A United Express Embraer 175 at Québec City airport on April 20, 2025. File photo: Wilfredor / Wikimedia Commons, CC0 1.0. Resized from the original.
In this article
- How airlines deal with AI agents
- From search to ticket
- Booking, payment and staying in control
- When travel gets complicated
- What shapes the offers
- What we expect from an agent
United’s change puts booking back at the center of the debate
Asking artificial intelligence to find a flight already seems almost ordinary. Asking it to buy the ticket for us is another matter. On November 11, 2026, United announced new restrictions targeting unauthorized personal AI agents. The change deserves more than a showdown between an airline supposedly opposed to progress and assistants presented as inevitable. It raises a practical question for travelers: who is allowed to act, which systems are they allowed to use, and what commitments can they make on the passenger’s behalf?
In its November 11 announcement, United targets actions by unauthorized third-party agents, whether directly on its platforms or through other businesses. The airline cites concerns including the protection of personal information and the risk of unintended actions. That is its explanation. We can take those concerns seriously while examining the response it has chosen and what it means for a customer who simply wanted to save time.
The original promise is appealing. We do not all want to repeat the same search on several websites, translate fare conditions, or piece together the schedules for a complicated trip. An assistant that understands a request, compares the options properly, and explains the differences would be useful immediately. The benefit is not just minutes saved. It can also be a better understanding of what we are buying.
An airline ticket, however, is more than an answer in a conversation. It relates to a specific journey, named people, and conditions that will continue to matter after payment. Passengers need to be able to find the itinerary they bought, its conditions, and a contact if something goes wrong. Some choices also require a personal answer from them; reassuring language cannot settle those decisions. A smooth conversation can help prepare the trip; it cannot, by itself, ensure that the arrangements are carried out.
That makes the debate much more interesting. An agent can understand our request perfectly well without being authorized to use its chosen channel. It can have authorized access without knowing all our preferences. It can also prepare an excellent comparison and deliberately stop before purchase. These are different services. Describing them all simply as “booking” obscures what actually works and what remains to be decided.
We are looking at this from the passenger’s perspective, beyond the technical demonstration. An action completed in seconds is an improvement only if it produces the intended trip and leaves the traveler with someone who can support the booking afterward. An extra step can be useful when it explains an important choice. A good assistant should spare us repetitive tasks while drawing attention to decisions that deserve our consideration.
There is also a date to keep in mind. This article examines the public positions checked as of November 12, 2026. An airline may change its agreements, a provider may obtain approval, and an announced feature may become available later. We will therefore describe what has been published today, distinguishing access rules, stated intentions, and the capabilities actually presented. The absence of a restriction worded like United’s will never, by itself, establish permission.
We have made no purchases or booking changes for this article. The scenarios later in the piece are imagined situations used to examine travelers’ choices. These examples examine what a service should show before and after an action, without claiming capabilities we have not verified. They also let us ask what we consider the essential question: once the excitement of the first exchange has passed, does the passenger still know what was decided, by whom, and with what consequences?
The issue extends beyond United. Airlines, agencies, and technology providers are not always describing the same thing when they talk about AI. Some offer conversational search, others contracted access to offers, and others service after purchase. Travelers do not need to master the entire industry. They do need to understand the service being offered.
That clarity, more than the claimed level of autonomy, should guide the development of digital travel agents. We previously examined the intermediary’s role in our analysis of Meta Muse in Singapore. United’s announcement makes the issue more concrete: we need to distinguish what a tool recommends from what the airline authorizes it to do.
A better conversation before the search can also help. “The best flight” means little without knowing the purpose of the trip. Arriving early for work, avoiding an extra night, or keeping an evening at the destination can lead to different choices. A good assistant should reveal those differences without asking unnecessary questions. It could help us identify what matters and keep those criteria in the comparison. Easier booking should not begin by forgetting why we are traveling.
What United prohibits and what remains allowed
The starting point is the contract of carriage revised on November 11. United uses a broad definition of an AI agent, potentially including autonomous or partly autonomous software even when supervised by the passenger. Its reservation and ticketing rules address bookings made, changed, or serviced by these agents. The airline specifically reserves the ability to cancel certain reservations that fall within its prohibited practices.
United’s explanatory page makes essential distinctions: a customer’s permission does not replace the airline’s; bookings made by an agent before November 11 will not be canceled on that basis; and assistive technology for customers with disabilities is unaffected. United also says it is not currently working with a third-party AI agent. Those points need to accompany any account of the change.
The most immediate consequence for passengers is straightforward: being able to supervise software does not establish that its involvement is allowed. A customer might believe they remain fully in control because they read the proposals and approve each step. United has chosen to make its restriction depend on the type of participant and its access, rather than only the freedom granted by the user. That difference explains why a final approval does not settle everything.
Going through an agency cannot automatically be presented as a solution either. The FAQ includes indirect access and distinguishes third-party tools from tools an agency might develop with the required permission. It would be misleading to promise that using a different screen necessarily makes the operation compliant. The agency’s permissions need to be checked, including whether they cover the tool it uses. The fact that the tool works, has been around for years, or is easy to use does not answer that question.
There is still room for help with planning. The airline distinguishes general research from unauthorized activity on its website or app. An assistant can help formulate criteria, explain terminology, or prepare a search that the traveler then carries out within the permitted framework. That is more limited than handling everything, but it can still be useful. Understanding a condition before choosing is sometimes more valuable than having a click performed for us.
The assistant’s provider needs to describe that boundary honestly. If its role ends at comparing information or preparing a request, it should say so from the outset. The customer may be perfectly happy with that division of work. Discovering the limitation only when they thought the trip had been handled would be disappointing. “Found,” “prepared,” and “booked” should not be used interchangeably to announce the same result.
Existing reservations also need careful wording. United’s clarification avoids unnecessary alarm for customers already affected. It does not remove the need to check the reservation and ticket, as the airline recommends. Checking the available details does not guarantee that a particular booking is free of other problems. The important point is not to transform a change concerning agents into a story about every AI-assisted trip being canceled retroactively.
Accessibility needs equally careful treatment. The airline says assistive technologies are not affected. That exception needs to be retained rather than treating every tool that helps someone as an agent negotiating or completing a purchase. Assistance with reading, expression, or navigation serves a different purpose. We will return to accessibility requirements without attributing to United a restriction it specifically says it is not applying.
United presents its position as a response to risks involving data, accounts, and actions taken on a passenger’s behalf. Those concerns should not be dismissed. They do not establish that every assistant would inevitably make the same mistakes either. The debate should address the conditions that would prevent, detect, and correct problems. An airline can close an access route; it can also explain what safeguards would allow it to recognize another one.
It is useful, then, to separate the current decision from the discussion about the future. Today, a provider should not describe itself as authorized by United without being able to establish that authorization. For the future, we can favor better-defined access, limited to specific tasks and controllable by the traveler. There is no reason to assume that the only possible automation is software using a customer’s credentials as though it were that customer. Other arrangements can be designed and negotiated.
Explaining the rules should not fall entirely on the passenger. A customer cannot necessarily see whether an offer comes from an authorized connection, pages being consulted, or a chain of intermediaries. Clear information about the authorized service would be more useful than a broad security promise. They should know whether the tool can only search, whether it can carry out a transaction, and who will handle service afterward. That information should be available before any request for account access.
We are discussing United’s published contractual position here. This is not individual advice about how a clause would apply in a particular dispute. For someone planning a trip, the published position nevertheless provides concrete information a serious service needs to consider. The most useful assistant will recognize the boundary in time, provide a clear handoff, and preserve the useful work already done, rather than implying that customer permission overrides every other requirement.
Rule changes also need to be tracked without making the customer start over. If a provider could perform a task yesterday but cannot today, its interface should say so before continuing. We would want it to preserve useful search details, state clearly what has not been done, and offer a workable next step.
A late failure message after a long conversation simply transfers the difficulty to the traveler. An assistant’s quality also shows in how it explains that it can no longer perform a task, without describing an interrupted process as a completed trip.
American, Delta, and Southwest are taking different approaches
Comparing airlines should start with their published terms, rather than a simple ranking of open and closed companies. The language, services, and commercial agreements differ. An airline can approve a particular integration and reject other uses. It can also allow certain searches while placing tighter controls on purchases. For travelers, that diversity primarily means an experience available with one carrier cannot be assumed to work with the next without checking.
American Airlines’ platform terms, effective January 1, 2026, require express written permission for third-party tools used in connection with its platforms and content. Their AI section provides an approval process. Customers using an approved tool retain responsibilities. This framework cannot be summarized as American accepting any assistant capable of connecting to its systems.
On November 11, Fast Company reported statements from American and Delta. American described some tools facilitating purchases and its monitoring of automated traffic. Delta described certain appropriately authenticated interactions, with additional safeguards for sensitive actions and passenger oversight. These are statements reported by the publication, rather than proof that a particular assistant is approved for every task.
That qualification changes how a commercial promise should be read. “Works with Delta” might refer to checking schedules, access to an account, or a particular action on a reservation. Those capabilities are not interchangeable. Before treating the claim as decisive, we would want to see what the service can actually do and under what arrangements. An airline name on a partner list is less informative than a clear description of the permitted actions.
Safeguards for sensitive actions particularly need to remain visible. An assistant able to recognize an account has not necessarily received the customer’s approval for every action available within it. Identifying the person, authorizing a provider, and deciding to carry out a transaction are three separate questions. They can be handled in a simple process. Treating them as one login, however, makes it much harder to understand what has just been allowed.
Southwest illustrates another route. Its October 6 announcement presents a ChatGPT plugin developed with OpenAI and AWS for discovering destinations and shopping for Southwest flights. The airline describes a first step in a phased rollout. The release does not establish that a ticket can already be purchased and issued entirely within the conversation. The search experience offered today needs to be distinguished from future capabilities.
An integration like this has an obvious attraction: it can shorten the path from a travel idea to options genuinely connected to an airline. Travelers no longer necessarily have to translate their question into several forms. They still need to know what the integration covers. A properly displayed selection of flights can be a good service even if the purchase finishes elsewhere. Moving to the airline’s website is not a failure when that step is explained and the selections carry over correctly.
