DMN Plus LLMs: Alex Porcelli Pitches Auditable Decision Models for Enterprise Agents

Alex Porcelli has a blunt diagnosis for enterprises deploying AI agents on consequential work: the models are non-deterministic, and when a decision goes wrong, nobody can say who owned it. His proposed fix, presented in a session on decision models in agentic architectures, is to stop asking the LLM to be the decision-maker and instead place a formal decision layer between the model and the action. That layer, built on the Decision Model and Notation standard, keeps business logic explicit and reviewable while the model handles the messy language work around it.

The pitch splits responsibilities along a familiar enterprise seam. Business leaders define and own the decision logic in DMN tables and models, while engineers govern how agents invoke that logic, manage context, and enforce boundaries at runtime. Porcelli pairs the decision models with agent skills, a pattern that gives an agent a bounded, declared capability rather than open-ended tool access, and with NeMo guardrails to constrain what the model is allowed to do before and after a decision is reached. The result is an architecture where a model's output feeds a deterministic rule set instead of replacing it.

The problem he is targeting is real and increasingly expensive. Regulated industries such as lending, insurance, claims processing, and clinical triage have spent decades building audit trails around rules engines precisely because decisions in those domains must be explainable after the fact. Dropping an LLM into that pipeline introduces variance that compliance teams cannot sign off on, and the common workaround, logging prompts and outputs, produces evidence of what the model said rather than why the business reached a conclusion. A DMN layer restores the artifact regulators and internal auditors actually ask for: a versioned decision definition with traceable inputs.

None of this is technically novel in isolation. DMN has been an OMG standard since 2013 and is embedded in business process tooling from vendors including Camunda, Red Hat, and IBM. Agent frameworks have been adding guardrail hooks and structured tool-calling throughout the past year. What Porcelli is arguing for is a specific composition: treat the LLM as a translator and classifier that maps unstructured input onto the inputs of a decision model, then let the model render the verdict. Guardrails sit at the edges to catch prompt injection, off-policy requests, and outputs that violate declared constraints.

The approach has costs worth naming. Decision models must be authored and maintained, which reintroduces the kind of governance work teams hoped agents would eliminate, and every new capability needs a corresponding model update rather than a prompt tweak. There is also the question of coverage: some decisions resist tabular expression because their logic is genuinely open-ended, and forcing them into DMN can be worse than leaving them probabilistic. Porcelli's answer is scope discipline, applying the pattern where stakes and auditability justify it rather than everywhere.

The open question is whether enterprises will accept two systems of record for agent behavior, one in the model and one in the decision layer, or whether they will keep pushing for a single end-to-end learned pipeline. Adoption will likely be decided less by architecture debates than by the first high-profile incident where an autonomous agent's reasoning cannot be reconstructed. Watch whether agent platforms ship first-class DMN execution alongside their guardrail and skills APIs, and whether auditors start asking for decision model versions in the same breath as model cards.