8 min read

C28
Operation · Governance & Compliance

Admitting or Refusing Third-Party Agents

You cannot tell whether an agent is holding the mouse, so your agent-access policy is only worth the layer you can reverse.

Write the rule at the HTTP layer and you have published an evasion target; write it at the layer where you can undo the action and you have published an enforceable one. United's revised Contract of Carriage, dated 11 November 2026, is the clearest worked example yet: it bans third-party AI agents from buying, changing or servicing a ticket, keeps fare search open, and — the part that makes it work — reserves the right to cancel any reservation an agent touched, without notice. The ban is not a bot detector. It is a voidability clause, and that is the only kind of agent policy a platform can currently keep.

Step 1Start by conceding the detection question, because you have already lost it.

Every agent-access policy written in 2026 rests on a predicate the platform cannot observe. United's definition reaches AI assistants, standalone agents and AI-enabled browsers, and — the detail that gives the game away — it applies even when the passenger oversees or approves what the agent does. That is a rule about who moved the pointer inside a session that otherwise looks exactly like a logged-in customer on a supported browser. Reporting on the change was unanimous on one point: the airline did not say how it intends to spot agent activity.

This is not an oversight in the drafting. It is the honest consequence of the fact that a browser-resident agent presents your own client, your own session cookie and your own customer's credential. Bot verification gives you signals about declared automation; it says nothing about an agent that never declares anything because it is driving a real user's real browser. Treat the predicate as unobservable and the design question changes shape: not "how do I detect this?" but "what do I do about an action I will only learn was an agent's afterwards, if ever?"

A policy whose trigger you cannot observe is not an access control. It is a basis for voiding a transaction later. Teams that miss this build a detection pipeline, watch it produce a false-positive rate they cannot defend against real customers, and quietly stop enforcing — which is worse than having published nothing, because the unenforced rule is now evidence you knew.

Step 2Pick the enforcement layer by its reversal cost, not by where the traffic arrives.

There are four places you can put the rule, and they differ on exactly one thing that matters: what it costs you to undo an action after you decide it should not have happened.

  • Edge / HTTP — cheapest to enforce, nothing to reverse, and nothing to enforce against: a browser agent is indistinguishable from the browser. Useful only against declared crawlers.
  • Authentication — this is where JPMorgan drew it. 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, and framed it as liability rather than security. Enforceable if the agent must use a credential you issue, which is why the sanctioned shape is always a token you can scope and revoke rather than a password the customer can hand over.
  • The action — the reservation, the order, the ticket. Reversible by you, unilaterally, at a cost you can calculate. This is where United put it, and it is the layer most platforms should use.
  • Settlement — money that has moved. Reversal here is a chargeback, a refund path and a regulator, and the cost is no longer yours alone to set.

The rule of thumb falls out directly: place the prohibition one layer upstream of the first irreversible step. United's agents may still display fares and availability — the read path costs nothing to allow and is not reversible-or-not, it simply has no consequence. Everything from the booking onward is covered, because the booking is the first thing the airline would have to unwind. Compare platform terms for agent access, which is this same boundary read from the agent builder's side of the table.

Step 3Write it as three separate decisions, because one sentence collapses them.

"No AI agents" reads like a single policy and is in fact three, each of which can be answered differently and each of which needs an owner. Keeping them apart is what lets you loosen one without reopening the others.

# agent-access.yml — three decisions, three owners, one review date
admission:                       # WHO may act through us
  default: deny
  sanctioned:
    - id: partner-agents         # named, contracted, credentialed by us
      identity: signed-request   # how WE recognise them, not how they describe themselves
      credential: scoped-token   # never the customer's password
  unsanctioned: tolerated-read-only

scope:                           # WHAT a sanctioned agent may do
  read:   [search, price, availability]
  write:  []                     # the expensive list — start empty, add with a contract
  never:  [change-payment-method, redeem-loyalty-balance]

consequence:                     # WHAT HAPPENS to an action that broke the rule
  action: void                   # void | hold-for-review | honour-and-bill
  notice: none-required          # the clause that makes 'void' actually available
  grandfather_before: 2026-11-11 # actions already taken are honoured
  appeal: human-queue            # mandatory if 'void' can hit a real customer

The consequence block is the one teams forget to write, and it is the only block that does any work on the day it matters. United's version is unusually complete: a prohibition, a cancellation right exercisable without notice, and an explicit grandfather date before which agent-made bookings are honoured. The grandfather line is not generosity — it is what stops the policy from being retroactive, which is what would have made it indefensible.