The comparison becomes misleading, however, if that search process is contrasted with United’s restriction as though both necessarily concern the same action. An airline-recognized plugin and third-party software taking control of a browser do not rely on the same relationship. They may look similar in a short video because both respond to a request in ordinary language. What the customer sees does not always reveal the permissions behind the service.
We would therefore like providers to move beyond sweeping descriptions. Saying an assistant “books flights” does not explain which carriers it covers, what it does, or when the user takes over. Useful information would identify the relevant markets, booking types, change options, and contact after purchase. That precision does not hold back innovation. It lets customers confidently use a limited feature rather than expect a universal service that does not yet exist in the promised form.
An airline’s own tool also needs to be distinguished from one chosen by the passenger. A carrier can use AI for support, recommendations, or operations without giving access to an outside service. That is not inherently inconsistent: access, available data, and responsibilities differ. The interesting question is whether the arrangement improves service for travelers and whether the access conditions for other participants remain understandable, rather than simply counting mentions of AI.
The current diversity could be a useful opportunity to compare services by their results. Search that explains its coverage limits can be more useful than an agent claiming to do everything. An airline opening a specific feature should be able to explain how it works and how to get help. An authorized assistant should be able to establish what it is allowed to do without making the customer investigate every connection. We expect that transparency on both sides.
These differences matter particularly on a trip involving several airlines. A single conversation may discuss offers subject to different access conditions. The same layout should not suggest that the same actions are possible with every airline. If the tool can complete only part of the request, it needs to explain this before the customer books part of the trip without knowing how to book or manage the rest. The aim is not to decide which airline looks most modern. It is to know which trip the chosen service can actually arrange and support properly.
An honest presentation should also distinguish an available feature from a trial or a planned next step. These labels tell customers what they can expect now. A successful demonstration on one route does not prove that every combination is supported. A service can start with a few clearly defined cases and still be very useful. It simply needs to explain which ones it handles and announce new capabilities when they are actually available.
Searching, recommending, reserving, and issuing a ticket are four different things
“Booking” is a convenient word in a conversation, but it can complicate everything else. It may describe a flight search, a prepared shopping cart, a reservation record, or a ticket actually issued. To understand what an agent offers, those stages need to be separated. Travelers should not have to learn an agency’s vocabulary; the tool should use words matching the result it has actually obtained. Clarity starts with the right verb at the right moment.
- Searching means obtaining options that match a set of criteria. That may produce schedules, itineraries, cabins, and fare offers. Their presence in a response does not mean they are being held for the customer. Search can also have limits in coverage or freshness. A useful assistant should make those limits visible: what it is comparing, when it obtained the information, and what will need to be checked again before a transaction.
- Recommending adds a judgment to those results. The agent explains why one option might suit the traveler better than another. That judgment depends on preferences and how they have been interpreted. The shortest flight, the cheapest, and the one best preserving a working day may be three different flights. A good recommendation reveals the trade-off. People should be able to understand why an option was selected without having to trust an unexplained ranking.
- Reservations are then made with a seller or airline. That action cannot be inferred from the quality of a suggestion. A tool may have prepared everything necessary without creating anything in a reservation system. At this point, the customer should be able to distinguish what has been sent from what remains an intention in the conversation. We will examine the evidence to expect later, without treating a congratulatory screen as sufficient confirmation.
- Ticket issuance is another distinct result. We will not go through every document different systems produce here; the important point is that a reservation record and a ticket are not two names for the same thing. The assistant needs to state precisely what has been achieved and what, if anything, remains. An interface using one sentence for every situation hides essential information: does the traveler still need to do something before being ready to fly?
These distinctions also help recognize the value of a partial service. An assistant preparing a well-explained comparison may have completed its promised task without buying anything. An agency issuing the ticket and providing ongoing support supplies a different service. Disappointment arises when both are sold with the same promise. We would like each participant’s role to be visible early enough to choose the level of help we actually need.
Moving from one stage to the next deserves more attention than a progress animation. If the offer changes along the way, the user should know what changed before continuing. A different arrival time, cabin, or less flexible condition can change the trip’s value. A smooth interface should not give the impression that nothing has changed. A pleasant conversation can support a new choice; it should not hide that choice in the next part of the exchange.
Ordinary requests quickly show why the vocabulary matters. “Find me a good option” allows room for exploration. “Prepare this trip so I can check it” asks for a more specific result. “Buy this offer” calls for a defined action. The agent should recognize those differences rather than turn a general intention into a purchase instruction. We do not want to repeat technical language in every sentence. We want the service to respect the ordinary meaning of the words we use.
The result should also make sense at a glance. A traveler may check a notification between meetings or send the details to a family member. If they need to reread the entire conversation to determine whether a flight was merely suggested or actually bought, the promised simplicity has not materialized. A clear, dated status, the seller’s name, and the next useful action would be more valuable than a long response repeating the attractions of the route found.
This is how conversation can make booking easier and help people follow its progress. It can explain complex terminology, connect the stages, and flag changes. It also needs to stop and acknowledge an incomplete result. A reliable agent should be comfortable saying it found an offer but did not book it. That precision protects the traveler’s trust and makes a proper handoff to the next participant possible.
The distinction remains useful when someone changes their mind during the conversation. Asking for another recommendation should not be interpreted as a request to change a trip already purchased. The same word can refer to a suggestion or an action with significant consequences. The assistant needs to know the current state of the request, recall it when necessary, and ask if the meaning becomes ambiguous. An instant answer is of little use if it addresses the wrong action.
Who is authorized to deal with the airline?
When we ask for help, several permissions can appear to merge into one agreement. We accept assistance, allow a provider to use certain information, and may permit a transaction. The airline, for its part, needs to recognize the method of access being used. These permissions concern different relationships. A traveler’s request alone cannot change the conditions under which a commercial service accesses a carrier’s systems.
The customer’s instructions answer an initial question: what do they want to entrust to someone else? Those instructions may cover only a search or a defined transaction. They do not establish how the provider is authorized to deal with the airline. Confusing the two would imply that anyone we ask to perform a service automatically has every tool and permission needed to do the job. In travel, as elsewhere, that shortcut needs to be avoided.
KLM’s website terms illustrate the distinction. They allow for genuine bookings for yourself or someone you are authorized to represent, while requiring prior written agreement for commercial retrieval of content. Being able to act for a traveler is therefore not the same as having a general right to extract and use website information. Nor does this establish that every kind of personal AI assistance is explicitly prohibited.
A distribution chain may involve several businesses. The customer sees an assistant’s name; behind it may be an agency, a reservation provider, and a technology partner. That arrangement is not inherently problematic. The roles still need to remain identifiable. One participant’s rights do not necessarily pass to everyone technically able to send it a request. A working connection alone does not establish what each participant may do with that access.
Ryanair’s March 2026 agreements with Vola and Fru show that a carrier can recognize specific intermediaries. They do not provide blanket permission for every provider or every piece of software further down the chain. Travelers would benefit from knowing the agent’s exact role in the sale, the seller’s name, and the relationship giving that seller access to offers. A general list of logos does not fully answer those questions.
Knowing whose account is being used is not the same as approving the provider using it. Logging in with a passenger’s details may tell a system which account is being accessed. It does not establish that the airline approves the software using that access. The issue is therefore not simply whether an assistant can get past a login page. A clearer process would reveal its own role and permissions, rather than act as though the passenger were personally performing every action.
We do not expect every customer to investigate distribution contracts. Replacing a complicated form with an even harder task would make little sense. We expect providers to explain their service in verifiable terms: what they can request, what they can carry out, and whom to contact afterward. That information should agree with the airline’s. A disagreement on something this fundamental should be resolved before asking for traveler data.
Airlines also have a part in making this understandable. If they recognize certain access arrangements, they would benefit from publishing a clear account of what those arrangements cover. Customers should be able to distinguish an integrated service, an authorized agency, and general planning assistance without interpreting scattered clauses on their own. A clear rule can protect travelers. A hard-to-find rule alongside an interface that appears to allow everything creates room for misunderstandings.
Clarity needs to extend beyond the initial purchase. Permission may cover search and sale without making every later action available through the same channel. Providers should then explain the handoff rather than suggest that their role remains unchanged until the traveler returns. We will examine these situations in the discussion of ongoing service. The principle matters already: the authority required depends on the task and when it is being performed.
Ultimately, we prefer a service that can state its limitations precisely to an assistant whose confidence substitutes for evidence. It can help choose another route, preserve the prepared criteria, or pass a request through an approved channel. That honesty does not make it less useful. It lets travelers know whom they are entrusting with what, rather than treating a single sentence of consent as the universal answer to questions of access, representation, and responsibility.
The public documentation for Ryanair’s Travel Agent Direct also shows why scope matters: it concerns approved brick-and-mortar travel agencies and offline sales, with requirements to keep passengers informed. It does not describe an interface open to any AI agent. Online agencies’ terms need to be considered separately, for the sales and tools they cover. The example shows why an agreement’s intended participants must be identified before using it to make a promise about another service, even when both involve booking the same flight.
Comparing complete offers rather than isolated prices
An assistant should offer more than amounts listed from lowest to highest. Travelers buy a way to get somewhere, with particular features and conditions. Two offers on the same flight can meet their needs differently. A trip that costs more on the screen may also be better for the stay. The quality of the comparison depends on what it brings together, what it explains, and what remains unknown.
- Schedules are an initial example. An early departure may require a night near the airport; a late arrival may make a day at the destination less useful. An assistant should not decide the value of those consequences by itself. It can reveal them and ask for a choice when the difference matters. The best recommendation is not the one using the most criteria. It is the one retaining the criteria that matter for this particular trip.
- The airports involved also need to be clearly identified. Grouping a city’s different airports can be convenient during exploration. It becomes insufficient when someone is planning a meeting, a transfer, or a short stay. The exact airport and whether the itinerary works as a whole need to be easy to check. Quietly replacing one arrival airport with another because they serve the same city can turn a reasonable trade-off into an unwelcome surprise.
