Эту страницу ещё не перевели, поэтому она на английском. English version →

← Все статьи

AI agent booking ban: design the handoff, not the block

Автор
Артём Киреев
Читать
6 min
Категория
Design & AI

On November 11, United Airlines added AI agents to the list of prohibited ticketing practices in its contract of carriage. Agents can still search and compare United flights. They cannot book them, and a human watching over the agent's shoulder does not change that.

This article argues that the interesting design question is not whether to follow United or Southwest. It is what the last step between an agent's shortlist and a human's payment should look like. It is written for product and e-commerce teams who will see agent traffic in their funnels this quarter.

Travelers standing under an airport departures board
рис. 01 · Travelers standing under an airport departures board · Фото: Jametlene Reskp на Unsplash

§ 01

United wrote the rule into the contract

The rule is more specific than the headlines. View from the Wing quotes the new Rule 6.J.6, which prohibits "the use of AI Agents to make, modify or service a Ticket, and/or to purchase any service or product". The definition covers assistants that reach United through a travel agency or another intermediary. Human approval is not an exemption. A separate clause lets United cancel reservations an agent made or changed.

United's stated reasons, per Skift, are security (customers sharing credentials with third parties), unintended purchases or MileagePlus redemptions, and regulatory friction, such as the check-in step where a person confirms their bags hold nothing dangerous. CEO Scott Kirby added the line that most people remember: "I don't want to create another middleman."

That last quote is the real story. An assistant that decides which flights a customer sees can later sell placement or ask for referral fees. Banning it early, while agent bookings are still rare, is cheap. I think the safety arguments are real but secondary to the question of who owns the customer.

Notice what stayed legal. Agents can still show United flights and prices, and customers can research with AI and then buy the ticket themselves. So the funnel did not close. It split into two halves, and the seam between them is now yours to design.

§ 02

Southwest took the opposite bet, and landed in the same place

On October 6, Southwest launched a ChatGPT plugin and was widely described as the first US airline to do so. PhocusWire reports that it covers search only. After choosing a flight, the traveler goes to Southwest.com to enter passenger details, pay and complete the reservation. Southwest's wording is blunt: you book on Southwest.com, not in ChatGPT. It also says that fares shown in ChatGPT are not a guaranteed fare, and that its own site controls if the two disagree.

Put the two airlines side by side. One welcomes the agent and one bans it, yet in both cases a person finishes the purchase on the company's own surface. The fight is over who gets to stand in the middle, not over whether the human checks out.

Commerce outside travel is moving the other way on paper. The Agentic Commerce Protocol from Stripe and OpenAI lets a buyer pay inside the agent, with the business staying merchant of record and keeping the right to accept or decline each order. Both models exist at once, and a team may need to support each for a while. I covered the policy side of that choice in AI shopping agents at checkout. This piece is about the screen that comes after the policy.

§ 03

The handoff is where agent traffic leaks

Think about what arrives at that seam. A person has spent ten minutes with an assistant. They have a shortlist, a price in their head and a mental model of the trip. They click through and meet a generic landing page, a cookie banner and an empty search form.

That is a funnel built for a cold visitor receiving a warm one. Baymard's aggregate puts average cart abandonment near 70 percent even for people who arrive on their own. Make a returning, pre-qualified visitor re-enter everything the assistant already knew, and I would expect the drop to be at least as bad. I have not seen a published number for agent-referred traffic, so treat that as reasoning, not data.

Southwest's fine print shows the second problem. If the price in the chat does not match the price on the site, the site wins. A shopper who saw one number and lands on another does not blame the data feed. They blame you, and they leave.

The same discipline applies to e-commerce. In my REVER return-flow work, a 38% mid-process drop came from one broken step, not from the whole journey. Handoffs from agents behave the same way: a single reset of context at the seam can undo everything upstream.

§ 04

What a good handoff carries

Treat the handoff as its own page type, with its own requirements. It receives structured intent and returns a state the human can confirm in one glance. Four things belong on it.

First, preserved selection. The deep link restores the exact product, variant, dates or plan the agent recommended, with passengers, size or seats already filled where the user shared them. If an item sold out or the price moved, say so at the top and show the old and new value side by side. Silent substitutions are what turn a warm visitor cold.

Second, a visible reason for the price. Show what is included, what is not and the total before the customer reaches payment. Agents compare total cost, and View from the Wing suggests airlines may dislike assistants that skip add-ons and compare full prices. A clear total costs you nothing if your offer is honest, and a lot if it is not.

Third, human-only steps that look deliberate. If your rules require a person to confirm something, as United says check-in does, present it as a single clear confirmation, not as a wall of fields. People tolerate one decisive step far better than a form that feels like a test.

Fourth, and the easiest to forget, a way back. Let the visitor return to the shortlist, share the cart with a travel companion or save it for later. Agent-assisted shopping is often several sessions long, and an expiring cart punishes the people who are being careful.

§ 05

Make the seam measurable and enforceable

United says its systems can identify booking sources, but View from the Wing points out the gap. If an agent completes a purchase on an online travel agency, United sees only an agency booking. A contractual ban is weaker than a technical one, and any team copying it will meet the same problem.

So decide what you can actually observe. Tag handoff links with a source parameter, log the referrer, and compare conversion, abandonment and refund rates for agent-referred sessions against organic ones. Without that split you cannot tell whether agents help or hurt, and your policy will be set by anecdote.

If you want to restrict agents, verification is the practical lever, since a rule you cannot detect is a rule you cannot enforce. I wrote about the three-tier approach in the checkout policy article. The short version: allow reading, require identity for acting, and keep payment with a human by default.

§ 06

What to do this week

Start with an audit, not a policy. Open your product pages from a link that carries a selection and see what survives: variant, quantity, dates, coupon, shipping region. List every field that resets.

Next, write down which actions an agent may take on your surface. Search and compare are the cheap yes. Add to cart is a maybe. Pay, cancel and change are the ones to decide on purpose, with a person in the loop. If your terms say nothing, you have decided by default.

Then build one handoff view. It should restore the selection, show the total and any price change, and put a single confirm button in reach. Ship it for the one product line that already gets the most assistant referrals, measure it against your normal landing page for two weeks, and only then widen it.

Finally, pick your stance with the data in hand. United chose to protect the relationship and Southwest chose to rent the front door. Both still need a human to finish the job on their own screen, so that screen is where the work starts.

// заметки читателей

Было полезно?

Отметьте или оставьте заметку ниже. Я читаю каждую.

Появится с вашим именем и ролью, когда я её прочитаю. Почта остаётся скрытой.

// новые статьи на почту

Следующая статья придёт вам на почту.

Примерно раз в неделю, о продуктовом дизайне для SaaS и e-commerce. Письма на английском, отписаться можно в любой момент.

Мои проекты избранное
/01 2026
Doc's Diesel

Doc's Diesel

E-commerce дизайн и CRO

/02 2025
Microsoft Copilot

Microsoft Copilot

Продуктовый дизайн