AI Agents Booking Flights: United’s Risk Line 🚀 (2026)12 min read2,452 words

AI Agents Booking Flights: United’s Risk Line 🚀 (2026)

Discover how AI travel assistants can search flights—and why United limits airline booking automation. Learn AI agent transaction risks and agentic AI steps.

AI agents booking flightsAI travel assistantsairline booking automationAI agent transaction risksagentic AI in business
AI Agents Booking Flights: United’s Risk Line 🚀 (2026)

Why United Won't Let AI Agents Book Your Flight

AI assistants are getting good at finding options, comparing prices, and handling the easy stuff. But once you move from "help me decide" to "help me buy," the line between useful automation and a real risk gets much harder to draw. United Airlines says that line runs right at the booking step: an AI agent may show you available flights, but it can't complete the purchase for you.

United points to privacy, the chance of unintended purchases, and the responsibilities that come with a booking. In air travel, a ticket isn't just a payment - it's a bundle of details a traveler has to review and confirm. That includes required disclosures, such as an attestation related to explosives in checked baggage. If an agent clicks through without the right authority - or without clear confirmation - it can create problems for both the customer and the airline.

This is also part of a bigger question: when should companies let third-party AI agents act on their websites and accounts? Here's what United's stance suggests, why airline purchases carry more stakes than ordinary searches, and what "safer agent access" can look like if you don't treat it as an all-or-nothing switch.

What United's decision allows - and blocks

United's approach is narrower than banning AI from travel entirely. As reported by The Wall Street Journal, an AI assistant can still help a user find United flights. The restriction is on finishing the transaction: the agent can't take the user through the booking flow and buy the ticket on their behalf.

That distinction matters. Search and comparison mainly involve showing information. Booking changes a customer's financial and travel situation. It can include payment details, personal information, fare and baggage choices, and confirmations that have legal or operational weight. In other words, it makes sense for a company to allow one step while drawing a harder line at the next.

United says it hasn't seen many customers trying to use AI agents to book flights yet. So this isn't only reacting to misuse - it's also about setting a boundary before the practice spreads. The company's concern is that agents could expose customer data, make purchases the user didn't intend, or mishandle the disclosures and attestations built into the process.

The airline also left the door open to change as third-party agents get better and safeguards improve. In plain terms: this doesn't read like "agents can never book." It reads like "the risk controls and responsibilities aren't clear enough yet for United to allow direct booking access."

United CEO Scott Kirby has also questioned the value of adding another intermediary between the airline and customers. If an outside assistant controls the interaction, United may lose visibility into how the customer makes choices - while still being responsible for the booking and the service that follows.

For customers, the immediate takeaway is straightforward: use an assistant if you want help researching, but plan to review and complete the booking through an authorized process. For businesses, the more useful lesson is that agent access can be broken into permissions instead of treated as a single yes-or-no decision.

Why booking a flight is different from finding one

An AI agent that finds a flight is recommending. An AI agent that books is acting. That switch changes what a business needs to guarantee. The system can't just produce a reasonable answer - it has to carry the user's intent through a series of choices and make sure the final step is actually authorized.

Airline purchases also mix different kinds of information and obligations. A traveler may need to select a route, date, fare class, and payment method, then review rules and provide required details. Some of those steps are easy to misunderstand, and some consequences show up later. A cheap fare might come with restrictions. A traveler might pick the wrong date. An automated flow might skip a required confirmation. These aren't bizarre edge cases - they're normal mistakes, and they become more serious when the system completes the purchase without a person checking the final screen.

United's mention of the baggage attestation gets right at the difference between filling out a form and making a declaration. The traveler is the person who knows what's in their bags and is responsible for answering accurately. An agent can surface the question, but it can't independently verify whether the answer is true. If the system submits a response based on an assumption or a default, it turns uncertainty into something that looks official.

There's also an authorization problem. "Find me a flight to Chicago next week" isn't automatically the same as "buy ticket X for Y dollars under Z conditions." The agent may have to choose among airports, times, fare rules, and prices. Without a clear approval point, it can satisfy the broad request while choosing something the customer wouldn't have picked on their own.

The comparison that matters most isn't a human assistant. Humans can ask follow-ups, notice context, and confirm intent. It's a software flow that has to recognize ambiguity, pause when needed, and distinguish between permission to research and permission to spend. The more expensive - or harder to undo - the action is, the stronger the boundary should be.

The business risks behind agent access

United's concerns show up as a set of risks that apply well beyond airlines. A third-party agent might interact with a company's site while using a customer's account or information. If that access isn't controlled, sensitive data could be exposed, stored in the wrong place, or used in ways the customer didn't expect. The company has to think not just about what the agent can see, but what it might keep - and what it might do with what it sees.

Another risk is unintended action. Agents can misunderstand instructions, miss a constraint, or pick an option that seems reasonable from the data but is wrong for the user. In a shopping cart, that could mean the wrong item or quantity. In travel, it could mean the wrong date or a fare with restrictions that don't fit the traveler. Even if a third-party assistant initiated the action, customers still tend to blame the airline or merchant.

A third issue is accountability. When something goes wrong, customers want to know who can fix it. The airline may have made the ticket available, the agent may have selected it, and the customer may have approved only a broad request. If the roles and boundaries aren't clear, support teams end up in disputes that are hard to investigate and expensive to resolve.