- Included services need equally clear comparison. A bag, seat selection, the ability to change a ticket, or cancellation conditions may influence the decision. We will not assume those features are identical across offers. The agent should show what is included, what still needs to be added, and what it could not verify. An empty field should never be presented as a favorable answer by default.
The differences then need to play their proper role. A traveler may choose a less flexible offer because the dates are fixed, or accept a higher cost to retain an option to change. Neither is automatically the best choice. The assistant’s job is to explain the consequences in understandable language. Any value it gives flexibility, time, or comfort should reflect an expressed preference rather than an invisible assumption.
A useful comparison also distinguishes trips that appear similar but are put together differently. Customers need to recognize a connecting itinerary sold as one arrangement and a combination of separate journeys. We infer no automatic rights or identical protection here. We ask that the offer’s structure be clear before selection. An attractive line on a map does not explain how the different pieces actually fit together.
The agent should also explain the coverage of its search. If it accesses only certain sellers or offer categories, the result is a comparison within those limits. It can be very useful without being exhaustive. “Best” becomes misleading when it implies that every option has been examined. A few properly compared, well-explained offers would be more valuable than a ranking whose omissions the traveler cannot see.
Travelers need to know how current the information is until they make a decision. An offer found and then discussed in a conversation may no longer be available on the same terms when the user returns. The assistant needs to check again without losing the reason for the original choice. If the result changes, customers should understand the difference and have a genuine opportunity to decide again. A continuous conversation does not guarantee an unchanged offer.
We do not want every search to become an endless spreadsheet. The presentation can begin with the few decisive differences, then offer access to more detail. This is a promising use of AI: explaining why two options look similar yet differ on a point that matters. A good summary reduces the amount people need to read while preserving information that could change the decision. It simplifies the traveler’s work without quietly choosing for them.
Once the comparison is understood, the next question becomes more precise: what will actually be obtained by continuing? This is where the quality of the recommendation needs to meet the quality of execution. An agent finding an excellent offer has not yet completed a booking. Moving to a verifiable reservation, ticket, and ongoing service requires evidence of its own. We will now turn to those concrete results.
Evidence that the trip is actually booked
An agent that ends its work with “you’re booked” needs to show what supports that statement. A reassuring message in the conversation is not enough, even with an attractive trip summary. The traveler needs to find the itinerary through a channel the airline recognizes, with the right passengers and an understandable status. We would expect the same from a human travel agency: documentation we can use to verify the booking and manage what comes next.
In particular, we need to distinguish a saved offer, a reservation that has been created and, where the airline uses tickets, ticket issuance. A booking reference helps retrieve a trip; it does not describe every step needed to complete the purchase. Air France’s Time to Think option provides a concrete example: subject to its conditions, it holds a fare before ticketing. The duration is stated when the option is purchased. Confirmation of that hold does not mean the ticket has been bought.
Confirmation that can be checked
We would therefore want a precise account of what is confirmed, what is pending and, where relevant, the deadline for taking action. Travelers should not have to understand reservation systems. The provider’s job is to explain the status in plain English. “Your reservation is held until this time; the ticket has not been issued” is useful information. Replacing that distinction with “everything is ready” leaves the traveler to discover what is missing, possibly too late.
References also need to lead to the right place. On its My trips page, United distinguishes the confirmation code used to view a booking from the ticket number that can be used at check-in. That specific process does not apply to every airline. It does show why an agent should label the references it provides rather than list several unexplained codes. Travelers need to know which one to use with the airline and which one to give the intermediary that sold the trip.
Next comes the actual content of the booking. Names, dates and airports need to match the accepted offer for every traveler. A return the following day, an arrival at another airport or a missing passenger is not a presentation detail. We would rather see the final summary flag a discrepancy, however inconvenient, than repeat the originally requested itinerary as if nothing had changed. The conversation needs to describe what was obtained, not what was promised a few minutes earlier.
Purchased services deserve the same clarity. A seat request, a paid bag and a saved preference do not necessarily have the same status. The agent should report the response received for each important item and retain the available documentation. It must not turn a request into a guaranteed service simply because that is easier to say. In a hypothetical example, two travelers might have confirmed flights while the adjacent seats they wanted still need to be checked.
Keep the evidence within the traveler’s reach
The useful booking information should remain accessible without the assistant. Confirmation sent directly to the traveler, the seller’s contact details and access to booking management avoid dependence on a single conversation. If the app disappears from the phone, the essential information should still be available. A few clearly identified documents are more useful than an endless history mixing abandoned searches, old proposals and bookings that actually went through.
The time of the check matters too. A reservation viewed at purchase reflects its status at that moment. It does not ensure that the schedule, gate or seat will remain unchanged until departure. An agent offering ongoing monitoring should distinguish the original confirmation from subsequent updates. An old document must not be presented as a recent check. If monitoring ends after the sale, that limit needs to be clear so the traveler knows to resume checking the booking.
An incomplete result also needs an explanation. If the outbound flight is confirmed but there is a problem with the return, an overall success or failure message is inadequate. The customer needs to know what already exists before approving another purchase. They can then decide whether to keep part of the trip, seek an alternative or contact the seller. The agent should not make those choices simply to display a successful outcome. How it reports a partially completed booking is part of the service.
Our expectation is straightforward: after the purchase, we want less uncertainty than before it. An assistant can spare us several screens, but it needs to preserve the information that lets us check its work. A summary is helpful when it explains the evidence and its limits. It becomes risky when it replaces them. Good confirmation is not defined by its confident tone. It is defined by what travelers can retrieve and understand without asking the agent to tell the story all over again.
What travelers need to decide before letting an agent act
Useful delegation starts with the decisions we want to keep. Asking for “the best flight” leaves too much open to interpretation once an assistant can take action. The best option for arriving early is not necessarily the best for sleeping, working or avoiding a connection. We need to state the priorities that make the purchase worthwhile. That should not take longer than booking the trip ourselves. The agent’s role would be to ask a few relevant questions and then explain what it has understood.
The first boundary is often a travel constraint rather than a price. Imagine needing to arrive before an appointment: the local arrival time matters, but so does the journey from the airport to the final address. An assistant should not treat the plane’s arrival and arrival at the appointment as equivalent. If it lacks the information, it needs to say so. The traveler can allow a comfortable margin or decide how much risk to accept. That becomes an explicit decision rather than a compromise hidden in the ranking of results.
Separate preferences from requirements
- Some requests are flexible. A preference for a departure after nine may allow a little room. Others are essential: a date, an accessible airport, a cabin or no airport transfer during a connection. The agent needs to know which can be relaxed and which require it to come back to the customer. Otherwise, a recommendation can satisfy almost every preference while missing the one that makes the trip possible.
- The budget needs the same care. A limit might cover the flight alone, the trip with bags or the whole order, including service fees. Once that is clear, the next question is whether any additional spending is authorized. Going a few dollars over budget to keep a nonstop flight may seem reasonable to the agent and unacceptable to the traveler. There is no universal answer. The right limit is the one the person chose, in a currency and amount they can recognize at payment.
- We also need to decide what the agent can do if no offer meets every requirement. Can it suggest another airport, an extra night or an earlier departure? Yes, as clearly identified alternatives. Can it buy them without asking again? That depends on the authority actually granted, not on its confidence that it has found a bargain. A creative solution is useful while it remains a proposal. Travelers need to be able to decline without discovering that a different ticket has already been issued to solve the problem.
We would want the assistant to keep a readable record of those trade-offs. That need not be a twenty-page agreement for every search. A few sentences can summarize what is mandatory, what is preferred and what requires another answer. Restating the request also helps catch ambiguity. “Friday evening” might refer to departure or arrival; “Paris” might include several airports. A well-timed question prevents a convenient interpretation from becoming an expensive mistake.
Leave room for the final decision
Permission to buy needs to relate to an offer the traveler can still recognize. A concise summary can bring together the dates, route, passengers, total price and restrictions that influenced the choice. We do not need every condition we have already read repeated. We need to know whether anything important differs from the accepted proposal. A price increase, a different operating airline or an additional connection warrants coming back to the person before proceeding.
Someone who knows exactly what they need might grant broader authority. They could specify a particular route, cabin and maximum price, then authorize a purchase when those conditions are met through a permitted channel. The service still needs to define how long that instruction lasts and how to stop it. A request made while preparing a trip should not remain active indefinitely, ready to trigger a purchase after the appointment has changed or the travel plan has been abandoned.
Silence should not expand that authority. If a new question goes unanswered, the agent can continue researching or preparing options within its permitted access. It should not invent a preference to finish faster. A traveler who is busy, asleep or offline has not necessarily changed their mind. An assistant that knows how to wait is more useful than one that treats the lack of a reply as approval for a compromise nobody wanted.
Some actions also deserve authorization separate from the ticket: adding insurance, buying an upgrade, changing account contact details or committing to other expenses. Their position in the booking flow does not make them incidental to the person paying. We would want to authorize transportation without automatically authorizing every other choice. The agent can point them out when they help a decision. It should not present them as necessary steps when the traveler does not need them.
Good delegation also leaves a way to take over. Customers need to know when an action has been submitted and whether it can still be stopped. We are not asking for a magic button that reverses something the airline has already confirmed. We are asking for an honest account of the operation’s status. “I can still stop the search” and “a purchase request has been sent” have different consequences. That distinction needs to stay visible even in a very simple interface.
This framework gives the assistant freedom to explore and prepare options. It needs to come back to us when the nature of the choice changes. A personalized service recognizes the limits of our agreement. It does not push them aside simply to finish a purchase.
Limit access to the data the agent needs
An assistant does not need to know everything at the first question. Looking up schedules requires little personal information. Building a reservation needs more. Managing an existing booking raises further questions. Access to data should grow with the task rather than begin by opening every account. That also makes the service easier to understand: we know why information is being requested and what it will be used for.
