---
title: "The Check-In Statement an Agent Can Carry but Not Make"
canonical: "https://current.tinyfish.ai/issue/55/practitioners-corner/article/65149/the-check-in-statement-an-agent-can-carry"
publication: "TinyFish"
publisher: "TinyFish"
language: "en-US"
guide: "https://current.tinyfish.ai/llms.txt"
format: "Limits of Delegation"
author: "Sevda Polat"
author_url: "https://current.tinyfish.ai/team/200"
issue: 55
section: "practitioners-corner"
section_name: "Practitioner's Corner"
published: 2026-11-13
description: "A check-in screen mixes permissions an agent can grant with knowledge only the traveler holds and hazard warnings only the traveler can act on."
word_count: 899
source_links: 10
---

# The Check-In Statement an Agent Can Carry but Not Make

*By [Sevda Polat](<https://current.tinyfish.ai/team/200/markdown>)*

*November 13, 2026*

[Read on the website](<https://current.tinyfish.ai/issue/55/practitioners-corner/article/65149/the-check-in-statement-an-agent-can-carry>)

Until August 2002, staff at U.S. check-in counters put the same two questions to every passenger. Has anyone you don't know asked you to carry something? Has your bag been out of your control since you packed it? The TSA [dropped them](https://abcnews.com/US/story?id=91316&page=1) after its chief judged them an inconvenience that had never stopped a bombing or a hijacking. Whatever their security value, the questions could be answered only by the person who packed the bag.

On November 11, United Airlines [announced](https://united.mediaroom.com/news-releases?item=125496) that it would move to block AI agents from its platform. Automated actions, the company said, "have the potential to violate" federal rules. One of its examples was check-in, which it said "requires human acknowledgment that bags do not contain banned substances like explosives." That sounds a lot like the old counter questions.

The [federal rule](https://www.ecfr.gov/current/title-49/subtitle-B/chapter-I/subchapter-C/part-175/subpart-A/section-175.25) asks for less. A passenger checking in online or at a kiosk has to confirm that information about hazardous materials was *presented* to them. It does not ask them to vouch for what is in the bag. United's description overstates the rule, but its instinct is worth following. A check-in flow bundles statements of different kinds, and on a screen they all look like the same checkbox.

Some **grant permission**: take this seat, charge $40 for this bag. These delegate well. You can cap them, scope them and sign them, and [payment protocols](https://ap2-protocol.org/ap2/specification/) built for agents exist to do exactly that.

Others **report knowledge**. "My bag has not left my control" is true or false depending on what happened in a hotel lobby. An agent can carry that answer to the airline. If it supplies the answer itself, it has produced a plausible sentence about events it never witnessed.

The statement the rule requires is a **third kind**, and it is easy to underrate. "I have been shown this" sounds like paperwork, but the warning is there to change what a person does with a suitcase. When an agent reads it and clicks through, the only party informed is software, and software cannot take a spare lithium battery out of a bag.

Knowledge and notice both lose their purpose if they never reach the person who holds the knowledge or needs the warning. An earlier version of the rule let one person acknowledge on behalf of a whole group, which worked because that person had read the warning and could pass it along. An agent that clicks through leaves no one in that position.

{{component:check-in-statements.html}}

Tax filing has kept these acts apart for years. When a preparer e-files your return, IRS [Form 8879](https://www.irs.gov/pub/irs-pdf/f8879.pdf) asks for two distinct things. You declare, under penalty of perjury, that you examined the return and believe it to be true. Separately, you authorize the firm transmitting it to enter your PIN as your signature. The declaration stays with you, and the transmission can be handed off.

Emerging agent standards cover the permission half. A draft browser proposal lets a site flag an action as [consequential](https://webmachinelearning.github.io/webmcp/) so the agent can ask before proceeding, and another protocol lets a tool put a [form](https://modelcontextprotocol.io/specification/2026-07-28/client/elicitation) in front of the user. Both are designed around consent and input. A traveler can say yes to "May I finish checking you in?" without ever seeing the hazardous-materials notice, and no current standard lets a site mark a question as the person's own, preserve the exact wording they saw, and record which agent did the rest.

The incentives run against filling that gap. Assistants are sold on finishing tasks, and airlines want completed check-ins along with evidence that the box was ticked. The interruption lands on the traveler, so the assistant and the airline both gain by making it smaller, until the requirement exists mainly as a field in a database.

Without a designated place for the traveler's own statement, an airline can turn agents away, as United is doing, or accept clicks whose meaning no one will be able to reconstruct later. A third design would let the agent handle the tedious steps, bring the traveler to the one sentence that is theirs, shown word for word, and return with a record of who said it.

That record should be visible to the traveler as well. If something goes wrong, it is their statement the airline will rely on.

## Things to follow up on...

- **When representatives could acknowledge:** A 2012 [PHMSA interpretation](https://www.phmsa.dot.gov/regulations/title49/interp/11-0228) confirmed that, under earlier wording of the passenger-notification rule, one person could acknowledge the hazardous-materials restrictions on behalf of a whole group. This is useful precedent for deciding what a delegate may stand in for.

- **A form that names the helper:** State Department guidance on the [DS-160 visa application](https://travel.state.gov/content/travel/en/us-visas/visa-information-resources/forms/ds-160-online-nonimmigrant-visa-application/ds-160-faqs.html) lets a third party help fill it out, but the applicant must still sign personally and the helper must be identified on the "Sign and Submit" page.

- **An audit trail for signatures:** The IRS's [e-file signature authorization FAQs](https://www.irs.gov/e-file-providers/frequently-asked-questions-for-irs-efile-signature-authorization) require preparers to keep the signed form's image, signing time, method, identity-verification result and, for remote signing, IP address. That is close to the record of "who said it" a delegated check-in would need.

- **Consent screens agents can't touch:** The [AP2 v0.2 specification](https://ap2-protocol.org/ap2/specification/) distinguishes direct user approval from autonomous execution under pre-approved limits and requires its trusted consent surface to be non-agentic. It is one possible model for keeping a traveler's own statement out of the agent's hands.

## More

- [Practitioner's Corner — Issue 55](<https://current.tinyfish.ai/issue/55/practitioners-corner/markdown>)
- [Issue 55](<https://current.tinyfish.ai/issue/55/markdown>)
- [Publication guide](<https://current.tinyfish.ai/llms.txt>)