A few years ago, the hotel industry put a voice assistant in the bedroom. It could dim the lights, order room service, call the front desk without picking up a phone. It worked, mostly. Guests unplugged it anyway.
They were not reviewing the features. They were answering a question nobody had asked them out loud: do you want an always-listening machine in the room where you sleep? The features were fine. What was unclear was whose side the machine was on. Out came the cord.
Hold that image next to the numbers. Skift and McKinsey asked travel executives about agentic AI, software a traveler can talk to that answers, plans, and increasingly acts: 80 percent plan to deploy it at scale within three to five years. It watches the journey, rebooks the missed connection, and it is learning to pay. My own corner of the business, the one that makes the advertising, has already started publishing guides on how a brand makes itself impossible for an agent to overlook. Everyone is building for the agent. Skift also counted the American consumers willing to let a fully autonomous agent do the booking. Two percent. An Expedia study of more than 5,700 adults found the same shape: two thirds would not trust an AI assistant to book for them.
Booking.com surveyed more than 37,000 people across 33 markets, and 89 percent want to use AI in planning future travel; most already have. What collapses is not appetite for the help. It is willingness to hand over the decision. Only 12 percent are comfortable with AI acting independently, and only 6 percent fully trust what it tells them. The newest survey, from this July, puts trust in an AI’s travel recommendation at one traveler in four. Travelers are not rejecting intelligence. They are refusing agency, and the industry is laying pipe at full speed for the one behavior they have specifically declined.
Travelers are not rejecting intelligence. They are refusing agency.
The industry reads that gap as a maturity problem. People hesitated over online booking too, and over mobile check-in, and both became the default. The argument has weight; every one of those earlier hesitations looks quaint now, and the people making it have been right before. Maybe they are right again. I read it differently, and not from a neutral position: I have spent thirty years building the screens that taught people to look down. The attention economy was not something that happened to my industry. It was our best work, shipped proudly, one delightful notification at a time. So when travelers hesitate in front of our newest machine, I do not hear ignorance. I hear pattern recognition.
When researchers ask what exactly worries people, the answers land in the same four places: loss of control, errors and liability, unease about machines deciding, and data privacy. Forrester, surveying consumers in the US, UK and Canada, found three quarters uncomfortable letting an agent buy anything for them even with spending limits they set themselves. Four concerns, two fears underneath.
The first fear: it will fail you. The booking will be wrong, the confirmation will be imaginary, and no one will be accountable. This fear is earned. Documented incidents from mid-2026 show the pattern: confirmed statuses issued for reservations that were still pending, charges made to stored cards without authorization, trips canceled without anyone asking. Behind all of it sits a question the industry has barely started to answer: when the machine books wrong, who is responsible? When a machine fails you at home, you lose an evening. When it fails you in a strange city at midnight, you understand exactly why people want a human on the other end.
The second fear is the one people say out loud less often, and I think it is the deeper one: it will farm you. Your data, your patterns, your attention, and eventually your price. In a late-2025 survey of US consumers, 82 percent called losing control of their data to AI a serious personal threat, and 81 percent suspected companies were already using their data in ways nobody disclosed. Here is the part that should stop us: in the travel context, 86 percent of Booking.com’s respondents said they were comfortable sharing personal data with AI. They will hand over the data. What they will not do is trust what it gets used for. The objection was never privacy in the abstract. Gartner predicts that by 2028 nine tenths of business buying will pass through AI agents, which is the world those brand guides are written for. We spent a year courting the agent before it existed. It arrived this month, from the company whose business is your data. It will book your travel for a monthly fee and, its chief AI officer said, possibly a cut of the purchase. Nobody is courting it on the traveler’s behalf.
Two fears, one question underneath, and it is a question of allegiance: who does this thing answer to? An assistant that answers to you is a concierge. An assistant that answers to someone else is a salesman holding your calendar. There is a third kind, and real estate already has a name for it: the agent who works for the buyer and the seller in the same deal. A dual agent. Most states allow it only if both sides are told in writing and agree, and a handful ban it outright, because an agent paid by both sides cannot fight hard for either. And here is the uncomfortable part: nearly everything shipping today is a dual agent, and the disclosure is in the terms of service. It answers, at least partly, to someone else. To an engagement metric. To a conversion target. To a data strategy. The fear of being failed and the fear of being farmed are the same correct suspicion, twice: that the machine’s loyalty is undisclosed.
An assistant that answers to you is a concierge. An assistant that answers to someone else is a salesman holding your calendar.
The airlines already run a dual agent everyone loves, and it wears your colors. It is the loyalty program. When flights were grounded in 2020, the airlines borrowed against their programs to stay alive: United took 6.8 billion dollars against MileagePlus, Delta 9 billion against SkyMiles, American 10 billion against AAdvantage. Two of those programs were appraised at more than the airlines that own them. United told investors MileagePlus was worth nearly 22 billion dollars at a moment when the whole airline was worth about 10.6 billion on the stock market. A thing worth more than the airline is not mainly a thank-you to you. It is a currency the operator prints, prices, and reprices. When Delta made its miles harder to earn in 2023 and its members pushed back, the chief executive said the airline had probably gone too far. That it could go that far at all was in the terms the whole time.