The French data protection authority CNIL describes data minimization as keeping personal data proportionate to the purpose. This principle in the European framework also gives travelers a useful way to assess a service. If the goal is to compare two departures, why request a passport copy or access to an entire email account? A specific explanation can justify information needed for the next step. A general promise of more personalized travel does not, by itself, answer that question.
Reading a booking is not authority to change everything
- The scope of access matters as much as the volume of information. A tool that reads a reservation should not automatically gain the ability to change it, cancel it or alter the associated loyalty account. The channel actually has to support that separation. If it does not, the service needs to explain the limitation and its consequences. A connection should not be presented as read-only if it also enables actions the traveler never discussed.
- Login credentials deserve particular attention. Handing over an account password is not the same as sharing a seat preference. The account may contain other trips, information about relatives or saved payment methods. Where providers offer them, we would prefer limited permissions to unrestricted account access. But those permissions cannot be assumed to exist. An assistant cannot promise narrowly defined access if it is actually using a session that opens the whole account.
- Email makes the issue easy to see. Finding a confirmation can be useful; reading all personal correspondence to do so is a different proposition. Travelers should be able to provide the relevant document or select the accessible messages where the product supports it. That matters especially when email is also used for work. A travel assistant does not need to learn the contents of a negotiation or a medical file simply because a booking confirmation is in the same inbox.
- Next, we need to understand where the information goes. The name of the service in the conversation does not necessarily identify everyone involved. An airline, a seller, a technology provider and a support service may play different roles. We expect a clear explanation of who needs the data to make the booking and who might receive it for another purpose. Customers should not have to reconstruct that chain from a series of logos or guess whom to contact with a question about their information.
Plan for access to end
A completed trip does not automatically justify keeping every permission used to prepare it. We would want to distinguish preferences we choose to retain from documents and permissions that are no longer useful. The service needs to explain its retention periods and any obligations that may prevent immediate deletion of certain records. Promising to erase everything at landing would be too simplistic. Leaving access open indefinitely because it might make the next trip easier would be no more satisfactory.
Revoking access needs a clear explanation too. Disconnecting an assistant should end its future access as described, without being presented as cancellation of tickets already bought. Travelers need to know what stops, what stays with the seller or airline and what remains in the service. That avoids two unpleasant surprises: believing a booking has disappeared when it still exists, or believing a tool can no longer act simply because its conversation window was closed.
People sharing a reservation may want different boundaries. One person can organize the trip without wanting to give everyone’s information to an additional service. Buying tickets together does not remove that concern. The assistant should request necessary details at the appropriate stage and clearly state who it is acting for. A name appearing in a conversation should not be treated as general authority to inspect that person’s accounts or documents.
We also expect discretion in the responses. A summary sent to a group conversation should not repeat sensitive information that serves no purpose for the group. A booking reference, an identity document or bank details do not belong in the same category as a flight time. An agent can help share the itinerary while keeping detailed records for the people concerned. An interface that knows a great deal should be able to display only what the situation requires.
The more an assistant can do, the more understandable and easy to withdraw its access needs to be. We would value a service that asks for information when it becomes useful, explains the purpose and then knows how to work without it. A permanent connection to all our accounts is not, in itself, a more efficient service.
Payment needs clear authorization
Saying we prefer a flight does not necessarily authorize payment for it. That difference needs to remain obvious when a conversation replaces several screens. Travelers can compare, ask questions and express a preference without making a final decision. An agent needs to recognize that point and clearly present the action it proposes. A fast service should not force us to watch every sentence in case an enthusiastic response is interpreted as an instruction to charge a saved card.
Approval needs to cover an identifiable amount, currency and order. If the price has changed since the preferred offer was selected, the summary must show it before payment is submitted. The initial budget helps guide the search; it does not always fully explain what will be charged. We also want to see who is selling the ticket and any intermediary fees. The conversation should make that easier to understand rather than separate the fare from several additions revealed late in the process.
Know whom you are paying
The airline’s name on the itinerary does not answer every payment question. Depending on the process, the seller collecting payment and the party responsible for handling a problem may differ. The agent should identify that arrangement before purchase and repeat it on the receipt. That helps travelers recognize a statement description, request an invoice and find the right contact. We want them to be able to tell whom they have just paid.
The payment method also needs to remain a choice. A personal card, a company card or another method may serve different purposes for the customer. The assistant should not select the first available payment method merely because it allows checkout to finish. It can use a preference already authorized for that transaction; it needs to ask when the circumstances fall outside that authority. The convenience of a saved card is useful as long as it does not obscure the decision of the person bearing the expense.
Bank verification steps should not be treated as obstacles to remove. If the intended payment process requires the cardholder’s involvement, the service needs to hand over control and explain what is being approved. This is not about multiplying confirmations for their own sake. It is about preserving the meaning of an action that commits money. A code, notification or authentication request should not become another detail collected in conversation without the traveler understanding the transaction it relates to.
We would also look closely at how payment information is retained. If a provider offers a wallet or limited authorization, it needs to explain what that actually permits. Customers should not have to paste full card details into an ordinary conversation to use an agent. A supposedly more convenient service also needs to make sense when someone wants to replace a payment method, withdraw authorization or stop using it.
Separate the bank activity from the travel outcome
A transaction visible in a bank account does not, by itself, explain the booking’s status. American Express, for example, distinguishes pending charges from posted transactions. That information needs to be read according to its status and the issuer’s rules. The seller, meanwhile, needs to confirm what was purchased. We do not expect an agent to infer ticket issuance from a bank message alone, or to declare definitive failure because the two systems do not yet show the same result.
In a hypothetical example, payment may appear to have progressed while the trip confirmation has not returned. The useful response is to check the existing operation and flag the uncertainty. Immediately starting another purchase to obtain a clearer response can create a second problem. The service should be able to reconcile the seller’s information with the payment information without unnecessarily exposing card data. If it cannot, it needs to give the traveler what they need to contact the relevant parties.
A possible refund requires the same precision. A request, its approval and money actually received are different stages. A conversation with an assistant cannot establish that a refund will be granted, completed within a universal deadline or automatically replaced by a successful card dispute. The ticket conditions, circumstances, seller, card issuer and applicable rules matter. An agent can help gather documentation and track a request. It should not turn that assistance into a guarantee that the money will be recovered.
Additional payments also need to stay connected to what they buy. If a seat or bag is charged separately, travelers need to recognize the transaction. An overall total can be clear enough for a decision but inadequate when something goes wrong. We would want a summary linking each expense to its result: ticket confirmed, service added or request still pending. That avoids confusing the main purchase with services whose confirmation or handling may follow another process.
We would happily delegate payment preparation, with a clear limit: we should always recognize the commitment being made in our name. The agent needs to stop when the agreement no longer covers the proposed purchase, then describe the result honestly, including any uncertainty that remains.
An interrupted transaction must not create two bookings
Demonstrations tend to show a booking that runs without interruption. Everyday use also brings lost connections, expired sessions and slow responses. The issue then is to establish what has already happened. An interface freezing does not prove the request failed with the seller. An agent needs to preserve that uncertainty rather than erase it by trying again. Travelers should not pay twice so the assistant can obtain a success message.
We need to distinguish a search that can be repeated from an action that may already have committed to a purchase. Refreshing schedules is not the same as resubmitting a reservation or payment request. The agent should record the stage reached and any references already received where the channel provides them. That record should help it check the existing result before acting again. It should also be transferable to a human without asking the customer to reconstruct the entire sequence.
Check before trying again
A duplicate is more than a nuisance in a trip list. American’s reservations FAQ states, for example, that duplicate trips on hold may be canceled. That does not predict the outcome of every repeated purchase. It does explain why an agent should not create extra records for reassurance. Before another action, we would want to know whether a reservation already exists, which passengers it covers and its payment or ticketing status.
Imagine a connection dropping just after a trip is approved. The assistant can show that the response is missing, seek confirmation through the authorized channel and ask for help if needed. It should not announce “nothing happened” without establishing that. If it proposes another attempt, it needs to explain the basis for it and what remains unknown. Transparency matters especially when the price changes between operations: the customer needs to understand which purchase they are actually approving.
Multiple parties make this more complicated. A traveler may check the airline’s website while the agent contacts the seller, or call an agency after losing patience. The service should allow further actions to be paused and the information already obtained to be shared. Good assistance avoids everyone trying to solve the problem independently by creating another booking. We want a coordinated recovery, with a clear account of what is in progress, rather than competing booking attempts.
Hand over an accurate record
- Handing over to a person therefore needs more than a screenshot of the final screen. It should convey the requested itinerary, accepted offer, time of the attempt and useful references, without sending more personal data than necessary. That account makes it possible to find the relevant operation. It must not portray an unverified outcome as certain. A precise sentence about what is missing is more helpful than a convincing story that leaves the next person to discover the booking’s real status.
- A partially completed booking needs care. If an additional service is pending, repurchasing the whole trip would be a disproportionate response. If two passengers have different outcomes, that needs to be shown separately. The assistant should preserve those differences and explain the remaining options. Customers can then choose the solution that fits, with the seller’s help where needed. A simplified interface must not collapse different outcomes into a single status that accurately describes none of them.
- The product should also know when to stop retrying. The acceptable number of attempts depends on the channel and the operation. We simply expect a persistent problem to change the service’s response. It can suggest resuming later, a contact or another authorized approach. Repeating the same action without explanation while the traveler watches an animation spin does not amount to handling the problem.
- The records need to remain readable after the incident. An abandoned attempt must be distinguished from a confirmed purchase, and a correction request from a correction completed. A dated history can prevent someone disputing the wrong transaction or confusing two similar amounts. We do not need every technical exchange. We need to find the decisions and their outcomes. That is especially valuable when a problem is not resolved immediately and someone else takes over the following day.
The ability to resume an operation properly matters to us as much as speed. An autonomous agent needs to preserve the existing booking when a response is missing, even if that means explaining an interruption. That caution should not depend on a traveler recognizing the danger of a second click unaided.
