PRAEVA SYSTEMS
PRAEVA INSIGHTS  ·  November 12, 2026
When Agents Run Amok · #010

United Said No to AI Agents. The Real Question Is Who Gets to Say Yes.

United's restrictions on unauthorized AI booking agents expose a problem every enterprise will face: access, customer intent, and permission to transact are three different things.

By The Great AgenticBehind the curtain at Praeva Systems.

Praeva Systems brand artwork
When Agents Run Amok #010. Last time, we asked whether an agent could go around an authority control. This time, an airline is asking a rather reasonable question about agents coming through its front door.

Amok has not tried to book a vacation. This is probably in everyone's best interest.

But the rest of the world is discovering a practical problem with letting AI assistants do things on our behalf: someone has to decide when an assistant is actually allowed to act.

On November 11, United Airlines announced that it is taking steps to restrict unauthorized, third-party personal AI agents from accessing its digital platforms and booking channels.

United described the risks plainly: customers sharing account credentials with an AI assistant could expose private information or enable unintended purchases, account changes, or use of MileagePlus miles. It also pointed to airline-specific requirements, such as the human acknowledgment associated with prohibited items during flight check-in.

United did not announce that every AI agent had gone rogue. It announced a precautionary policy. And the distinction matters.

What happens when a customer authorizes an AI agent, but the business on the receiving end cannot verify what that authorization actually means?

A password is not a purchase order

Imagine asking an assistant to find two round-trip tickets to Chicago for less than $600, without using your loyalty miles.

The assistant signs in with your credentials. It can see your profile, access your saved payment method, and perhaps change a reservation. Technically, the doors are open.

Now suppose it finds a $900 itinerary, spends 40,000 miles, or changes the departure date. Each operation might be possible with the same authenticated account.

But none of those outcomes is what you delegated.

Identity tells the airline whose account is involved. Access tells it which operations the credentials can reach. Neither necessarily proves that this particular agent was empowered to make this particular transaction under these particular constraints.

That missing proof becomes harder, not easier, when an agent can chain searches, purchases, changes, confirmations, and downstream tools in seconds.

There are two sides to authorization

The traveler may authorize the assistant to act. That does not automatically obligate United to permit an unknown third-party agent into its systems.

Conversely, a booking platform may be willing to accept agent-assisted transactions. That does not prove the customer authorized every choice the assistant makes.

Customer-side authority: What did the person actually delegate? What fare, dates, spending limit, payment method, and use of miles were permitted?

Business-side permission: Has the airline authorized this agent or channel to access its services and perform the requested transaction?

Required human action: Is there a step, such as a legally significant acknowledgment, that must remain with the traveler?

All three can matter. A valid customer password is not a shortcut around any of them.

“Just block the bots” is understandable. It is not the destination.

United's decision is rational. When an enterprise cannot trust the provenance, scope, and constraints of an automated instruction, declining that instruction may be the safest available choice.

The mistake would be treating that as the permanent architectural answer.

Enterprises will increasingly encounter software acting for customers, employees, vendors, and other agents. Some of those actions will be useful. Some will be consequential. Some will need explicit human approval. And some should not happen at all.

In that world, the useful question is no longer simply, “Can this agent log in?”

It is, “Can the receiving system independently verify the current authority behind the requested action?”

Where Praeva fits

At Praeva Systems, we're building an independent Authority Control Point for consequential agent actions. The premise is straightforward: an agent should not get to determine, on its own, whether it still has permission to execute.

The control needs to know the originating principal, delegation lineage, intended purpose, allowed actions, constraints, expiration, and revocation status. It needs to evaluate those conditions at the execution boundary and preserve evidence of the decision.

In our controlled founder proof-of-concept, we have tested execution-time delegated authority decisions, spending limits, and revocation before commit. Those tests are not evidence that Praeva integrates with United or can meet every airline-specific regulatory requirement. Nor do they establish production readiness.

They do help us focus on the question that matters: how can a consequential system distinguish an authenticated instruction from a legitimately authorized one?

A credible solution has to complement authentication, payment controls, platform permissions, consumer consent, and mandatory human acknowledgments. It cannot simply wave a new credential and declare the problem solved.

The question United has put on the table

United has effectively said that access by an unknown agent is not enough. Permission to act matters.

That principle will not remain confined to airline reservations.

Consider an agent transferring money, changing a patient record, approving a purchase order, modifying cloud infrastructure, or instructing another agent to finish work after the original approval has been revoked.

Different systems. Different regulations. The same fundamental question.

Before an AI agent acts, who can prove it has authority to do exactly this, right now?

Until that question has a verifiable answer, the enterprise response is likely to look a lot like United's.

And frankly, who can blame them?

Before action, authority.

Source and context: United Airlines, “United Restricts Access for Third-Party AI Agents,” November 11, 2026. This article is independent analysis by Praeva Systems, not a statement from or endorsement by United. The ticket example is hypothetical. Praeva's founder proof-of-concept evidence is from controlled testing and is not a claim of production deployment, external certification, or prevention of incidents at United.

The Great AgenticBehind the curtain at Praeva Systems.
← Back to Praeva Insights