Skip to main content
← Back to Insights
Architecture7 min read

BPMN + DMN + Agent Cognition: The Three-Layer Execution Model

The single most differentiating concept in agentic architecture. Here's why mixing these three layers is the primary cause of automation project failure — and how to separate them correctly.

Paul Roma·2026-01-28

There is one architectural concept that, once understood, makes the failure mode of most enterprise AI projects immediately obvious. That concept is the Three-Layer Execution Model.

Layer 1: BPMN Process Flow

Business Process Model and Notation answers one question: WHO does what, IN WHAT ORDER?

BPMN is the right tool for orchestration. It is explicit, visual, auditable, and changeable by business analysts without developer involvement. When a process changes, you update the BPMN diagram — not the agent code.

A BPMN flow in a well-architected agentic system looks like this:

  1. Receive invoice (trigger)
  2. Extract line items → Agent (cognitive extraction task)
  3. Validate against PO → DMN (deterministic rule)
  4. Route for approval if over threshold → BPMN gateway
  5. Human review if anomaly detected → Human task
  6. Record to ERP → System task

The agent appears exactly once — where genuine cognitive work is needed.

Layer 2: DMN Decision Tables

Decision Model and Notation handles all deterministic logic. If a rule can be expressed as a truth table — IF condition THEN result — it belongs in DMN, not in an LLM prompt.

Why not just put the rule in the prompt? Three reasons:

  1. Auditability: DMN decisions are fully explainable. LLM decisions are probabilistic.
  2. Changeability: A business analyst can update a DMN table in minutes. Updating a prompt requires understanding AI behavior under distribution shift.
  3. Reliability: DMN is deterministic. LLMs are not. For compliance-critical decisions, determinism is mandatory.

Layer 3: Agent Cognition

LLM agents should handle only tasks that genuinely require cognitive capabilities — tasks where the input is ambiguous, context-dependent, or requires synthesis of multiple sources.

Examples of genuine cognitive tasks:

  • Extracting structured data from unstructured documents
  • Classifying customer intent from free-text
  • Synthesizing insights from multiple knowledge sources
  • Drafting communications that require contextual judgment

Everything else — every rule that can be written down, every process that can be diagrammed — should be in BPMN or DMN.

The Diagnostic Question

For every task in your automation system, ask: "Is this orchestration, deterministic logic, or genuine cognition?"

If it's orchestration → BPMN. If it's deterministic → DMN. If it's genuinely cognitive → Agent.

If you can't answer that question for a given task, your architecture review is incomplete.

Ready to apply this in your organization?

Book a free discovery session. We'll show you exactly how A² Agentic Architecture applies to your stack, your domain, and your automation targets.

Book a Discovery Session →