Change a trip without losing what mattered
Changing a ticket often begins with a very short request: leave earlier, return a day later or avoid a connection. Behind that sentence, travelers generally want to keep the rest of what they bought. The agent therefore needs to understand what changes and what stays. It should not treat the reservation as a blank slate. The original fare, passengers, services and still-relevant constraints are the starting point for the new search.
The distinction between a change fee and a difference in fare needs to stay explicit. Delta’s flight-change page explains, with conditions depending on the ticket and route, that having no change fee does not eliminate a possible fare difference. We are not turning that into an identical rule for every airline. Our expectation for the assistant is to explain the calculation applying to the booking, with its conditions, rather than describe a change as free because one fee line is zero.
Compare the trip before and after
- The summary should show the old itinerary, the new one and the additional payment required, or how any favorable difference would be handled. Money returned to a payment method, a credit subject to conditions and no remaining value returned are not equivalent outcomes. The agent must not choose whichever word sounds most reassuring. It needs to report what is actually offered for that ticket and obtain approval before committing to a change that falls outside its delegated authority.
- Additional services need to be checked at the same time. A seat, bag or requested meal on the old flight should not be described as transferred without confirmation. The service needs to explain what appears in the new booking and what requires separate handling. This is also a matter of choice: a seat accepted for a short trip may no longer suit if the duration or cabin changes. Travelers need to see those consequences before treating the change as complete.
- A partial change can affect the rest of the journey too. Leaving earlier has a different value if it means arriving at another airport or making the next connection awkward. We expect the assistant to examine the itinerary the traveler will actually be left with. That does not mean it can guarantee every event during the trip. It means it needs to flag a consequence visible in the new proposal rather than focus only on the segment it was asked to replace.
- Who can make the change also depends on the booking and channel. Delta notes, for example, that particular conditions and fees may apply to some tickets issued by a third party. Promising the same process for every reservation would therefore be misleading. Before offering to execute a change, the agent needs to establish who can act and which details are accessible. If the seller has to handle it, that limitation should appear promptly, with the references needed to continue the request.
An announced capability is not a universal rollout
The initial use case announced for Amadeus Max for Travelers is voluntary rebooking. The vendor describes a connection between conversation, airline-specific decision-making and the systems handling rules and transactions. That is a concrete approach to this need. It does not prove an outside assistant can already change any ticket, or that every client airline has deployed the same functions. The capability needs to be checked for the service actually being offered to the traveler.
Final approval needs to cover a quote that is still valid when the action is submitted. If the offer is no longer available or the additional cost changes, the agent needs to return to the boundaries agreed with the customer. It should not silently substitute another option to finish the change. In a hypothetical situation, travelers may prefer to keep their original flight rather than pay more for an advantage that has become marginal. That choice should remain available for as long as the operation’s actual status allows.
After the change, the documentation needs updating. Travelers need to know the new booking is confirmed and, where applicable, that the ticket has been reissued or the required documents are available. They need to recognize that the old itinerary has been replaced, without confusing it with the new one in their calendar or email. If part of the process remains pending, the service needs to identify it. A requested change is not yet a change someone can rely on to travel.
The right outcome is a different trip whose price and consequences we understand. We should retain the choices that mattered or have clearly accepted changing them. An assistant can make this much less frustrating by gathering information and preparing options. It becomes truly useful when it supports the move from one confirmed booking to another. Between those points, travelers need to retain control of the trade-offs and receive a result they can verify without starting the whole investigation again.
When a flight is canceled, the assistant has to handle more than the booking
A canceled flight quickly reminds us that booking is only part of a trip. Here we mean a disruption to the flight, not a reservation removed because the airline disputes how it was made. In this situation, an assistant expressly authorized to act could be a great help. Its role would go beyond announcing the cancellation: it should help us understand which options still fit our trip.
Consider an entirely fictional example. A couple needs to get home to their children. Their evening flight is canceled. A connecting itinerary would arrive the next morning, while a later nonstop departure would preserve a full night’s sleep but delay their return. The agent cannot identify the right choice from the arrival time alone. It needs to know whether someone can look after the children, whether either traveler works in the morning, and whether spending part of the night in transit is acceptable. A preference saved several months earlier will not necessarily answer those questions.
Time matters, but rushing can make an already difficult trip more stressful. Sometimes a decision is needed before the last seats disappear. That calls for brief answers, clearly organized options, and someone able to act. It does not justify presenting two very different journeys home as equivalent. We would expect the decisive differences to be stated plainly: arrival that day or the next, a night at the airport, a change of city, or the group traveling separately. Under pressure, one precise sentence is more useful than an attractive itinerary whose complications emerge along the way.
A replacement itinerary also needs to account for the rest of the trip. A new flight may look suitable while making a booked train or car pickup impossible. In our example, arrival at another airport could move the problem rather than solve it. We would not expect the agent to guarantee every ground connection. We would want it to identify what it has not checked and let us decide whether that uncertainty is acceptable. Reaching the right city is not always the same as getting home.
Refund rules make that choice more important. For flights to, from, or within the United States, the DOT explains passengers’ refund rights after a cancellation or significant change when they decline the offered alternatives. Accepting other transportation or a credit is therefore more than a way to dismiss a notification. The agent needs to make the decision clear, without presenting a commercial offer as the only available outcome.
Where European passenger rights apply, a cancellation can also entitle passengers to a choice between a refund and rerouting, as well as assistance. Compensation has separate conditions. We would not want an assistant to promise a payment before establishing the facts, or to consider everything settled because it has found a seat. Transportation, care, and possible compensation are different questions. Finding an immediate solution should not make the others disappear.
At the airport, the most urgent need may be very ordinary: knowing where to wait and who can confirm a hotel room for the night. A general explanation of rights does not provide a room key. If the agent simply directs us to a document after saying it is handling the situation, the service will fall short. Its value will depend on what it can actually obtain or help us request. When a person needs to take over, they should receive the information already gathered so that we do not have to tell the whole story again.
Seat availability also needs an honest explanation. A seat found at the start of a search may no longer be available when we respond. The agent should not defend its first suggestion as though it were secured. It needs to return with the current options and explain what has changed. A traveler may accept a different compromise when they understand why the better option disappeared. They will be much less understanding if, after several reassuring confirmations, they discover that nothing was ever secured.
Switching airlines raises further questions. The new itinerary may change baggage arrangements, seating, or requested assistance. We want to avoid having a replacement presented as an exact copy of the original booking. If the agent cannot confirm something that matters to us, it needs to say so before we decide. Sometimes getting a seat is the overriding priority. At other times, that seat is of little use if an essential condition of the trip can no longer be met.
The handoff to a person will therefore matter a great deal. During a disruption, a customer does not necessarily want to defend a software system’s reasoning at an airport counter. They want to explain what has happened and understand their options. A useful assistant should be able to prepare that conversation: which flight was canceled, which alternatives were offered, and what decision remains. That summary does not need to recount every step of the search. It needs to help the person taking over understand the problem quickly.
We would pay particular attention to the final confirmation. Saying that a way home has been found is not enough if we still do not know which itinerary we will actually take. The agent needs to distinguish an option under consideration from a completed change. Only then can it help notify the hotel, rearrange a transfer, or change other plans, within the agreed limits. Things become dangerously confused if each service receives a different version of a journey home that remains uncertain.
The best outcome will not always be the earliest departure. It may be the one that lets people travel home together, sleep, or retain essential assistance. We expect an agent to help us choose when plans have gone wrong, without replacing our priorities with whatever seat is easiest to obtain. Once the journey home is secured, the remaining paperwork can wait for a calmer moment. The traveler still decides what is urgent, even when a service helps deal with it.
Families, companions, and individual needs
Booking for several people is not simply repeating the same purchase. A group may include adults with different constraints, a child who needs an accompanying adult, and someone who would find a long connection particularly difficult. The agent needs to understand what these passengers need to do together. Displaying four available seats does not yet answer that question. A trip can be technically possible for each traveler and still be poorly arranged for the group.
Imagine a family accepting a connection to travel during school vacation. The parents would prefer not to arrive too late, but do not need everyone in the same row. Having a young child alone at the other end of the cabin, however, is not part of their plan. Simply saying that seats will be assigned later leaves a decisive point unresolved. A request submitted, a possibility described, and an assignment actually confirmed are different things. The conversation should not compress all three into a reassuring “everything is arranged.”
The DOT’s family seating dashboard shows that airlines’ commitments differ and have conditions. One carrier’s policy should not become a general guarantee. The useful answer for a traveler concerns their own flight and reservation. An agent should not cite an old announcement instead of checking what is possible now.
The number of reservations deserves the same attention. Booking passengers separately can sometimes look easier on a screen, particularly when not all the seats appear in the same way. A family may nevertheless want to keep one reservation so that they can follow the trip and handle changes together. If a different arrangement is being considered, it needs explaining before purchase. Customers should not discover during a disruption that the assistant divided them among several bookings without explaining the consequences.
Individual needs extend beyond seating. A stroller, equipment, a requested meal, or assistance can matter more than a small difference in flight time. A good assistant should let travelers explain those needs without implying that a sentence in the conversation amounts to the airline’s acceptance. It should be clear which requests have merely been recorded and which have received a response. We are not asking for an impossible guarantee of every detail. We are asking for a distinction between what the customer wants and what has actually been arranged.
An aircraft or itinerary change may require checking again. What worked on the original flight does not automatically carry over to the next one. For our fictional family, the quickest solution might not preserve the conditions they originally accepted. After a disruption, the agent should therefore retrieve the group’s constraints, not just their names and destination. It becomes useful if it saves us from repeating an important request while saying clearly when that request needs to be confirmed again.
It also matters who is speaking for whom. An adult can book for a spouse or organize a trip with their parents without having unrestricted authority over everyone’s choices. A cabin preference, a long connection, or a changed return date may need their agreement. We would not want an agent to treat the group as one person because one account started the search. It can prepare a shared response, but a neatly organized presentation does not resolve real disagreements.
