United will limit AI agents to flight searches: they won't be allowed to complete purchases

You ask an AI assistant to find a morning flight, choose a seat and buy a ticket. It handles the search, reaches checkout — and stops: the airline won't let it complete the transaction. That is precisely the boundary United is drawing. For developers of personal assistants, this is a practical problem: being able to control a browser is not enough to carry out a user's request from start to finish.

According to The Verge, citing The Wall Street Journal, United will not allow third-party personal AI assistants to buy tickets on passengers' behalf. They will be able to display flights. They will not be able to complete the transaction. The company attributes the decision to concerns about personal data, the risk of unintended purchases and the need to comply with passenger disclosure requirements.

Where an agent's assignment ends

Consider a routine task: fly to a meeting and return that evening. The user sets a budget and schedule, and the agent compares options. At this stage, it helps make sense of the offers. But checkout introduces actions with consequences: accessing an account, sharing personal data, choosing a fare and charging a payment.

These also involve confirmations that cannot casually be reduced to another click. The report mentions, for example, a passenger confirming that their baggage contains no explosives. United is concerned that automated checkout could bypass mandatory disclosures and procedures.

It helps to separate two questions here. Has the agent been instructed to buy a ticket? And is it authorized to complete every step of checkout on its own? To the user, this may look like a single task, but the seller has to consider the consequences of individual actions. A broad instruction to "arrange the trip" does not explain how the assistant should respond to a price change or an additional condition.

The original report provides no technical details about the blocking mechanism. It is unclear how United will identify agents, which interfaces the restriction will affect or how it will handle disputed cases. It would therefore be premature to conclude that the airline has already learned to reliably distinguish humans from automation.

Beyond safety, there's the question of intermediaries

The decision also has a commercial dimension. United CEO Scott Kirby said he does not want to create another intermediary. That position matters for founders: even a safe agent may face resistance from a company that wants to maintain a direct relationship with its customers.

If an assistant selects a supplier, compares offers and guides the user toward a purchase, some control over the customer experience shifts to the assistant. Kirby's remarks suggest that United is concerned about this scenario too. But it would be wrong to portray commercial considerations as the sole reason: the company separately points to passenger protection and checkout requirements.

For now, this is not about a large volume of automated purchases. According to the report, United has not yet seen many attempts to book through agents. The airline is therefore setting boundaries in advance. The article also describes restrictions on agents at banks and eBay, although their terms cannot be assumed to apply to United without additional information.

What to change in your product now

For a team building a travel assistant, the practical takeaway is to design the handoff to a human at checkout in advance. The agent gathers suitable flights, explains the differences and saves the selected option. The user then continues the purchase with the airline. How smooth that transition can be depends on the available interfaces; developers cannot promise that every field will be preserved without checking the integration.

That handoff should also be reflected in the product's promises. "We'll find a flight and help you book it" describes the available workflow more accurately than an unconditional "we'll buy the ticket for you." Teams also need to consider what the user will see if access is blocked midway through the task: which results will be saved and where the user can resume on their own.

United is open to reconsidering its position as safeguards evolve, but has given no timeline or specific criteria. For developers, that currently means something simple: they need to test not only whether the model can perform an action, but also whether a particular service is willing to accept that action from an agent. That determines which promises an assistant can actually deliver on.