United’s AI Booking Restriction: Build the Human Handoff

▶ Watch the video version of this article
Your travel assistant can find a flight, collect an approval, and still be unable to book the ticket. If your workflow calls that task “done” before a reservation is confirmed, someone may arrive at the airport with an itinerary brief instead of a flight.
CBS News reported on November 11, 2026 that United Airlines is banning AI agents from booking flights on travelers’ behalf. The practical question for a small business is not whether an agent can click through checkout. It is how to finish the work reliably when that step is off-limits.
What United’s announcement does—and does not—settle
United says it is taking steps to prohibit unauthorized third-party personal AI agents or assistants from accessing its services. Its announcement, titled “United Restricts Access for Third-Party AI Agents,” says the restrictions cover United’s digital platforms as well as access through travel agencies and aggregators.
That scope matters. Moving an agent from United’s website to an intermediary is not automatically a permitted workaround. United also says it is updating its booking, ticketing, and distribution policies as part of these restrictions. A builder needs to check the current terms and the authorization for the specific route an integration uses before deploying it.
The announcement does not establish that all AI-assisted trip planning is banned. Nor does the available reporting tell us exactly how United detects agents, which integrations may be authorized, or whether every attempted agent booking is technically blocked. Those are questions to resolve with the provider or an authorized integration partner—not assumptions to bury inside a prompt.
For a small business, the immediate design decision is straightforward: do not make an unauthorized agent-operated purchase the only path to completion. The assistant can still have useful work to do, provided it gets travel information through permitted sources and hands checkout to a person or a separately confirmed, authorized channel.
Separate permission from approval
An employee approving a trip and a provider allowing an agent to operate its booking service are different permissions. You need both internal authority to spend company money and an allowed method of interacting with the external service; one does not grant the other.
Imagine a client-facing employee needs to travel for a meeting. Your workflow collects the destination, arrival deadline, baggage needs, and company budget. A manager approves the selected itinerary. That approval answers, “May we buy this trip?” It does not answer, “May our agent sign in to United, use the traveler’s account, and complete checkout?”
Treat those as separate gates:
| Gate | Question | Evidence to retain |
|---|---|---|
| Business approval | Has the appropriate person approved the itinerary and expense? | Approver, selected option, decision, and time |
| Provider permission | Is this method of accessing and acting on the service allowed? | Current terms or integration authorization reviewed for that route |
| Purchase completion | Did the provider actually issue a reservation? | Booking reference and confirmation reviewed against the traveler and itinerary |
United specifically warns about sharing account credentials with personal third-party AI assistants because it can expose sensitive information. The reported risks also include unintended purchases, account changes, and use of MileagePlus miles. That makes “give the browser agent the traveler’s login” a poor fallback even before you consider whether the access is permitted.
A human approval button cannot fix a prohibited checkout method. Conversely, using an authorized booking integration would not remove the need for your company’s spending approval. Build both checks into the workflow rather than asking the model to infer them from a message such as “Looks good.”
Make the booking brief useful without making it a booking
A booking brief is a handoff document: it gives the traveler what they need to complete checkout without pretending that a seat has been reserved. The brief should make the remaining action unmistakable and keep time-sensitive information labeled as such.
Here is a practical format. The entries below are illustrative, not a live fare or availability claim:
traveler:
name: "Name as supplied for travel"
trip:
departure_airport: "Origin airport"
arrival_airport: "Destination airport"
departure_date: "Requested date"
required_arrival: "Before the client meeting"
baggage: "Carry-on and checked-bag requirements"
selection:
itinerary: "Chosen flight details"
source_link: "Link obtained through a permitted source"
price_checked_at: "Timestamp of the observed price"
displayed_price: "Price observed, not guaranteed"
approval:
owner: "Person authorized to approve the expense"
status: "Approved"
next_action:
owner: "Traveler"
instruction: "Review the current fare and complete checkout personally"
status_after_click: "Awaiting human booking"
The source link is a navigation aid, not proof that a fare is held. Prices, availability, baggage terms, and flight times can change between preparation and checkout. The traveler should review the live details before paying; the workflow should not silently substitute a different itinerary if the original option disappears.
Only collect data the preparation step needs. A planning assistant usually does not need an airline password, payment card details, or access to a loyalty account. If the traveler must enter sensitive information at checkout, that can remain inside the provider’s normal checkout flow rather than passing through your AI system.
The brief also gives the traveler a way to catch a common automation error: the almost right flight. A destination city may have multiple airports; a flight that lands on the meeting date may still land too late; a low displayed fare may not match the required baggage. Put those constraints next to the selected itinerary so a person can compare them before purchase.
Use a state machine that cannot confuse a click with a ticket
A reliable travel workflow needs distinct prepared, awaiting human booking, and confirmed states. Only evidence of a provider-issued reservation should move the task to confirmed; opening a checkout link, approving a budget, or receiving an agent’s “success” message is not enough.