Separate itineraries may themselves be a perfectly legitimate choice. One traveler goes home early; the others stay a few days. In that case, the service should clearly show which parts of the trip remain shared and which do not. The aim is to respect the travelers’ intended arrangement rather than impose the same schedule on everyone. A change affecting one person should not be described as a change to the family’s trip before it is clear exactly who is affected.
Finally, practical information needs to remain available to the person who will use it during the trip. Keeping everything in the booking traveler’s conversation can be awkward if people separate at the airport. Each person should be able to retrieve their schedule and useful details without relying on an improvised explanation. That is an expectation about organizing a trip, not permission to share everyone’s data freely. We want the service to help the group travel while leaving each person in control of their own information.
We will therefore judge an agent by how well the family’s trip holds together, not merely its ability to fill out several passengers’ details. Did it preserve what needed to remain shared? Did it make important limitations clear? Did it let everyone understand the chosen arrangement? This is where an assistant should improve on a form without assuming the right to decide for the whole group.
Miles, status, credits, and vouchers: the price is not the whole picture
For a frequent traveler, the displayed price does not always settle the decision. Miles can be spent or saved, a flight may contribute to status, and a credit may be approaching expiration. Their usefulness is personal and depends on how someone travels. An agent that ignores them may suggest an option that makes sense for somebody else. One that assigns them an automatic value risks the opposite mistake: justifying any detour for a benefit we may never use.
Consider a hypothetical situation, without assigning a cash value to a mile. A traveler has a balance they intended to use for a trip for two. The agent finds an option for a business trip and presents the lack of a substantial cash payment as an obvious saving. Spending that balance now, however, removes an option from the personal trip planned for later. It may still be a good decision. That cannot be established simply because miles are available. A loyalty currency is not free simply because spending it does not appear as a cash charge.
The service therefore needs an understandable preference: include mileage options, use miles only after comparing them, or save them for another trip. We do not need to agree on a universal value for every point. We need to know which balance will be used and what it will buy. If a search cannot find the necessary seats for all travelers, that limitation needs to appear immediately. An offer for one person is not yet a solution for a couple, however attractive it looks.
Redeemable miles are different from the points and flight counts used to qualify for status. United’s page on Premier qualification distinguishes Premier qualifying points (PQP) and, for 2026, Premier qualifying flights (PQF); it announces that PQF will be eliminated on January 1, 2027. An agent therefore needs to check the kind of benefit and the year involved. Saying a flight earns a lot of miles does not establish that it will help us retain status.
It also matters who receives each benefit. On its checked bag fee calculator, United makes some baggage-fee exemptions for companions of Premier members conditional on being on the same reservation. That detail is enough to show why the frequent traveler’s account does not explain everything. An agent should not promise the same conditions to every companion without examining the rules that apply to the booking.
Credits and vouchers have their own terms to check. They may have an expiration date, a named holder, or restrictions on use. We will not assume that all credits work alike because they carry an airline’s logo. The service needs to read the conditions of the particular credit it intends to use. We should also be able to see how much will be spent and what will remain afterward. Using a credit to complete a booking is not necessarily the best use of it, especially if that changes the itinerary that suited us.
The choice between a refund and a travel certificate is particularly sensitive after a disruption. In the information accompanying its refund form, United says that an eligible refund goes to the original payment method unless an electronic travel certificate is requested. We would want the agent to keep that distinction clear. A credit may suit someone traveling again soon. It can be far less useful to someone with no next trip planned. Presenting one as automatically equivalent to the other would take a decision away from the passenger.
A ticket change can also call loyalty expectations into question. The marketing carrier, operating carrier, and fare conditions may affect the benefits a traveler hopes to receive. Recognizing a familiar brand in the itinerary is not enough. If the agent could not verify the expected earnings or a particular benefit, that uncertainty needs to remain in its answer. A useful estimate is better than an unjustified assurance. Sometimes we prefer an itinerary that is less rewarding for the program but better for the trip. We still need enough information to make that choice.
Easy account access should not determine how rewards are used. An available balance is not permanent permission to spend it on every new request. We would expect an explicit decision when an action uses miles, a certificate, or a personal voucher. This would also distinguish between travelers: the person organizing the booking is not necessarily the person whose rewards will be used. A conversation can simplify planning. It should not obscure whose rewards are being used.
Loyalty is worthwhile when it improves our travels. It becomes a burden if an assistant turns it into a compulsory goal, or uses it to portray unplanned spending as a bargain. A good service should be able to explain a proposal with those benefits and without them. We can then decide what is worth using today.
Business travel: the traveler’s preferences are not the only rules
A business trip can look like a personal trip until someone asks how it was booked. The employee chooses a departure time and boards the same airplane, but may not be free to choose the booking channel, fare, or payment method. That does not make their preferences unimportant. It means an assistant has to work with several constraints set by different people. A pleasant itinerary can still be unusable if nobody is authorized to buy it on those terms.
Take a fictional company that requires a travel agency and prior approval for certain trips. The agent finds a flight that suits the employee perfectly through another channel. Buying first and offering to sort out approval later does not help them. Whether the price was attractive is not the question. The purchase bypassed an arrangement the traveler will have to account for afterward. An apparent saving can turn into extra paperwork, uncertain reimbursement, or time spent getting the booking corrected.
Some business travel tools already support this kind of approval. SAP describes an arrangement in its Concur Approve then Book course in which a request must be approved before booking. It is one possible setup, not how every company operates. The example nevertheless illustrates a practical limit: being able to complete a purchase does not confer permission to skip the step that authorizes it. The assistant needs to know the rules for this particular trip, without applying a supposedly universal policy to every employee.
The time needed for approval must remain clear. A flight found now may no longer be available when the manager responds. We would prefer the service to explain that uncertainty rather than present a pending request as a secured trip. It could prepare useful options for the decision, identify what has not been booked, and check availability again after approval. That preparation has value even when it does not immediately end with a ticket purchase.
Company rules do not say everything about fatigue either. Imagine an early meeting following a late arrival. The cheapest permitted option is not necessarily the most sensible way to prepare for the next day’s work. The agent could help explain an alternative and the reasons for it, without inventing an exception to the rules. That would make the disagreement easier to address: here is the itinerary that fits the existing policy, here is one that would allow more rest, and here is what needs further approval. The choice would no longer be hidden in a ranking.
Support after booking deserves just as much attention. A company may have arranged after-hours support with its agency, established ways to track trips, or familiar contacts. We do not assume those services exist in every contract. Where they do, however, booking elsewhere may change how they can be used. The traveler should know this before discovering abroad that the person they call cannot see the reservation. The agent needs to explain who will actually help if the meeting is rescheduled or the return flight needs to be moved earlier.
Personal benefits add another limit. An employee may prefer an airline to retain status; the employer may have other priorities. The assistant should not settle that disagreement silently by favoring one side. It can show the options and leave the decision to someone authorized to make it. Nor should it spend a personal credit because that makes an employer-funded trip easier to book. The source of the money and the beneficiary of the trip are not necessarily the same, and the conversation needs to keep that distinction clear.
After purchase, information needs to reach the right people. That does not mean sharing an entire work conversation with another provider. It means knowing where the trip will be tracked, which documents will be available, and how a change will be reported. An agent that books correctly but leaves the employee to assemble the paperwork merely shifts some of the work. We would instead expect it to simplify the company’s existing process, within the access it has been given.
There is another familiar business travel situation: the purpose changes before departure. In a fictional example, a meeting moves to a video call. An assistant focused on finding a better flight time might keep refining a trip that is no longer needed. The employee should be able to explain the change without automatically canceling every service. There may still be another appointment at the destination. Understanding why someone is traveling can sometimes prevent a purchase; it does not confer the right to abandon the trip without discussing it.
We therefore see a useful role for an assistant before purchase: preparing an understandable request, comparing constraints, and identifying what needs a decision. The impressive act of booking in a few seconds is not always the most useful contribution. On a business trip, saving time can also mean presenting a plan that the employee, manager, and agency can actually carry out, then retrieve and change without unnecessary difficulty.
Accessibility: getting help while remaining in control
A service that makes booking easier for some people can make it harder for others. A conversation can help someone read an offer, organize information, or explain a request. It can also become an obstacle if it imposes its pace, loses previous answers, or replaces precise information with reassuring language. We would not measure accessibility simply by the ability to type a sentence into a window. People need to understand the response and act without the service forcing them to depend on a companion.
United’s exception for assistive technology matters in this respect. In its third-party agent FAQ, the airline says these disability-related uses are not affected by its new policy. That clarification is not blanket permission for autonomous purchasing. Equally, using technical assistance should not automatically be portrayed as an attempt to bypass the rules. The distinction needs to survive all the way through the customer’s interaction with the service.
Imagine a traveler using a screen reader to check a connection; this is not an account of a real person. A lengthy answer that sounds pleasant is not enough if flight times, airport changes, and the connection time are buried in the narrative. The service needs to let them retrieve those details, compare them, and ask for clarification. That is no reason to decide on their behalf that an itinerary would be too complicated. Help should make choosing easier rather than arbitrarily narrow the available options.
Consider another hypothetical traveler who needs airport assistance. That need cannot be reduced to a label stored in a profile. What matters is what was requested for the actual itinerary and what has been confirmed. After a change, an earlier response may no longer be sufficient. We would want a clear record of what needs checking on the new itinerary, without having to explain the entire situation again to each person. The agent can prepare that handoff without claiming to deliver a service it does not itself provide.
In the United States, the DOT’s published Bill of Rights brings together rules that include accessible information and assistance. Their scope and applicable conditions still matter. We would not use them to promise the same response to every request at every airport worldwide. They do, however, remind us that accessibility is more than a pleasant optional feature in an interface. The agent must not make it harder to reach the information and people the passenger needs.
Getting help from a person should remain straightforward. If a traveler says the automated answer does not resolve the question, asking them to rephrase indefinitely is not assistance. A handoff should preserve what has already been explained while letting the customer decide which information to share. Medical details are not something to collect out of curiosity or infer from a seat preference. We expect questions relevant to the current request, asked understandably, without improvised diagnoses.
