United's new rule bans AI agents from buying a ticket even when the passenger watches the agent do it and clicks the button — and that refusal to carve out the supervised case is the most useful sentence any platform has written about agents this year. It tells you the airline knows it cannot detect an agent, so it stopped trying, and moved the control to the one thing it can do unilaterally: cancel the booking afterwards, without notice.
What the rule actually says
United's revised Contract of Carriage is dated 11 November 2026. Rule numbers below are as reported by the outlets that covered the change; the operative text is the contract itself, and the reporting was consistent on every substantive point.
| Change | Where | What it does |
|---|---|---|
| AI-agent use added to prohibited ticketing practices | Rule 6.J.6 | Using an agent to make, modify or service a ticket, or to buy any product or service, is a breach |
| Cancellation without notice | Rule 5.H | United may cancel a reservation an agent made or managed, without telling anyone first |
| Definition of the thing banned | Contract definitions | Reaches AI assistants, standalone agents and AI-enabled browsers |
| No supervision carve-out | — | Applies even where the passenger oversees or approves the agent's actions |
| Reaches intermediaries | — | Applies to agents offered through travel agencies and aggregators, not only united.com |
| Fare search untouched | — | Agents may still show United fares and availability; the customer checks out |
| Grandfathered | — | Agent-made bookings before 11 November 2026 are honoured |
Read as a list of prohibitions it looks like a familiar anti-bot measure. Read for what is missing and it is something else: there is no identification requirement, no registration path, no approved-agent list, and — as every outlet covering it noted — no stated method of detecting agent activity at all.
The supervised agent is the tell
Every reasonable draft of this policy contains an exception for the supervised case. An agent that assembles the itinerary and hands it to a human who reads it and confirms is, by any sensible account of responsibility, a human booking a flight with a tool. United covered it anyway, and the reason is not that the airline thinks oversight is worthless.
It is that oversight arrives as an assertion from the party being regulated. It is carried in no header, signed by nobody, and backed by no artefact the server can look up. At the moment the booking request lands, a supervised agent and an unsupervised one are byte-identical: same browser, same session cookie, same customer credential, same TLS fingerprint. The distinction United declined to draw is a distinction it has no instrument to measure.
This is why the gap is a protocol gap rather than a policy failure. An oversight claim becomes checkable only when something outside the agent attests to it — a signature over the request naming the principal and the grant it acts under, as in Web Bot Auth and signed agent requests, or a consent record the platform itself issued and can query. Neither is on the wire for a browser-resident assistant today. Until one is, the cheapest correct response to an unfalsifiable claim is to ignore the distinction it asserts, which is exactly what the contract does.
Nobody writes an exception for a promise they cannot check. The agents that will eventually get carve-outs are the ones that arrive with something verifiable attached.
Enforcement moved to the layer United can undo
Detection was never on the table
A browser-resident agent presents the platform's own client, the customer's own session and the customer's own credential. There is no signal of automation to find, because from the server's position there is no automation — there is a logged-in customer whose clicks happen to be generated. The only platforms that can meaningfully detect agents are the ones whose agents must announce themselves to work at all, which is a property of a gateway you built, not of traffic you receive.
Teams that miss this build a detection pipeline, discover a false-positive rate they cannot defend when it cancels a real family's holiday, and quietly stop enforcing. The unenforced rule is worse than no rule, because it is now documentary evidence that you knew.
A void is the only control that survives an unobservable predicate
If you cannot see the thing you are prohibiting at the moment it happens, your options collapse to one: make the action undoable, and undo it when you later form the view that it broke the rule. That is what Rule 5.H is. It is not a security control, it is a remedy, and it works precisely because it does not require detection at the time of the request — a complaint, a pattern noticed a week later, or a partner's tip-off is enough.
The layer choice follows mechanically. Blocking at the edge costs nothing and achieves nothing against a real browser. Blocking at authentication is what JPMorgan did — on its 13 October 2026 earnings call the bank said it will not let outside agents log in to customer accounts or act on a customer's behalf, framing it as liability rather than security — and it works to the extent that the agent needs a credential the bank issues rather than one the customer hands over. Blocking at settlement means a chargeback, a refund path and a regulator, and the cost of reversal stops being yours alone to set. The booking sits in the sweet spot: consequential enough to matter, cheap enough to unwind, entirely within United's gift.
Stated as a rule for anyone drafting one of these: put the prohibition one layer upstream of the first irreversible step. Which is also the test for whether you have a policy at all. If your agent rule sits downstream of the first thing you cannot undo, you have a preference.
