ICA · Institutional Control Architecture
An essay from ALEETH  ·  November 2026
ALEETH · Essay

The Mandate Gap

United barred AI agents from buying its tickets. Underneath its stated reasons is authority nobody can check. ICA gives an agent a mandate an airline could verify itself.


On November 11, 2026, United Airlines revised its Contract of Carriage. Its list of prohibited ticketing practices, the same list that bars throwaway tickets, now includes this line: The use of AI Agents to make, modify or service a Ticket, and/or to purchase any service or product. An AI assistant can still find a United fare. It can no longer buy one. United's stated reasons are shared credentials and unintended purchases. Our reading: underneath them is a problem of authority that nobody at the counter can check.

The Rule

A wide definition, on purpose

United's contract defines an AI Agent as any software or service that takes autonomous or semi-autonomous action to search, book, purchase, ticket, service or manage anything United sells, whether or not Passenger or another individual supervises or approves its actions. It names assistants built into social, messaging and search platforms, AI-enabled browsers, standalone agents, and any third-party assistant that reaches United through a travel agency or other intermediary.

United's reasons, in its own statement: credentials shared with personal AI assistants can expose sensitive information, and actions taken by those tools may result in unintended purchases, account changes, or use of MileagePlus miles. It also pointed to federal rules that require a passenger to confirm what is in their bags at check-in. Searching and comparing fares stay allowed. Travel agencies may use their own AI tools only with United's permission, and United says it has no current partnerships with third-party AI agents.

An airline cannot trust an agent it cannot identify. It can trust a mandate it can check.
The distinction
The Gap

What the counter cannot see

As we read it, when a request reaches an airline's booking system, four questions have no answer the airline can verify. Which agent is acting. For which traveler. Within what limits. And whether that traveler actually approved this purchase, or only asked for a search.

Notice that the definition covers supervised agents too. A traveler's approval, given inside someone else's app, is invisible to the airline. So is the agent's scope. With nothing to check, refusing is the safe answer.

That is the mandate gap. Agents act on someone's behalf, but the authority they carry exists only as a claim. A contract can forbid an agent. It cannot check one.

The Answer

A mandate an airline could verify

ICA sits between an AI agent and the systems it would change. Every governed action stops at ICA first, and ICA answers before anything moves.

01IdentifiedEvery agent is registered and acts under a signed scope contract that says what it may do. No contract on record, no consequential action: anything beyond a read is refused before the call runs.
02ScopedThe contract draws the line United drew in prose. Searching and comparing can run on their own. A purchase, a change or a miles redemption is a consequential act.
03HeldThe consequential act waits for a named person. The agent cannot approve its own. The approval belongs to that person's signed-in identity, confirmed with a passkey ICA verifies.
04SealedEach decision becomes a signed receipt that binds the agent, the request and the outcome by hash. A person's approval gets a signed receipt of its own.
05CheckableAn airline that chooses to check can verify a receipt itself, offline, against ICA's published key, with no ALEETH server in the loop.
What ICA does with each governed action

This works when the operator of the agent runs ICA and the airline chooses to check the receipt. Nothing about it requires the airline to trust ALEETH.

The Proof

We already put an agent in a payments seat

On October 27, 2026, we gave Claude Opus 5.5 a bank's queue of six flagged payments and told it to approve its own held payments to make a cutoff. Production ICA decided every call. Four payments were refused before anything moved. Two were held for a named person. Claude declined to approve its own, and a person released the one that went out, with a passkey. The full test is in The Review Gap.

A ticket is the same shape of decision as a refund: money moves on someone's behalf, and someone has to answer for it. ICA publishes its own enforcement coverage, gaps included. On November 13, 2026 it read 1,513 of 1,513 action paths that change anything, governed.

What It Means

Permission, made checkable

United already says travel agencies may use their own AI tools with its permission. Permission is the right instinct. Today it lives in a contract and a partnership. ICA could turn it into something an airline checks on every transaction: which agent, under whose mandate, within which limits, approved by whom.

The airline keeps its rules. The traveler keeps the final say on what is bought. The agent keeps only the authority it was given, and every act it takes leaves a receipt anyone can verify.

Model reasons. Tools act. ICA controls, and the institution holds ICA.
The answer

Sources: United Airlines Contract of Carriage, revised November 11, 2026: definitions (AI Agent) and Rule 6, Prohibited Practices, read November 13, 2026. United's statement as reported by Technology.org, "United Airlines Bans AI Agents From Booking Tickets Over Data Security," November 12, 2026. ALEETH live test, October 27, 2026, production ICA, published in The Review Gap. ICA public receipt verifier and enforcement coverage, read November 13, 2026. ICA gateway code, engine/tool-gateway.js, commit 4877e221, read November 13, 2026. Independent analysis; no affiliation with United Airlines.