Working category definition

What is a Decision Agent?

A Decision Agent is an AI-enabled system that treats a consequential external decision as a first-class object: it assembles evidence, represents alternatives, reasons against objectives and constraints, communicates uncertainty, and produces a recommendation or governed action with a trace that can be reviewed.

A decision—not a task—is the object

Decisagent uses “decision” narrowly. The choice has an accountable owner, real alternatives, objectives and constraints, uncertainty, and a cost of error. An agent choosing which internal tool to call is not automatically making the external decision its user cares about.

Decision Agent means the system has evidence for decision capabilities. It does not mean the decision is good, autonomous, or independently tested.

What does not qualify by name alone

Decision Agent vs AI agent

An AI agent may plan, call tools, or complete tasks. It reaches the Decisagent category boundary only when it supports at least a named consequential decision with evidence—and normally only at DA-2, when alternatives are explicitly evaluated into a recommendation.

Decision Agent vs execution agent

Completing a workflow or invoking a service is execution, but it may not be decision support. Conversely, a governed Decision Agent can be DA-3 while returning its decision through an API without performing an external operational action.

Decision Agent vs Decision Intelligence

Decision Intelligence is a broader discipline and platform category. A platform may supply modeling, rules, analytics, simulation, or runtime infrastructure. Decisagent classifies only the exact product, component, configuration, deployment, or research artifact for which the evidence applies.

The ingredients Decisagent looks for

Evidence

Information supporting the decision, including explicit gaps, freshness, provenance, and conflict.

Alternatives

First-class courses of action, including abstaining, deferring, or gathering more evidence.

Objectives

What the decision seeks to improve, balanced across competing outcomes where relevant.

Constraints

Hard boundaries such as policy, eligibility, capacity, risk limits, regulation, and consent.

Authority

The human approval or pre-authorized policy that permits a consequential external effect.

Outcomes

Observed results attributable to the originating decision, not a generic monitoring signal.

DA-0 through DA-5

The maturity architecture describes kinds of evidenced capability, not a quality score or ranking. Each level is cumulative under Methodology v0.2.

  1. DA-0Insight Assistant

    Information or analysis without a first-class external decision.

  2. DA-1Evidence Agent

    Evidence assembly for a named decision, without alternative evaluation into a recommendation.

  3. DA-2Decision Support Agent

    Explicit alternatives evaluated into a recommendation or justified abstention.

  4. DA-3Governed Decision Agent

    A reproducible trace plus explicit objectives, constraints, or uncertainty.

  5. DA-4Decision-Action Agent

    A consequential external action under explicit policy or action-time authority.

  6. DA-5Closed-Loop Decision Agent

    Attributed outcome, expected-versus-actual comparison, and a governed later-decision update.

See every criterion and evidence rule →

When humans retain judgment

Human accountability does not disappear as decision support becomes more capable. People should retain or explicitly authorize judgment when evidence is incomplete, objectives conflict, consequences are hard to reverse, attribution is weak, rights or safety may be affected, or the action exceeds a bounded policy. A lower-authority design can be the more mature choice.

Decisagent Fit therefore separates useful AI participation from permissible action authority. It can recommend evidence support, analysis, recommendation, proposal, human-approved action, or tightly policy-bound action without treating the highest DA number as the goal.

The term is entering enterprise AI discourse

Current enterprise discussion increasingly distinguishes agents that orchestrate tasks from systems designed around accountable decisioning. An IBM Community series on decision-agent design emphasizes explicit decision models, while Pega’s 2026 decisioning discussion emphasizes data, governance, observability, and accountability as AI approaches execution.

Those sources show current usage of the category language; they are not Decisagent acceptance of a vendor capability or classification. Decisagent’s definition remains its own evidence-and-authority framework.

Why uncertainty matters

A system that hides missing evidence or false precision can make decision-making worse. Decisagent records UNKNOWN, CONFLICTING, and CONTRADICTED separately. No public evidence is normally unknown—not false. A recommendation may also be a justified abstention.

Assess whether AI belongs in your decision