Authority in AI-Enabled Work

AI can identify a problem in seconds. It can recommend a response immediately. It can trigger a workflow. It can execute a bounded action. It may even be legitimately delegated authority to make specified Organizational Decisions.

The technology may be ready.

The authority architecture may not be.

The governing question is:

Can legitimate Decision Authority, escalation, and intervention operate at the speed and scale of AI-enabled work without becoming either meaningless control or uncontrolled autonomy?

That is Authority in AI-Enabled Work.

AI does not create a new source of authority

The Wave 5 rule is explicit:

AI capability, technical access, workflow permission, model Confidence, historical use, repeated success, or recommendation quality does not independently create or enlarge Decision Rights, Decision Authority, Decision Jurisdiction, accountability, or governance legitimacy. AI may exercise specified authority or execution only where that authority or execution has been legitimately established through the applicable governance architecture.

Capability tells us what the system can do.

Governance tells us what it may do.

Recommendation and authority remain different

AI may analyze, predict, prioritize, recommend, synthesize, or identify a Decision requirement without possessing Decision Authority.

This matters because strong recommendations can become de facto Decisions if the organization routinely accepts them without meaningful challenge.

The formal architecture may still say:

The manager decides.

But if recommendations are effectively auto-accepted or execution begins before meaningful intervention is possible, the enacted authority arrangement may differ from the documented one.

Authority therefore has to be evaluated in operation.

Runtime Authority is the operating test

Decision Authority asks whether legitimate organizational power exists.

Runtime Authority asks whether that authority is identifiable, accessible, exercisable in time, supported by sufficient Decision Context, and preserved in enacted operation.

If the only legitimate authority path takes hours while the AI-enabled activity can create material consequences in seconds, the organization has an operating-model problem.

The issue is not that governance takes time.

The issue is whether legitimate authority can still operate inside the material Decision Horizon.

Human presence can become nominal authority control

Adding a human approval step can be appropriate.

It can also be fiction.

If a human is expected to approve hundreds of machine-speed actions without sufficient context, time, or ability to stop the activity, the organization does not have meaningful human authority at that point.

It has a workflow label.

Human presence does not establish substantive governance.

KO-061 owns that general rule.

For authority, the practical test is:

Can the designated authority holder actually exercise authority before the consequence outruns the Decision Horizon?

Bounded delegation can make authority usable

The answer to AI-enabled speed is not necessarily more centralized approval.

Centralization may preserve formal control while making authority inaccessible in practice.

Bounded delegation can place legitimate authority closer to the relevant Decision Context, operational condition, customer, Decision Horizon, or AI-enabled mechanism itself.

AI may exercise explicitly delegated authority within defined boundaries.

But it may not enlarge those boundaries itself.

KO-056 owns how delegation is established.

KO-063 owns whether the resulting authority remains usable and enforceable in operation.

Faster work requires clearer boundaries

As execution speed increases, vague authority becomes more dangerous.

AI-enabled work benefits from explicit definition of Decision Class, Decision Mandate, Decision Jurisdiction, Authority Thresholds, Execution Thresholds, included actions, prohibited actions, escalation conditions, stop conditions, intervention conditions, and review or termination criteria.

The objective is not to send every AI action to a committee.

It is to make the authority boundary precise enough to operate at the speed of the activity.

Runtime controls can preserve authority without becoming authority holders

Organizations may use transaction limits, rate limits, hold points, automated suspension, mandatory handoff, escalation triggers, or Threshold-based containment.

Those controls can prevent activity from outrunning legitimate authority.

They do not themselves possess Decision Authority.

Technical controls implement governance established elsewhere.

The system may be able to stop something. That does not determine who is authorized to decide what happens next.

Intervention need does not create authority

A manager may recognize that an automated action should be stopped.

That recognition does not itself create authority to stop it.

A technical team may have credentials to disable a system without possessing authority to change the underlying Organizational Decision.

The correct question is:

Who may legitimately stop, redirect, suspend, or contain the activity, under what conditions, and can that authority operate before the consequence becomes difficult to reverse?

Existing authority architecture should govern the intervention.

Escalation should remain jurisdictional

AI may identify more exceptions and more situations outside normal authority.

If every one is escalated upward, the organization recreates the bottleneck automation was meant to reduce.

Escalation should occur where the current Decision Jurisdiction cannot legitimately resolve the matter.

That may be because authority is insufficient, a Threshold is crossed, a material cross-scope consequence appears, a Mandate boundary is exceeded, or a different jurisdiction is required.

Better AI should make those conditions more visible. It should not make escalation indiscriminate.

Authority boundaries must survive enacted operation

Where humans, AI systems, automated systems, suppliers, or hybrid arrangements can authorize or execute materially consequential action, the operating mechanism should preserve the applicable authority boundary in practice.

Where feasible, systems may prevent, constrain, intercept, or require additional authorization for activity outside legitimate authority.

Where prevention is impractical, proportionate controls may include monitoring, detection, intervention, suspension, containment, and traceability.

This is where authority stops being a policy statement and becomes an operating capability.

Material change can require reassessment

Delegated AI authority should not be treated as permanently valid because it was approved once.

Material changes in capability, configuration, system access, evidence environment, operating behavior, or relevant performance characteristics may affect the assumptions supporting the existing authority arrangement.

Possible responses include continued operation, qualified operation, added safeguards, narrowed authority, suspension, renewed validation, or legitimately revised authority.

The point is continuing legitimacy, not one-time approval.

What management should look for

Take one consequential AI-enabled activity.

Ask:

What is AI actually doing?

Recommending? Prioritizing? Deciding? Executing? Monitoring?

Then:

Which Decision Rights apply?

Who or what holds legitimate Decision Authority?

What Mandate, Jurisdiction, and Thresholds bound it?

If authority is delegated:

Is the delegation explicit?

Can the AI expand or reinterpret its own boundary?

It should not.

Then test runtime usability:

Can the relevant authority be identified and exercised within the Decision Horizon?

Can the activity be paused or contained before irreversible consequence where required?

If the primary authority path fails, is there a legitimate alternate path?

Finally:

Does the authority architecture operate at the speed of the AI-enabled activity, or does it become unusable precisely when it matters most?

Boundary

Authority in AI-Enabled Work is not a new canonical AIMT construct, a new source of Decision Rights, technical permission, recommendation quality, workflow access, human presence, or a separate intervention-authority category.

Runtime controls are not authority holders. Escalation is not the universal response to uncertainty. Faster AI does not justify weaker governance.

The governing rule is:

In AI-enabled work, legitimate authority must remain explicitly grounded in Decision Rights, bounded through Mandate, Jurisdiction, Thresholds, and delegation, and operationally usable within the speed, consequence, Reversibility, and propagation of the activity being governed.

The management implication

AI does not remove the authority problem. It exposes it.

The practical discipline is:

Do not ask only whether AI can act. Ask who or what may legitimately authorize the action, what boundaries apply, how those boundaries are enforced, and whether legitimate intervention can still occur before the consequence outruns the Management System.