This is not a plea for digital minimalism. Solving allegiance takes more technology, not less: systems that check live inventory instead of guessing, that move a refund toward you unasked, that know the moment to produce a human being. Loyalty you can verify.
The worldview I am describing has a lineage, and honesty requires naming it. In 1995, Mark Weiser and John Seely Brown at Xerox PARC called it calm technology: “that which informs but doesn’t demand our focus or attention.” Designers have been trying to ship it for thirty years. The most spectacular recent attempt was a small pin worn on the chest, sold on the promise that you could finally stay present with the people around you. The company wound down, the devices went dark, and my industry drew the wrong lesson: presence doesn’t sell. You cannot sell presence with a camera pointed outward from your chest, any more than you can sell rest with a microphone that never sleeps on the nightstand. The products contradicted the promise; the worldview was never disproven. It is unclaimed where it actually lives: not in a gadget you buy, but in how the services you already use decide to behave.
The newest attempt is on the wrist, and it gets the order right. This month Apple announced a watch that will listen through your day and hand you notes on your own conversations, and it published the limits before the feature exists: the watch ships this week, the listening arrives later in the year, and the rules came first. Opt in, not opt out. No audio kept. Notes deleted after a week unless you save them. Set it to listen at work and never at night.
Travel is where it should be claimed first, because travel is the only thing you buy in order to be somewhere else. The entire point of the purchase sits on the far side of the screen. Every screen between you and the somewhere is charging you twice.
Look at what travelers actually reward. The record is not subtle. The same hotel industry that shipped the unplugged speaker also shipped one of the most widely adopted guest technologies of the past decade, and it was a key. Hilton’s digital key opened two million doors across four hundred hotels in its first year, because it removed the line at the front desk and asked for nothing in return. The feature that has outlasted almost everything around it is simpler still: a chat thread with an actual staff member on the other end, open from the day before arrival until a week after checkout. Narrow, opt-in, and accountable beat broad, ambient, and always-on. Every time.
In the air, the most reassuring AI in the industry is one you never see. At one US carrier, a system reads weather, taxi times and gate availability, then tells the people on the ground whether a departing plane can be held long enough for a tight connection to make it. Nobody thanks it. You just make the flight. That is what allegiance looks like at altitude.
Air travel is also the one place the allegiance question has already been to court. In late 2022, a man booked a last-minute flight to his grandmother’s funeral after an airline’s chatbot told him he could apply for the bereavement fare retroactively, up to ninety days after flying. He could not. The airline’s written policy said the discount had to be approved in advance. When he asked to be made whole for the difference, the airline refused, and then argued before a Canadian tribunal that its chatbot was a separate legal entity, responsible for its own actions. The tribunal called that submission remarkable and held the airline to what its machine had told him. The damages came to about six hundred and fifty Canadian dollars. The finding is worth more than that: the assistant’s promises are the company’s promises, and a brand that ships one of these is underwriting every word it says, whether or not it has thought about it that way.
The pattern is worth spelling out, because it is not the pattern the category is building toward. Travelers reward assistance that is bounded, quiet, on their side, and finished. Bounded means it does one job and does not audition for a bigger one; the digital key opens the door and has no opinion about your evening. Quiet means it does not announce itself. The airline system that holds a plane for your tight connection never sends you a message saying it just saved your trip; you make the flight and never learn why. On their side means that when the trip goes wrong, someone tells you what is happening, and the refund comes to you before you have to ask for it.
Finished is the rarest of the four, and the one worth building for. Nearly every assistant in travel stops one step short. It tells you the train is delayed. It may even offer a rebooking. Then it hands you back to yourself, and you spend the worst hour of the trip stitching systems together: the app that knows about the delay, the phone line that can actually change the ticket, the form that handles the compensation, the hotel that still thinks you are coming. Four systems, four tellings of the same story, and you are the only thing connecting them. Finished means the assistant carries it the whole way and closes it. Seat moved, money moving, hotel told, one message to say so.
Four systems, four tellings of the same story, and you are the only thing connecting them.
Rail is where all four already exist, scattered across operators that mostly do not know they are building the same thing. It is also the mode where the argument is easiest to see, because a train is the only vehicle that gives you a window for hours and knows to the meter where you are.
Bounded, in Switzerland, is a fare rule. Most standard tickets are not tied to a train. Miss your connection and you take the next one, usually inside half an hour, and there is no case to open because nothing has gone wrong. The most advanced rebooking assistant in Swiss rail is the absence of a reason to rebook.
Quiet, on the busiest corridor in America, was invented by riders. In 2000 a group of early commuters asked a conductor to make their car a no-peep zone, and within a year the quiet car was on most trains. Nobody designed it. People organized to protect their own attention and asked the railway to enforce it, which is the clearest signal in all of travel about what attention is worth to the person who owns it. No operator I can find has built that behavior into an app.