The same person will not always use the same kind of help. They may want to plan a simple itinerary independently and call about a more complicated one. The service should neither confine them to a dependent role nor refuse a human contact because they used the tool before. That flexibility would benefit many travelers. Fatigue, a poor connection, or a stressful situation can make a normally easy task difficult. A good solution remains usable when conditions are no longer ideal.
Useful answers should also be available to reread outside the original conversation. A departure time, meeting point, or booking reference should not become impossible to find because the exchange has grown longer. That helps travelers prepare independently or ask for help with one particular matter without sharing their entire history.
Independence therefore means being able to choose the help you need, understand what is confirmed, and reach a person without starting the booking over. We would distrust a system that celebrates the absence of human intervention while the customer is asking for it. We would be much more receptive to an assistant that knows when to step aside. That would be a practical way to respect the traveler rather than ask them to adapt to a technology demonstration.
Commercial interests do not disappear in a conversation
An agent can feel entirely on our side because it speaks in the first person and remembers our preferences. That familiarity says nothing about how the service is funded. It might be paid by the customer, a supplier, or through another arrangement. We will not attribute a particular business model to an organization without evidence. We do, however, want to know whether a commercial interest can influence a proposal presented as the best choice for our trip.
Consider two fictional services. One earns money for each booking; the other through a traveler’s subscription. The first might have an incentive to complete a purchase, the second to persuade the customer to keep subscribing. Neither model proves that advice is poor or independent. They simply raise different questions. In both cases, we would look at what happens when the most useful answer is to buy nothing, wait, or keep a perfectly satisfactory existing ticket.
Presentation matters a great deal. A preference we have supplied is not the same as a provider’s commercial preference. If we ask to prioritize an airline to use a credit, the ranking can reasonably reflect that. If the airline appears first because it pays the service, we want to know. The two results may look alike on screen while having very different explanations. A conversation should make those explanations easier to understand, not use its natural tone to blur them.
In France, information requirements for comparison sites cover matters including rankings and the influence of remuneration. We are not saying that every assistant automatically falls under the same rules. The editorial question remains valid beyond that legal classification: why does this proposal come before the others? Saying that the system understood our needs is not enough when the order may also reflect a commercial interest. A short, visible explanation would be better than a passing reference buried in the terms.
The issue extends beyond possible payments from an airline. A service may be better equipped to handle certain carriers or transactions than others. Limited coverage is acceptable. Presenting it as a complete comparison is much less so. If a recommendation really means “the best option among those this service can handle,” that limitation needs to accompany the result. It changes the scope of the advice without necessarily making the service useless.
Travelers also need to understand the consequences of saying no. Imagine the agent proposes a ticket with an additional service we do not want. Declining should not make the entire booking incomprehensible or cause us to lose the comparison already prepared. Decision support should not become a series of subtle pressures to retain every option. Where the offer permits it, we want to be able to accept the itinerary and decline an extra while clearly seeing what remains and what changes.
When a provider gets paid may also help explain its behavior after purchase. A service paid at the point of sale still needs to explain what it will do if a problem arises later. That does not prove it will lose interest in the customer; it gives us a reason to ask before needing help. Who will respond on a Sunday? Is another support request covered by what we paid? We are not asking for every task to be free. We are asking to understand the service actually purchased with the ticket.
An assistant should also be able to explain a recommendation without hiding behind its own complexity. Travelers do not need every calculation to understand the main reasons: timing, duration, fare conditions, a stated preference, coverage limits, or a commercial relationship. If the answer cannot distinguish between those factors, the familiar tone becomes misleading. We would be asked to trust a conclusion without knowing what carried the most weight in a decision that commits us.
We would also want to compare elsewhere without losing the work already done. A clear summary of the criteria and selected offer would let us seek another opinion. The service could explain the limitations of that comparison. It should not suggest that its own conversation is the only way to understand the trip. Trust does not require customers to give up checking.
This transparency will not resolve every disagreement. We may understand the business model and still find the advice unconvincing. That is normal: an explanation is there to enable a choice, not automatically secure our approval. A service that accepts that discussion seems more credible to us than one that presents every proposal as the indisputable result of personalization. The next question is which offers it can actually consult, and on what terms, because that access also shapes its recommendations.
NDC, reservation systems, and the cost of searching
Behind a simple conversation, booking involves systems that are more than a web page. Airlines assemble offers, distribute them through different channels, and then need to keep track of the services sold. Technical terminology can make a modern connection sound like the solution to everything. It needs careful explanation instead, so people can understand what it makes easier and what still depends on the partners’ decisions.
IATA describes NDC as a data-exchange format built around offers and orders. That shared language can help systems communicate. Its existence alone does not provide access to every airline’s offers. Using the standard and being authorized to use a connection are different things. Providers should not present their technical expertise as evidence that all carriers have opened their systems to them.
Lufthansa Group’s passenger NDC terms make the distinction tangible: they provide revocable, nontransferable access that cannot be shared with unauthorized third parties. They also require credentials to be kept confidential. An airline can offer a modern distribution channel while placing conditions on its users. This is not a specific prohibition on AI; it sets access limits that exist independently of an assistant’s marketing.
The same document covers shopping costs up to a ratio of 500 searches per net ticket issued, with possible measures beyond that, including access deactivation. This is a provision of that contract, rather than a universal IATA rule or a promise of free searches for each traveler. The number primarily shows that automated search is not treated as costless by the systems producing the offers.
An agent might be tempted to explore every date, airport, and combination endlessly. That abundance can seem attractive to a customer who sees only the result. It becomes less convincing if the same queries are repeated without producing anything useful. A good search should move forward: incorporate a confirmed criterion, rule out an irrelevant option, and avoid repeating exactly the same work. The number of searches alone does not measure the comparison’s relevance.
In its October 7, 2025 article on search volumes, IATA links their growth to costs and system load. It presents approaches involving performance metrics and caching as a starting point for discussion. We infer no mandatory industry-wide limit from that. The issue instead invites us to judge an agent by its ability to find a relevant offer without multiplying unnecessary operations.
Caching also needs an honest explanation. Reusing information can avoid an unnecessary search, but should not turn an old offer into something presented as freshly confirmed. An assistant can retain a preference while checking the offer again before taking action. For travelers, those two forms of memory serve different purposes: one preserves the meaning of the request; the other describes an offer that may change. The service should distinguish them without asking the customer to monitor its technical architecture.
Travelers can benefit directly from that discipline. An assistant that understands the request better does not need to present a hundred almost identical alternatives. It can explain why a few options deserve attention and which it ruled out. Efficient searching should not, however, become an excuse to hide narrow coverage. The balance needs to combine useful searches, sufficiently current information, and an honest account of what was actually compared.
Existing distribution routes do not disappear when a new format arrives. The 2021 agreement between Air France-KLM and Travelport already illustrated NDC distribution based on contracts and defined access. That historical example is not a current access offer to any developer. It shows that a conversational interface can rely on an organized distribution chain rather than automated visits to a public consumer website.
What happens after the sale also matters. The ONE Order initiative aims to bring information currently spread across different records and documents into an integrated order. IATA describes a gradual transition. It would therefore be wrong to assume those separate references have already disappeared everywhere. An assistant needs to work with the systems actually used by the participants, rather than the promised simplicity of a future model.
That explains why an integration may be good for search but less complete for changes or refunds. The exchange format does not guarantee that every function exists on every connection. Travelers do not need to know every detail, but the seller needs to describe what it handles. A promise of universal support should not be inferred from a mention of NDC compatibility. Operational capability and available contacts still need to be established.
Distribution economics also help explain airline positions without automatically excusing them. In its 2026 annual report, Ryanair discusses possible costs from new AI intermediaries and also describes its own support bot with human escalation. These positions are not inherently contradictory. Objecting to the cost of new intermediation does not mean rejecting every use of AI. Travelers, for their part, can ask whether the chosen arrangement actually improves their experience.
We therefore want providers to explain the limits of their connections and the results obtained. Good distribution technology can make comparisons richer and ongoing support easier. It does not decide commercial terms, access rights, or responsibilities by itself. Progress will be real when travelers face fewer complications because the participants work better together, rather than because those complications have merely disappeared from the conversation.
Three travel scenarios to consider
The following situations are entirely hypothetical. They describe neither our purchases nor tests of agents. Their purpose is to bring the previous questions together in a coherent journey. We will not assign behavior to a particular airline or product. We will consider what passengers should understand to make a decision, and how a service could help without treating a preference as an implied instruction.
A business round trip where arrival time is decisive
Imagine someone who needs to be in New York for a Wednesday morning meeting. They request a round trip from Paris and want to keep Tuesday evening free to rest. A search focused only on price might suggest a very late arrival. A search attentive to the purpose would recognize arrival time as essential, while still letting the traveler accept an exception if the difference becomes significant.
The agent can then present two or three genuinely comparable offers. It explains the local arrival time, airport, and the requested flexibility for the return. If it can purchase only from certain sellers, it says so without presenting that coverage as the entire market. Travelers retain a clear choice: pay more to preserve the plan or deliberately change the plan. The service’s value comes from that explanation, rather than the number of results displayed.
Now suppose the selected offer changes before purchase. The agent should not silently replace a late-afternoon arrival with a late-night arrival, even if the price remains similar. It returns to the criterion behind its recommendation. The customer can decline, accept, or restart the comparison. Pausing is useful here because it protects the trip’s purpose. Automatic execution that forgets the meeting would mainly accelerate a poor decision.
Getting to the meeting also needs to remain part of the reasoning. A flight’s arrival time is not the time you will reach your meeting. In this example, an attentive assistant would simply recall the desired margin, without inventing a guaranteed transfer time. It would help compare two arrivals against the purpose of the trip, rather than treat the difference as a display detail.
Evaluating the result would finally require looking at what was actually completed. An attractive proposal left pending is not a trip that has been handled. A properly completed reservation needs to remain understandable to the person traveling and whoever helps afterward. We would not want this scenario to become a flattering account of the agent’s autonomy. We would want the trip to match the accepted plan and the passenger to know where to find useful information.
