature, pressure, flow, etc.). What is the typical approach in real industrial deployments?

i)Does CDF automatically trigger the next course of action (for example, adjusting process parameters through an integrated control system)?

ii)Or does it first generate an alert/notification or workflow for the relevant engineer/operator, who reviews the recommendation and approves the action before anything is executed?

I'm trying to understand how Cognite is typically used in production environments, especially the balance between AI-driven recommendations, human-in-the-loop decision making, and automation. Any insights or real-world examples would be greatly appreciated.

Peter Arwanitis · Answer

Hi ​@Prince Vekariya and thank you for reaching out to our Cognite Community!My name is Peter Arwanitis, and I’m with Cognite working as a Solution Architect / Technical Account Manager.Short answer: in real production deployments it's almost always (ii) — human-in-the-loop. Fully automated closed-loop control from CDF into a control system (i) is rare, for architectural and safety reasons, not product ones.Why: CDF operates at the IT/analytics layer, while process control (DCS/SCADA/PLC/SIS) lives in the OT domain, typically with read-only flow OT→IT and functional-safety constraints. So CDF normally reads process data and produces insight; writing setpoints back is a deliberate, customer-owned, safety-gated integration — never a default.Typical pattern:CDF detects/scores the degradation (a calculation, simulation-model, or Atlas AI agent).It surfaces a recommendation — notification, dashboard, a Workflow task, or a work order.An engineer/operator reviews and approves.The action is executed — manually, or by the control system in its own domain.CDF gives you the building blocks for all of this: workflow triggers (signal must match a supported triggerRules type), CDF Functions for arbitrary Python in a workflow task, alerts via extraction-pipeline runs, or Records & Streams as an event queue external systems can subscribe to.But CDF doesn't pre-define the operational process — it can be composed.Happy to hear if that explanation works for you. Or if you want to go into more detail about your specific case.best regardsPeter(=PA=)

Sign up

Welcome to Cognite Hub