On their side, in Britain, has a number attached. The passenger watchdog Transport Focus asks riders how satisfied they were with a journey. On a delay the railway handled well, 92 percent say satisfied. On a delay of the same length handled badly, 26 percent. Sixty-six points, decided by whether anyone told you what was happening. The money is the other half. One small operator east of London has paid delay compensation automatically since 2016: the gate sees when you tapped in and out, the system does the arithmetic, the refund arrives, and there was never a claim to make. In Japan, a bullet train more than two hours late returns the express fee to a digital ticket on its own, the next day. Deutsche Bahn, on the same continent as that gate, will still ask you to post a paper form to an office in Frankfurt and wait six weeks.
Finished is the one the category keeps promising, and rail has already shipped it, in the least glamorous form possible. The best rail assistant on the market can tell you that you are owed money and then has to send you somewhere else to collect it. The gate in Essex does not tell you anything. It pays you. The finished assistant already exists in travel. It looks like a fare gate refunding you, not a chatbot performing helpfulness.
The finished assistant already exists in travel. It looks like a fare gate refunding you, not a chatbot performing helpfulness.
What rail has not done is program the window. One European operator caches its maps on the train, so the screen keeps telling you where you are under the Channel, at 186 miles an hour, with no signal. A Swiss scenic line put a GPS commentary on a tablet, and when a rider complained it was thin, the operator replied that the system provides points of interest and is not the same as live narration. That is the whole gap, in one sentence, from the people who built it. One American operator has installed satellite connectivity on its trains and put nothing on it. The best screen in travel, the one you actually paid for, is still almost entirely unprogrammed, and the industry is spending its money on the other screen.
None of this is new. The oldest service tradition in travel worked it out long ago. The concierges of Les Clefs d’Or, the international society of hotel concierges, take a pledge to do for a guest whatever is “morally, legally, and humanly possible.” Read the order of that sentence. The limits come first, and the favor comes after them. We launch AI the other way round: the capability goes on the billboard and the limits go in a settings menu. Japan’s omotenashi tradition asks the same thing of a host in another form. Anticipate what the guest needs, and never let the guest see you doing it.
The same tradition has the opposite edge written down too. At the Ritz-Carlton, any employee, from housekeeping to the front desk, can spend up to 2,000 dollars per guest, per incident, to put a problem right on the spot, with no manager to ask and no form to fill in. Almost no fix costs anywhere near that, and the number is rarely reached. It is not a budget. It is a statement about who the staff answer to, made before the guest has a problem. The concierge writes down the limit. The Ritz writes down the mandate. Either way, you know before you ask.
The strongest model of working trust we have is not software at all. A guide dog holds a person’s safety at every curb, and the surprising thing about the training is that obedience is not the goal. Guide dogs are taught intelligent disobedience: when the handler gives a command the dog can see is dangerous, step into traffic, board a bus across a gap too wide to clear, the dog refuses and holds its ground. Trainers will tell you that a dog which only obeys is a poor guide. The refusal is not a flaw the schools tolerate. It is taught, on purpose, before the dog is ever matched with a person, which is the concierge’s pledge in another form: the limit comes first. And research on first-time handlers found the emotional bond forms largely on its own track, only loosely tied to how well the animal performs the work. The dog has bad days. It is loved anyway, because its allegiance was never in question. Nothing shipping in travel has earned that kind of trust yet, and the reason is not capability.
Great assistance has always been a relationship with stated limits. We have been shipping the opposite: unlimited claims, undisclosed loyalties.
So here is the standard the next decade of travel should be built against, stated plainly enough that we can be held to it. With you, not used by you. An assistant whose allegiance is clear. It tells you what it will never do before it shows you what it can. It speaks rarely and finishes what it starts. When you ask it for a place, it gives you one, not a menu to grade. It never guesses when money or a reservation is on the line. It knows when the right answer is a person. And its measure of success is not minutes of your attention captured but the opposite: the moment you put the phone down, because everything is handled, and the canyon fills the train window.
There is a harder question under that standard, and it deserves a straight answer: who pays for the loyal assistant? The old business model answered with your attention. That is the dual agent’s answer, and it is how we got here. But travel is unusual in that the operator’s economics already run through trust. A traveler who believes the rebooking will actually happen buys the flexible fare. A guest who trusts the app checks in through it, and comes back. A rider who knows the refund arrives unasked rides again in the rain. Loyalty to the traveler is not a cost the industry absorbs to feel good about itself. It is the rare interface where the traveler’s interest and the operator’s interest point the same direction. The attention economy monetized the conflict. Assistance, done right, monetizes the alignment.
The attention economy monetized the conflict. Assistance, done right, monetizes the alignment.
Some operators have already noticed. Brightline, the private railway in Florida, calls its way of working the hospitality mindset, and every new hire, in every job, spends two weeks on that culture before learning anything else. If a late flight makes you miss your train, its published promise is to put you on the next one at no extra cost, without making you argue for it. That is the front-desk reflex, fix it before the guest has to ask, carried onto a train by people who decided it was the first thing worth teaching.
Software learned to have a face. Then it learned to hold a place. Now it is learning judgment, and judgment always serves someone (UNDERTOW 021: Who’s Standing at the Pass). The question of the next decade of travel is not what these assistants will be able to do; the capability is coming regardless. The question is who they answer to. Travelers have already told us how they will grade the answer. They told us with a power cord.