What the open read path tells you
United left fare search alone, and that is not leniency. A read has no consequence to reverse, so a prohibition there would be unenforceable for exactly the same reason detection is: nothing to undo and nothing to observe. It is also commercially free — an agent that shows a customer United's fares is doing distribution the airline would otherwise pay for.
The asymmetry generalises. Across the platforms that have moved, the read path stays open and the write path closes, and the closure lands wherever the platform holds the instrument. Walmart's approved-agent gateway, announced 21 October 2026, is the other shape of the same judgement: an agent may search and buy, and never receives card details or the ability to change a payment method. One platform refuses all agents and voids their bookings; the other admits named agents and keeps the card. Both have concluded that the thing they will not do is let an agent hold the credential.
What neither has published is a door for the unaffiliated agent acting for its own user. United says it works with no third-party AI agents at all. For an airline with its own app and loyalty programme that is coherent; for most platforms it selects for the worst traffic mix, because the agents willing to identify themselves stop — identifying yourself is now self-incrimination — while the ones that were never going to identify themselves carry on, wearing your customers' logins.
What to change this week
If you run a platform:
- Find your first irreversible action for a logged-in user, and check whether your terms say anything about agents upstream of it. Write the consequence — void, hold for review, or honour and bill — before you write the prohibition.
- Decide what a void must be supported by, and log it at the time of the action: which clause, which specific observation, what a human doing the same thing would have looked like, and who can un-void it within what window. The first month's records will be mostly false positives, and that is the number that tells you whether the rule is enforceable.
- Publish the door with the ban, or accept that you have selected for unidentifiable traffic. A scoped token, a rate and a revocation switch convert an unobservable predicate into an observable one, because a sanctioned agent then has something to lose.
If you build agents:
- Stop treating the read path as a foothold. It is open because it is consequence-free, and building a booking flow on top of an open search endpoint is building on the part of the policy that was never the point.
- Make your oversight claim verifiable by someone else. A confirmation dialog your own product renders is not evidence to a third party. A signed request naming the principal and the specific grant is. That is the difference between being the agent a platform writes an exception for and the agent it writes a rule about.
- Assume the grandfather clause is the generous version. United honoured bookings made before the rule changed. Plan for the version that does not, and make sure your user finds out from you rather than at the gate.
FAQ
Did United ban AI agents from its website entirely?
No. The restriction covers making, modifying or servicing a ticket and buying products or services. Agents may still search and display United fares and availability; the customer is expected to complete the purchase themselves.
Does the ban apply if I supervise the agent and click the confirm button myself?
As reported, yes — the rule applies even where the passenger oversees or approves the agent's actions. That is the clause that reveals the policy is about an unverifiable claim rather than about autonomy.
How would United know an AI agent made my booking?
It has not said, and for a browser-resident agent there is no reliable signal: the request carries the customer's own session and credential. That is why the enforcement mechanism is a cancellation right exercised after the fact rather than a block at the time of booking.
What happens to a booking that was made by an agent?
Under the cancellation provision United may cancel it without notice. Bookings made before 11 November 2026 are reported as honoured, and the airline advised passengers to check such reservations for errors.
Is this the same thing JPMorgan did?
Same goal, different layer. JPMorgan refused outside agents the ability to log in to customer accounts or act on a customer's behalf — enforcement at authentication. United left authentication alone and took the right to void the resulting action instead.
Would Web Bot Auth or a signed request have earned an exception here?
Not on its own. A signature establishes which operator sent the request, not that a specific human authorised this specific purchase. Closing the gap needs the grant itself to be attested — a consent record the platform issued, or a signed statement naming the principal and the action — which is the part no deployed protocol currently carries for a browser assistant.
Further reading
On this wiki:
- Admitting or refusing third-party agents — the full operational version of this argument, including the policy artefact and what a void has to be supported by.
- Platform terms for agent access — the same boundary read from the agent builder's side.
- Bot verification and agent access — what the available signals actually tell you.
- Third-party agents logging in as your users — the credential-sharing problem United cites.
- AP2 and agent commerce — the standardisation attempts aimed at this gap.
- Travel and booking agents — what the counterparty is trying to accomplish.
Sources:
- Bloomberg — United blocks AI agents from booking flights (11 November 2026)
- United Contract of Carriage — the operative text
- Forrester — United Airlines bans unauthorized AI agents