Should This Be an AI Agent?

Eight possible answers

Choose the smallest sufficient participation and authority mode
Mode	Use when	Boundary
Deterministic automation	The rule and permitted action are already explicit; AI judgment adds little.	Use tested rules, validation, idempotency, trace, and exception handling.
Copilot	A person needs faster drafting, retrieval, summarization, or exploration.	The person interprets the material and makes the decision.
Evidence support	Evidence is distributed, time-sensitive, or difficult to reconcile.	AI assembles and flags gaps; it does not evaluate alternatives into a recommendation.
Decision support	A named recurring decision has real alternatives, objectives, constraints, and uncertainty.	AI evaluates alternatives and recommends or abstains; authority remains separate.
Governed recommendation	The recommendation needs a reproducible Decision Record and explicit policy context.	AI produces a reviewable proposal; an accountable person retains the decision.
Human-approved action	A recommended action can be executed after a specific approval.	The action, limits, approver, authorization, and execution remain separately recorded.
Bounded action	A narrow, reversible, repeated action can be pre-authorized under a precise policy.	Permit only named actions within scope, limits, time, trace, override, and stop controls.
Do not automate	Evidence is structurally inadequate, accountability is absent, harm is hard to reverse, or the decision should remain human.	Improve the decision process or evidence first; AI may be inappropriate.

Five tests before choosing an agent

Name the decision. What external choice is made, by whom, when, and within what boundary?
Name the alternatives. Are there real options—including defer, abstain, escalate, or gather evidence—or only a task sequence?
State the decision basis. Can evidence, objectives, hard constraints, uncertainty, and meaningful tradeoffs be made explicit?
State the consequence. What is the error severity, reversibility, blast radius, and rights or regulatory posture?
State the outcome. Can a later result be observed and plausibly related to the originating decision at a useful horizon?

If these questions are not answerable, adding an agent usually hides the missing decision design rather than solving it.

Participation is not authority

A system can be highly useful at evidence support or recommendation while having no permission to cause an external effect. Authorization must name the action, target scope, conditions, limits, required approval, trace, exception path, and stop or revocation control. The ability to call a tool is not an authority grant.

Higher maturity is not automatically the better architecture. For irreversible, high-error-cost, rights-affecting, or weakly attributable decisions, analysis or proposal with accountable human judgment may be the responsible choice.

How much authority should an AI agent have? →

When not to automate the decision

No accountable decision owner can be named.
Material alternatives, constraints, or affected parties are hidden or disputed.
Required evidence is routinely unavailable and the system cannot abstain safely.
The action is irreversible or broadly harmful while authority and escalation remain vague.
Outcome attribution is so weak that the organization cannot learn whether the intervention helped.
The decision is regulated or rights-affecting and qualified domain review has not established an acceptable participation boundary.

Assess the real decision, not an idealized process

Decisagent Fit separately evaluates useful AI participation, recommended operating maturity, action-authority fit, evidence gaps, human control, and outcome observability. It returns a completed artifact, not an opaque score.

Run Decisagent Fit