A family whose preferences cannot be reduced to a price
Now imagine two adults traveling with children for a week’s vacation. Their request appears simple: leave on Saturday and return the following Saturday. But they want to stay together, avoid a very late arrival, and reduce fatigue. The agent should not decide for them how much those preferences are worth. It can clarify them and compare options that respect them, clearly showing the extra cost or time involved.
An apparently attractive offer may turn out to be less suitable if it does not include the desired bags or if seating arrangements remain uncertain. The assistant should distinguish what it has checked from what still awaits confirmation. It does not need to promise comfort it cannot control to be helpful. It can explain a difference, find another option, or let the adults make an informed choice.
The trip may also involve a particular request for one participant. That should not disappear among the group’s general preferences. A vague assurance that everything is arranged would not help if the agent cannot establish the request’s status. A good process preserves the important information and identifies the next contact. The aim is a workable trip for everyone, rather than just a shopping cart containing the right number of people.
The comparison then needs to remain understandable to both adults. If one prepared the trip and the other needs to take over, the decisive details should make sense without reconstructing every exchange. In this imagined situation, a shared summary of choices would matter more than a long explanation of how the agent calculated its ranking. The service would also help travelers coordinate with each other.
This scenario also shows that speed does not offer the same benefit to everyone. A few minutes of explanation may prevent a decision that is difficult to revisit after purchase. Quality would be measured by the adults’ understanding and whether the selected services work together for the family. An assistant asking a relevant question at the right moment can help more than a tool instantly turning a family request into a reservation.
A disrupted return while the traveler is already away
Our final imagined situation involves someone learning during a stay that the return journey cannot proceed as planned. They ask for help finding a solution. The priority changes. The task is no longer simply comparing an ideal trip, but understanding the existing reservation, the options actually offered, and the decisions still open. The agent should retain known constraints without assuming they have all stayed the same since the original purchase.
A new option may appear acceptable by schedule alone while changing something important: the airport, ground travel, or arrangements for the next part of the journey. The assistant should show those consequences and identify who can carry out the action. If it cannot do so itself, handing over the work already prepared can still be very helpful. Having to go through the whole conversation again at the wrong moment would lose part of that benefit.
The person’s priority may also have changed. Initially they wanted to return early; now they prefer a solution easier to arrange from where they are staying. The assistant should let them express that new trade-off. Remembering the trip does not mean mechanically applying an old preference, particularly when the possibilities and circumstances have changed.
Three questions matter here: which option did the traveler accept, what does the booking show afterward, and who is still available if things change again? We assign no particular right or automatic outcome to a carrier here. We are examining service quality. A useful agent should reduce the work without leaving passengers with incompatible answers and nobody able to say which one corresponds to their trip.
The trust we expect airlines and agents to earn
After examining the policies and examples, one simple question remains: how would we recognize a service we would want to entrust with part of our trip? A broad promise of reliability is not enough. We expect understandable commitments that can be checked against what actually happened. They concern the offer selected, the actions taken, and the way the service responds when a problem arises.
The first commitment should concern who is involved. Customers need to be able to find their seller’s name, the airline’s, and the assistant’s role. That information need not dominate every screen. It needs to be accessible, consistent, and lasting. When a question arises, nobody should have to guess which logo represents the right contact or reconstruct the path taken by their request on their own.
The second concerns what the service promises to handle. We would like a plain-language description of the tasks covered: searching, preparing, purchasing, changing in certain situations, and handing over to a person in others. Those limits should reflect the service actually being used, rather than everything its provider hopes to offer someday. A new product can be interesting with limited capabilities. It loses credibility when it implies that current limitations concern only unimportant details.
The third commitment would be to preserve the reasoning behind decisions. The assistant can learn from the conversation, but should not lose what motivated a choice. If arrival time mattered more than a small price difference, a changed result needs to return to that criterion. If a preference changes, the service needs to understand that the previous rule no longer applies. We expect useful memory of trade-offs, rather than an indiscriminate collection of everything said.
That memory should also allow corrections. A traveler may notice a misunderstood criterion before any action is taken. Correcting it should be simple, with a visible effect. The service should not keep recommending the same offers because an initial interpretation has become fixed. Personalization is useful only if people can understand what it retains and make it change its mind. An assumed preference should not become an invisible restriction.
We also expect a clear distinction between what the system knows and does not know. Missing information should not become a reassuring answer. The assistant could explain that one point is confirmed, another depends on selecting an option, and a third requires the seller’s response. That precision gives travelers a useful basis for deciding what to do next. The service need not recite its limitations in every sentence; it should place them where they could actually change the decision.
Another essential commitment concerns failure. Providers should distinguish a search with no result, a rejected action, and a result whose status remains uncertain. These situations call for different next steps. One message inviting people to try again can appear simple while transferring all the difficulty to the customer. We would prefer a short explanation of what was attempted, what is known, and what to do next, without asking people to repeat the entire process unnecessarily.
The quality of human support should be part of the original promise. We are not asking a person to approve every schedule lookup. We want a realistic next step when a case exceeds the automated service’s capabilities. Travelers should know whether someone can take over their case, when, and with what information. Saying support exists is less useful than showing how it connects to the work already done. A transfer should not erase the conversation that helped identify the problem.
Airlines, for their part, could help make permissions understandable. A dated explanation of recognized services, covered tasks, and essential restrictions would help customers and providers alike. We do not expect every commercial agreement to be published. We expect travelers to be able to verify a promise made to them. Permission treated only as a secret among professionals is difficult for the public to use as a basis for trust.
Policy changes deserve the same attention. If a feature is no longer available, customers need to know before building an entire trip around it. If the framework changes between searching and a later action, providers should explain the effect on ongoing service. The point is not to guarantee that nothing will ever change. It is to avoid users discovering a service break alone when they were counting on help that is no longer accessible.
We would also like quality measures that reflect the reality of a trip. The time needed to produce a first response is interesting, but insufficient. We should also look at whether the selected trip meets the traveler’s needs, whether changes are clear, whether transactions succeed, and whether problems are resolved. An assistant can appear slower because it asks a good question, while preventing a series of calls or a time-consuming correction later.
These expectations should remain proportionate to the service sold. An inspiration tool does not need to promise the same support as a seller handling an entire trip. Both should describe their roles clearly. Customers may choose a limited service because it suits them, costs less, or leaves them in control of certain actions. Problems arise when the level of support is discovered after commitment instead of forming part of the choice.
The language of the promise matters as well. An assistant can make the conversation sound natural in French while referring customers to wording they cannot understand for decisive details. We would want it to retain precision when explaining conditions, limits, and what happens next. A smooth translation should not add certainty that is absent from the offer. Simplicity needs to extend to the information people base decisions on, rather than just the sentences welcoming the request.
Trust finally requires a serious response to mistakes. We do not prejudge the responsibilities applying to every situation. We expect participants to do more than send passengers back and forth while invoking technical boundaries. Anyone selling a seamless experience should have planned how to explain a disputed result and guide the customer toward resolution. An agent’s quality will also show in that accountability, without turning customers into investigators of their own bookings.
These commitments do not prevent a service from being pleasant. They can make it simpler because the roles were clarified in advance. Passengers should be able to describe their needs in their own words, receive an understandable proposal, and know what happens next. Conversation then becomes a useful way into a coordinated service, rather than making disconnected steps look like one service when nobody oversees the whole process.
Our position: delegate actions without giving up your choices
We see real value in assistants that reduce the work involved in preparing a trip. Understanding a request, explaining a fare difference, or organizing a coherent comparison can already be a substantial help. We feel no nostalgia for repetitive searches and scattered information. But booking becomes more useful when it helps us understand the trip we have bought, rather than simply requiring fewer actions on a screen.
United’s change demands that permissions be taken more seriously. A passenger’s wishes, an airline’s access conditions, and a provider’s technical abilities cannot substitute for each other. An agent recognizing those differences can offer an honest, useful service. One hiding them behind a promise of autonomy ultimately leaves travelers with uncertainty the assistant was supposed to help resolve.
We therefore favor specific delegation. The assistant should handle repetitive tasks and highlight important decisions. It can ask fewer questions by preparing better proposals. It also needs to recognize when a personal answer remains necessary. A good question about a schedule, condition, or new expense is not a failure of automation. It may show that the service understands what matters.
This approach leaves room for different ways to travel. Some customers will want to make the purchase themselves after an assisted comparison. Others will prefer a service that can act under approved arrangements. Others will need a human contact for part of the journey. Progress should not impose one model. It should allow people to choose suitable help without losing their understanding of the reservation along the way.
Airlines should not regard a restriction as the end of the discussion either. If they want to retain the customer relationship, their own processes need to be understandable and support accessible. Seeking another interface may also reveal a need the current service meets poorly. Addressing that requires more than a clause: readable information, reliable actions, and help available at the right moment.
Agent providers have a comparable responsibility in how they present results. We expect them to name partners, describe what they cover, and distinguish a recommendation from a completed action. We also expect ongoing service to be discussed as precisely as search. A booking is not an isolated moment in a demonstration. It begins a relationship that may last until the return, involving changes, questions, and sometimes difficulties.
Much remains to be built, and it is too early to name a winner based only on current announcements. We can nevertheless judge the direction. A system making offers comparable, respecting permissions, and keeping practical help available is moving the right way. A system producing a confident answer without establishing what has been done mainly makes the problem harder to see. Good writing must never stand in for good service.
The best agent may be the one whose mechanisms we notice least. We will know what we chose, what was bought, and whom to contact. We will be able to correct a preference, understand a new option, and take over without losing the thread. That simplicity requires work behind the scenes, clear agreements, and reliable coordination. It deserves more attention than the promise of a trip arranged with a single sentence.
We want to delegate actions so we can give more attention to the journey. That means keeping our choices, understanding their consequences, and having a service that remains available when plans change. Under those conditions, AI can make a concrete improvement to airline booking. The decisive criterion remains the one we apply to any travel service: does it genuinely help us travel on the terms we accepted?
