Before Agentic AI Can Act, Define Its Authority
Jump to Section
Agentic AI governance starts with a practical question: What is an AI agent authorized to do?
Booking a flight may sound like a straightforward task to delegate. But an assistant that moves from comparing itineraries to completing a purchase may spend money, access personal information, use loyalty miles, and accept travel terms. Completing the booking does not establish that each of those actions was authorized.
United Airlines highlighted this issue on November 11, 2026, when it announced restrictions on unauthorized third-party personal AI agents across its digital platforms and distribution channels. United cited concerns about credential sharing, unintended purchases, account changes, use of loyalty miles, and automated actions that could interfere with regulatory requirements during check-in.
Organizations adopting AI agents should ask similar questions about their own workflows. What can the agent do? Which accounts and data can it access? When does it need approval, and who can stop it?
Define those limits before the agent begins acting on someone’s behalf.
What Does Authority Mean in Agentic AI Governance?
For this discussion, an AI agent’s authority is the scope of actions an AI agent may take on a person’s or organization’s behalf. That includes what it can do, what information it can access, and when it needs further approval. Whether an action creates a binding legal obligation is a separate question that depends on the facts and applicable law. Permission to search is not permission to purchase, and permission to purchase should not be an open-ended mandate.
Consider a traveler who asks an AI Agent to book a flight under $500. The assistant identifies a $475 fare and purchases it. The price meets the stated limit, but the fare is nonrefundable, excludes a carry-on bag, and arrives at an airport two hours from the intended destination. Had the traveler known about the caveats, they would not have gone forward with purchasing it.
The agent completed the task as it interpreted it. Whether the traveler authorized that outcome is a different question.
Before deployment, teams should define six elements of authority:
- Allowed action: What the agent may or may not search, recommend, change, submit, or purchase.
- Data access: Which information the agent may access, share, and retain.
- Limits: Thresholds for price, destination, contract terms, risk, or other material conditions.
- Approval triggers: Changes or thresholds that require human confirmation before proceeding.
- Evidence: The instructions, approvals, actions, and outcomes the organization must record.
- Revocation: Who can stop the agent, how access can be removed, and what happens to actions already underway.
When human approval is required, the person reviewing the action should see the relevant terms and have a practical opportunity to reject or change it. These decisions belong in the product requirements. Privacy, legal, security, compliance, and business owners should help define them before deployment.
How Should Regulatory Requirements Shape AI Agent Workflows?
United’s stated concerns also illustrate why organizations cannot assess agentic AI solely through the lens of emerging AI legislation.
An automated workflow may involve existing obligations for personal information, consumer disclosures, payments, contracts, or regulated activities. Which obligations apply depends on what the agent does, where it operates, and who is involved.
United cited potential conflicts with aviation regulations as one reason for its restrictions. That is the airline’s stated rationale, rather than a general legal conclusion that AI-assisted travel violates federal rules.
For organizations building or adopting AI agents, the task is to connect applicable requirements to specific actions:
- What information must reach the customer before the action proceeds?
- What acknowledgment or authorization is required, and who must provide it?
- What data can the agent access, share, or retain?
- What records will demonstrate that the required steps occurred?
The answers should shape the workflow. If a requirement calls for a person’s acknowledgment, for example, the process should deliver the relevant information to that person and obtain the acknowledgment before proceeding. Teams should revisit the design when applicable requirements change.
How Should Organizations Govern Third-Party AI Agents?
Organizations also need to consider AI agents they did not build.
A customer may use an AI agent to make a purchase. An employee may use one to interact with a vendor. A service provider may add an AI agent into an existing process.
In each case, more than one organization’s rules may apply. A customer’s authorization does not, by itself, establish that every participating platform permits the agent’s actions or that every applicable requirement has been satisfied.
Organizations should identify the systems and data an outside agent can access, confirm whether the relevant platforms permit that access, evaluate the available safeguards and evidence, and assign an owner. The assessment should also define the conditions for allowing, limiting, or blocking the agent.
Restrictions may be appropriate when the organization cannot establish sufficient control. When access is permitted, the decision should document the use case, permissions, limitations, and reasons for approval.
Why AI Governance Needs an Incident Response Process
Governance does not end when an AI use case is approved. An authorized agent can still make an unintended purchase, disclose information to the wrong recipient, or act outside its permitted scope. The response process should be designed before autonomy expands.
Teams need to know who can suspend access, who assesses customer and business impact, which records must be preserved, and who evaluates potential privacy, contractual, regulatory, or reporting obligations. An AI-related event may require privacy, legal, compliance, security, product, and business teams to investigate together. Whether the event is reportable depends on the facts and applicable requirements.
Governance records give investigators a starting point: the approved use case, the agent’s permitted actions, the safeguard that should have operated, and any conditions attached to approval. Investigators can compare the expectations with what actually happened.
Radar AI Risk supports centralized AI system inventory, structured assessment against internal policies and regulatory frameworks, and traceable governance decisions. RadarFirst AI Incident Management supports consistent intake, coordinated investigation, impact assessment, regulatory intelligence, and defensible response documentation. Together, these capabilities help teams connect the decision to permit an AI use case with the evidence needed when an outcome requires investigation.
Explore RadarFirst AI governance and AI Incident Management.
A Practical Framework for Expanding AI Agent Autonomy
United’s announcement illustrates a decision other organizations will face: which actions to permit, under what conditions, and when to restrict access.
Organizations can evaluate expanded autonomy in deliberate stages:
- Start with one workflow. Define the business purpose, expected outcome, system owner, and the people, data, and platforms affected.
- Define the agent’s authority. Specify allowed actions, prohibited actions, data access, limits, approval triggers, evidence requirements, and revocation.
- Connect requirements to the workflow. Identify applicable internal policies, contractual terms, regulatory requirements, and platform conditions.
- Test the controls and response process. Check that limits work, approvals are informed, activity is recorded, and accountable teams can investigate and suspend access.
- Expand only when the evidence supports it. Reassess the use case as the agent, workflow, data, dependencies, or requirements change.
A successful transaction shows that an agent completed a task. Teams should also check whether it stayed within its permissions, followed the required steps, and produced a result the organization was prepared to accept.
Expand autonomy based on those findings, with an owner responsible for reviewing the decision.
Turn AI Authority Into a Defensible Operating Model
RadarFirst helps teams inventory AI systems, apply internal policies and regulatory frameworks, document governance decisions, and coordinate the response when an AI-related event requires investigation. See how Radar AI Risk and AI Incident Management connect oversight with action.
Frequently Asked Questions
What is agentic AI governance?
Agentic AI governance is the set of policies, decision criteria, controls, ownership, monitoring, and records used to govern AI systems that can take actions. It should define both the agent’s permitted authority and the organization’s response when the agent acts outside that authority.
How is permission to search different from permission to purchase?
Search permission allows an agent to retrieve and compare information. Purchase permission authorizes a transaction and should include limits for price, terms, destination, data use, and any condition that requires human confirmation.
When should an AI agent require human approval?
Require human approval when an action crosses a defined threshold, changes a material term, creates a legal or financial commitment, uses sensitive data, or requires an acknowledgment from a person. The approval interface should show the information needed to make an informed decision.
What records support a defensible AI governance decision?
Useful records include the use case and owner, approved authority, applicable policies and regulatory requirements, risk assessment, control design, approvals and overrides, system changes, incident evidence, and the rationale for final decisions.
Let’s Get Started
Trusted by leading organizations, RadarFirst enables teams to manage incidents with speed, consistency, and defensibility by standardizing how incidents are captured, assessed, and actioned.