Step 4"A human approved it" is not a claim you can check, which is why supervised agents get banned too.

The most instructive clause in the United change is the one that refuses the obvious carve-out. A supervised agent — one that drafts the booking and waits for the passenger to click — is exactly what a reasonable policy would permit, and it is covered by the ban anyway. Ask why and the answer is not that United thinks supervision is worthless. It is that supervision arrives as an assertion from the party being regulated, carried in no field, signed by nobody, and indistinguishable at the server from no supervision at all.

This is the real gap in the agent stack, and it is a protocol gap rather than a policy one. An oversight claim becomes checkable only when something outside the agent attests to it: a signature over the request that names the principal and the grant it is acting under (signed agent requests), or a consent record the platform itself issued and can look up (delegated access and consent records). Until one of those is on the wire, every platform that wants to distinguish supervised from autonomous agents has to take the agent's word for it, and the cheapest correct answer to a claim you cannot verify is to ignore the distinction.

If you are the agent builder reading this as an injustice: the fix is not a better User-Agent string. It is being the one agent in your category whose requests carry a verifiable statement of who authorised this specific action, so that a platform drafting a ban has a reason to write an exception and a mechanism to enforce it. Nobody writes an exception for an unfalsifiable promise. See human in the loop and approval and confirmation UX for what a gate has to emit to be worth anything to a third party.

Step 5Publish the sanctioned path in the same document as the prohibition.

A ban with no door produces the worst possible traffic mix. The agents that would have identified themselves stop, because identifying themselves is now self-incrimination; the agents that were never going to identify themselves carry on unchanged, wearing your customers' credentials. You have selected for the population you least wanted and lost your only channel for talking to the other one. As of November 2026 United says it works with no third-party AI agents at all, which is a coherent position for an airline with its own app and loyalty programme, and an expensive one for most platforms to copy.

The alternative shape is an approved-agent gateway: the platform holds the payment instrument, issues the agent a scoped credential, publishes a rate, and keeps a revocation switch. It costs real engineering and it converts an unobservable predicate into an observable one, because a sanctioned agent now has something to lose. Operationally that means three things you must own before you announce it — an onboarding queue with a named approver, per-agent rate and scope enforcement you can change without a deploy, and the revocation path tested on a live partner. See travel and booking agents and shopping and checkout agents for what the counterparty is actually trying to accomplish, and AP2 and agent commerce for the standardisation attempts.

Step 6Instrument the void, or you will not be able to defend the one that hits a real customer.

The day your policy does something is the day you cancel a booking a human being was relying on. If your answer to "why was this one voided?" is a model score or an analyst's judgement, you have a customer-service incident and a regulatory one. Decide now what a void has to be supported by, and log it at the moment of the action rather than reconstructing it later from traffic you no longer have.

  • The basis — which rule, which clause, and the specific observation that triggered it. "Automation signals" is not an observation.
  • The alternative — what a human doing the same thing would have looked like, so the false-positive case is visible in the record rather than only in the complaint.
  • The notice — what you told the customer and when, even where the clause says none is required. Exercising a no-notice right silently is legal and reputationally expensive.
  • The reversal — who can un-void, within what window, on what evidence. A void with no documented undo becomes permanent by default at 2am.

Route all four into the same place as the rest of your audit trails, and expect the first month's records to be mostly false positives — that is the signal that tells you whether the rule is enforceable at all. Related from the receiving end: third-party agents logging in as your users.

Do this today: open your terms and find the first irreversible action a logged-in user can take. If your agent rule sits downstream of it, or there is no rule, you have no policy — you have a preference. Write the consequence block first, get it reviewed by whoever would have to defend a cancellation, and only then decide what you want to prohibit. The prohibition is the easy half; the right to undo is the whole thing.