Here’s exactly how I would model the boundary. This Python example is a workflow guard, not a United integration and not a complete confirmation-verification system:
from dataclasses import dataclass
from enum import Enum
class Status(str, Enum):
PREPARED = "prepared"
AWAITING_HUMAN_BOOKING = "awaiting_human_booking"
CONFIRMED = "confirmed"
@dataclass
class Trip:
traveler: str
selected_itinerary: str
status: Status = Status.PREPARED
booking_reference: str | None = None
confirmation_reviewed: bool = False
def hand_off_for_checkout(trip: Trip) -> None:
if trip.status is not Status.PREPARED:
raise ValueError("Trip is not ready for handoff")
trip.status = Status.AWAITING_HUMAN_BOOKING
def record_confirmation(
trip: Trip,
*,
booking_reference: str,
provider_confirmation_reviewed: bool,
) -> None:
if trip.status is not Status.AWAITING_HUMAN_BOOKING:
raise ValueError("Checkout handoff has not occurred")
if not booking_reference.strip() or not provider_confirmation_reviewed:
raise ValueError("Provider confirmation must be reviewed")
trip.booking_reference = booking_reference.strip()
trip.confirmation_reviewed = True
trip.status = Status.CONFIRMED
The important part is the transition rule, not the amount of code. In a working system, provider_confirmation_reviewed must represent an actual verification step—not a Boolean the agent sets because it believes checkout succeeded. Depending on the permitted channels available to your business, that step might involve a person checking the provider’s confirmation or a verified confirmation received through an authorized channel. A random email containing a reference-shaped string is not sufficient evidence.
Keep an owner on every open task. If the traveler has not completed checkout, the task remains awaiting human booking and the owner receives a reminder. If checkout fails or the fare changes, record the reason and return the decision to the traveler or approver. Do not automatically try another checkout route merely because the first one was blocked; changing the route can change the permission question.
The same state-machine pattern applies outside travel. A drafted invoice is not sent. An approved lead response is not delivered. A prepared account change is not applied. Name the external event that proves completion, and prevent earlier events from borrowing its status.
Test the failure before anyone travels
The most valuable test is to stop the workflow immediately before checkout and verify that it does not send a “trip booked” notification. This requires no purchase and exposes the exact failure that polished agent demos tend to hide.
Run the test with a sample trip. Have the assistant prepare the brief, collect the internal approval, and hand it to the traveler. Then do nothing. The task should remain open, visibly assigned, and labeled awaiting human booking. Accounting should not receive a booked-travel receipt, and the traveler should not receive a confirmation message.
Next, test the less tidy paths:
- The traveler opens the link but does not pay. Status stays unchanged.
- The fare or itinerary changes at checkout. The traveler chooses whether to proceed; the system does not claim the earlier selection was purchased.
- A message arrives that appears to be a confirmation but lacks a reviewable provider reference. Status stays open.
- A valid confirmation is reviewed, but its traveler or itinerary does not match the brief. The discrepancy goes to a person.
- No confirmation arrives. The owner is notified rather than letting the task disappear from a dashboard.
For measurement, track prepared trips, handed-off trips, and confirmed reservations separately. A high preparation rate can be useful: it may save research and coordination work. It is still not a booking-completion rate. This distinction lets a small business judge whether its business automation removes work or merely moves it to an untracked human inbox.
Also decide what happens after confirmation. With appropriate access and permission, a workflow can route the receipt to accounting, update an internal travel record, and notify the team that the trip is confirmed. Those follow-up steps are where workflow automation earns its place—but they should trigger from reviewed confirmation, not from the assistant’s prediction that a purchase probably worked.
Where bizflowai.io fits
At bizflowai.io, my work is building AI automations that ship as working systems, not stopping at a generated recommendation. For a workflow like this, that means making the handoff explicit: prepare the information, assign the human action, keep the task open, and only mark it complete when the required evidence exists. It does not mean claiming an agent can book through a provider that has not authorized its access.
Want more like this?
BizFlowAI is dedicated to practical AI automation for solopreneurs and small teams. I’m Lazar Milićević, a senior engineer who builds AI automations that ship.
For more about this work, visit bizflowai.io.
Frequently asked questions
What is United Airlines’ reported policy on AI agents booking flights?
CBS News reports that United Airlines is banning AI agents from booking flights on behalf of travelers. The report, as described here, does not establish an industry-wide ban or say that all AI-assisted travel planning is prohibited. It also does not explain how United detects agents or whether exceptions exist. Builders should check what interactions are permitted before connecting an agent to a booking process.
How do I handle a flight booking that my AI assistant cannot complete?
Have the assistant prepare an itinerary brief and make the next action explicit. The brief can include the traveler’s details, dates, budget, selected itinerary, source link, and the time the price was checked. The traveler then completes checkout through a permitted route. Keep the task open until a provider confirmation supplies a booking reference; if that does not arrive, notify the task owner.
Why does employee approval not give an AI agent permission to book a flight?
Expense approval and permission to use a provider’s booking service are separate checks. An employee may approve an itinerary and its cost, but that does not authorize an AI agent to use a booking method the provider prohibits. A human approval button cannot fix a prohibited integration. The workflow needs both internal authorization and a permitted way to interact with the provider.
When should I mark an AI-assisted travel task as prepared versus confirmed?
Mark the task as prepared when the assistant has assembled the booking details but no purchase has been completed. Use awaiting human booking while the traveler still needs to check out. Mark it confirmed only after receiving a booking reference from the provider’s confirmation. Clicking a booking link or receiving a success message from the assistant is not enough to establish that a ticket was booked.