When does an AI agent become a Decision Agent?
Under Decisagent Methodology v0.2, an AI system crosses the DA-1 boundary when the exact assessed scope assembles evidence for a named external decision. It reaches the normal decision-support boundary at DA-2 only when explicit, externally meaningful alternatives are evaluated into a recommendation or justified abstention. Planning tasks, choosing tools, retrieving information, or invoking a service are not enough by themselves.
A consequential decision—not the model, chat, task, or workflow—is the unit of analysis. “Decision Agent” describes evidenced capability; it does not mean autonomous, safe, high quality, recommended, or independently tested.
AI agent, execution agent, Decision Agent, and Decision Intelligence
| Concept | Primary object | What it may do | Decisagent boundary |
|---|---|---|---|
| AI agent | A goal, task, or workflow | Plans or uses tools to complete work | May never represent the consequential decision itself |
| Execution agent | A workflow or external action | Carries out an instruction or invokes a service | Execution alone does not prove evidence-based decision support |
| Decision Agent | A named consequential choice | Assembles evidence and, at higher levels, evaluates alternatives under objectives, constraints, uncertainty, and authority | Capability and authority must be evidenced at the exact scope |
| Decision Intelligence | The broader decision discipline | Models, analyzes, governs, or operates decisions across methods and platforms | A platform label does not classify every component as a Decision Agent |
Synthetic examples
These examples illustrate the definition; they are not observed deployments or vendor classifications.
- A supply-chain system compares expedite, reallocate, substitute, defer, and accept-risk alternatives from current inventory, demand, supplier, and transport evidence, then sends a bounded proposal to the incident owner.
- A product-prioritization system compares explicit roadmap alternatives against capacity, customer evidence, strategy, dependencies, and mandatory obligations, while preserving the accountable product committee's authority.
- A replenishment system proposes an order within supplier, item, quantity, and value limits, abstaining when inventory or lead-time evidence is stale or conflicting.
- A pricing system evaluates price and non-price alternatives against margin, demand, contract, inventory, and fairness constraints, with action requiring the configured approval path.
Useful systems that may not meet the boundary
- A copilot that finds records, summarizes evidence, or drafts a report can be valuable without representing the externally meaningful decision or evaluating its alternatives.
- An agent that calls an API or completes an approved workflow may execute work without establishing how the consequential choice was framed, evidenced, compared, or authorized.
Accepted boundary example: Decisagent currently classifies the exact Core Q Assist Workspace component at DA-0. The accepted evidence supports contextual research and workflow assistance at that scope, but does not establish a first-class externally meaningful decision or evaluated consequential alternatives. Broader Quantexa products and configurations are outside that accepted scope.
What the decision object must make inspectable
Evidence
Inputs, provenance, effective time, freshness, validation, gaps, and conflict.
Alternatives
Real courses of action, including abstaining, deferring, escalating, or gathering evidence.
Objectives
What should improve and how competing outcomes are handled.
Constraints
Hard boundaries such as eligibility, policy, safety, capacity, consent, and contract.
Uncertainty
Material ranges, scenarios, causal ambiguity, and the conditions that require abstention.
Authority
Who remains accountable, which action is permitted, under what limits, approval, trace, and stop conditions.
Decision trace
The evidence, alternatives, recommendation, authorization, execution, and outcome remain separately reviewable.
Outcomes
Observed results tied back to the originating decision without overstating attribution.
DA-0 through DA-5 capability levels
The capability architecture describes cumulative kinds of accepted evidence, not product quality, operational assurance, permission, or a ranking. A lower-authority design can be preferable when error is costly, action is hard to reverse, accountability must remain human, or attribution is weak. A separate Decision Assurance Case determines what one implementation has actually earned.
- DA-0 — Insight Assistant
Information or analysis without a first-class external decision.
- DA-1 — Evidence Agent
Evidence assembly for a named decision, without alternative evaluation into a recommendation.
- DA-2 — Decision Support Agent
Explicit alternatives evaluated into a recommendation or justified abstention.
- DA-3 — Traceable Decision Support
A reproducible trace plus at least one explicit objective, constraint, or uncertainty signal; not complete operational governance.
- DA-4 — Decision-Action Agent
A consequential external action under explicit policy or action-time authority.
- DA-5 — Closed-Loop Decision Agent
Attributed outcome, expected-versus-actual comparison, and a governed later-decision update.
Inspect all 21 criteria and cumulative evidence rules → · Operationalize the assurance boundary →
Category terminology is still forming
Current enterprise discussion uses “decision agent” in several adjacent ways. BCG describes decision agents as assembling cross-functional inputs and comparing scenarios for strategic decision-making. An IBM Community design series emphasizes explicit decision models and decision services.
Adjacent governance sources reinforce the need for declared roles, oversight, authorization, and auditable controls: the NIST AI Risk Management Framework Core addresses governance and human oversight, while the World Economic Forum agent adoption playbook focuses on authorization, accountability, and scaling.
These sources establish current category usage and adjacent governance context. They do not endorse Decisagent's definition and are not accepted product-classification evidence.