Companies also need to protect service integrity. If an agent can browse or transact in ways the site wasn't built to handle, it may create strange traffic patterns, bypass checks, or interact unpredictably with account features. Other businesses have reportedly taken a cautious approach too - reporting has described limits on unauthorized agents at companies including eBay and major banks. The details differ, but the underlying point is similar: letting something access a customer-facing interface isn't automatically the same as giving permission to automate every action on it.

A practical risk review should focus on questions like:

  • What customer data can the agent view, and does it need all of it?
  • Which actions are read-only, and which can change an account or spend money?
  • Where does the customer have to review and approve decisions?
  • Can the business identify what the agent did and explain it after the fact?
  • What happens when the agent runs into ambiguity or a failed transaction?

These questions are more useful than arguing about whether a company is "pro" or "anti" AI agents. The real decision is which actions are safe to delegate, and under what controls.

What this signals about the agent economy

Personal AI assistants promise something simple: you describe what you want, and software does the steps. Instead of bouncing between sites, comparing options, and filling out repetitive forms, a user could hand the job to an assistant. Products such as Meta's Muse and OpenAI's Dots have been discussed as part of that broader category - moving from answering questions to completing tasks.

But useful agents depend on cooperation from the businesses they interact with. They need access to information, the ability to maintain a session, and sometimes the ability to submit data or handle payments. Businesses, in turn, need to know who is acting, what the customer approved, and how to limit what the agent can do. If every company blocks unapproved automation, agents might remain helpful for research but stop being useful at the moment they're supposed to save time. If companies allow unchecked access, they take on privacy, fraud, and customer-service risks they may not be ready for.

This tension isn't just technical. It's also about distribution. If assistants become the main way people search and buy, they may shape what options users see and how those options get compared. Businesses may worry about losing the direct relationship with customers. Customers may prefer a neutral assistant that helps them decide. United's concern about creating another middleman reflects that commercial reality as well as the safety concern.

The most likely path isn't unlimited agents and it isn't permanent exclusion. It looks more like narrower permissions and clearer handoffs. Businesses can let agents search public information, then require a user to sign in, review key terms, or approve payment. They can also build supported ways for agents to identify themselves and operate within defined limits, instead of relying on improvised behavior inside a website.

The catch is that a permission model only works if the boundaries are clear. A customer should be able to tell whether an assistant is just searching, preparing a booking, or actually authorized to buy. A business should be able to stop or reverse an action when it needs to - and figure out what happened when things go wrong. Without that foundation, agent adoption can simply shift work to customer support rather than removing it for the user.

A practical framework for enabling AI agents

If a company is deciding whether to accept agent activity, it helps to focus on the action - not on the label. A human-facing assistant, browser automation, and a purpose-built agent can all create similar risk if they can access an account and change things. And an agent that only reads public flight information has a different risk profile than one that submits payment details.

A workable policy can be built around four levels of permission:

  1. Discover: The agent can search public information and show options. It cannot access an account or change anything.
  2. Prepare: The agent can assemble a draft or proposed transaction, but it cannot submit it.
  3. Confirm: The user reviews the exact choice, price, terms, and required declarations, then gives explicit approval.
  4. Execute: The agent completes the approved action within a limited scope, with a record that can be reviewed later.

Not every company needs all four levels, and not every task should reach the final stage. The point is to make the progression obvious. A business might allow an agent to compare flights and prepare a selection, but require a human to confirm fare rules, payment details, and any baggage declarations. That keeps some convenience without asking the system to make the high-stakes decision by itself.

Before moving from one level to the next, define the controls. Limit what data the agent can access to what the task requires. Set transaction thresholds when that's relevant. Require clear confirmation when the choice is ambiguous, expensive, or hard to reverse. Keep an audit trail that connects the user's instruction, the options shown, and what actually happened. And make sure there's a support path that doesn't turn the user into a messenger between the company and the agent provider.

On the user side, the same framework can guide product decisions. An assistant should show what it plans to do before it does it, explain the trade-offs that actually matter, and ask for confirmation when broad intent becomes a specific commitment. If it can't verify a required answer - like a declaration about what's in a traveler's bags - it should ask the person rather than guess.

Key takeaways:

  • Treat search, preparation, and purchase as separate permissions.
  • Require explicit approval for choices with financial, legal, or operational consequences.
  • Give agents only the data and access the task truly needs.
  • Make actions traceable so customers and support teams can understand what happened.
  • Pause and ask when the user's instruction - or a required declaration - is unclear.

Conclusion

United's restriction boils down to a simple idea: automating a transaction is not the same as automating research. Agents can reduce effort, but the business still has to protect customer data, preserve intent, and handle required confirmations properly. If you want a durable path forward, grant access in stages - let agents find options, limit what they can prepare, and require clear approval before anything consequential happens. For both customers and companies, that's a more stable starting point than either unrestricted delegation or a blanket rejection of AI.

Frequently Asked Questions

United’s stated position is that third-party AI agents may show users available flights but cannot complete a booking on their behalf. The company has cited privacy, unintended purchases and regulatory concerns, while leaving open the possibility of reconsidering as protections develop.
Searching presents information; purchasing commits the customer to a transaction. Booking can involve payment, personal data, fare conditions and required declarations, so United keeps a person involved in decisions and confirmations with greater consequences.
United has pointed to information it must share with customers and attestations travelers must make, including a declaration related to explosives in baggage. An agent cannot independently know whether a customer’s answer is accurate.
Possibly. United says it sees potential in third-party agents but wants protections before they interact directly with accounts or bookings. Explicit approval, limited access and traceable actions could help businesses assess the risks.

Related articles