The Scarce Resource Is the Right to Act
A practical authority model for approval, pre-authorization, intervention, exceptions, and proof when AI changes an enterprise system.
It is safe to approve delegated action only when the organization can name the authority ceiling, meaningful intervention path, and accountable owner. The model does not grant institutional permission. A credential, consent screen, or nominal human reviewer cannot substitute for an authorized delegator and a staffed recovery path.
The action-rights contract
The exact action class, scope, value, volume, time, data, and target the system may reach.
The decision that remains human, the evidence shown, the time available, and what happens when approval expires.
A person with information and authority can pause, override, stop, or revoke the action path.
Owners for the outcome, resource, exception, recovery, evidence, and supplier obligations remain named.
Owner, briefing, proof
Owner
One accountable business owner, with separate resource and control owners where the action crosses duties.
Briefing
The mandate, action ceiling, exclusions, impact class, approval design, exception service level, and legal or policy basis.
Proof
A record of who delegated, what the system saw, which action was authorized, what changed, and how it was corrected.
What leaders should take from it
The organization must grant a permitted purpose and action. Model capability and confidence do not answer whether the act is allowed.
An approval field is weak when the reviewer cannot understand the case, intervene before harm, or reach every execution channel.
A mature read-only workflow can create more value than an immature autonomous one. Rights can narrow as risk or evidence changes.
Missing data, ambiguity, denial, distress, policy conflict, and failed recovery each need an owner, priority, and response target.
When suppliers perform consequential work, buyers need evidence, service duties, remedies, change controls, and liability allocation closer to managed services than pure software access.
EU AI Act Article 14 applies to high-risk systems in scope, not every enterprise agent. NIST AI RMF is voluntary guidance, not United States law. Mayer Brown is legal analysis rather than empirical outcome evidence. Jurisdiction-specific obligations and liability allocation still require counsel for the actual use case.
First moves before hiring anyone
Record the owner, purpose, actions, data, target systems, credentials, limits, approvals, and deactivation path.
Use a representative case to prove the approver has the context, time, and authority to stop the actual transaction.
Name the queue, staffing, response time, fallback behavior, recovery duty, and redress path before production write access.
Take one write path and classify each action separately: read, recommend, execute after approval, narrow pre-authorization, or prohibited. Keep impact and maturity on separate axes.
Claim ledger
- Field studies compare synchronous approval, asynchronous review, and sampling on intervention effectiveness and human workload.
- Courts, regulators, or contracts establish clearer responsibility for agent-caused loss in representative enterprise use cases.
- Independent audits show which authority and exception designs reduce harm without erasing economic value.