S
Shahmeer Khan
Guest
Recently, I came across a LinkedIn post that made me wonder whether AI agents help businesses as much as they help users. Here's what I came across: Kiruba Shankar wanted a flight. He gave Claude Code the tedious part, and according to his account, it analyzed 881,076 fare options. One traveler. One trip. A computer that had evidently never heard of “good enough.”
Six flight options are usually enough to send me back to the baggage rules to figure out which "cheap" ticket is actually cheap. Shankar had code sift through hundreds of thousands of possibilities and hand him a usable list. The 881,076 figure counts fare options; the number of airline requests remains unknown. I kept picturing the view from the other side of the search. If one traveler's software explores route after route, date after date, what does that activity look like to a company using searches to estimate demand?
At 8 a.m., picture a hotel revenue manager with rooms to sell, a competitor discounting across the street, and a weekend creeping closer. Rates need setting before guests arrive. Bookings help. So do cancellations, searches, and rejected prices. Lighthouse describes using hotel and flight searches to spot demand ahead of reservations. At Outrigger Hospitality Group, Jenna Villalobos has described using data on shoppers who couldn't get their preferred stay or walked away from a price to inform rate decisions. These clues arrive while customers are still making up their minds.
Let’s say I plan a weekend in Istanbul. One agent suggested a hotel near the tram, which seems fine, but there’s a problem. I notice the booking is non-refundable, which sends me back to the search results. Meanwhile, the agent I forgot to shut off has located a room that is $12 cheaper. And it is actually better than the previous one. Forty minutes outside town. So no hustle and bustle of city life. Brilliant. I try to remember why I've got five of these things running in the first place. Wasn't one supposed to check another weekend? I haven't even bought a plane ticket, and hotel sites are already handling inquiries for dates and rooms I may never book.
Suppose a demand model sees those repeated checks and assigns too much weight to them. A manager might decide that the weekend is filling up and raise rates. Or interpret all those unaccepted offers as resistance to the current price. Either error is possible. In both cases, the manager gets a distorted picture of the customer.
Five tireless assistants can create five trails of commercial activity while pursuing one purchase. Treat those trails as five independent buying intentions, and the forecast has gone off the rails. The customer who prompted all of this could be sitting at home eating cereal, blissfully unaware that a piece of software is trying to interpret their collective enthusiasm.
The funny part? Hotels were dealing with an earlier version of this problem when our biggest shopping innovation was keeping eighteen tabs open, then forgetting which one had the decent rate.
IDeaS once described the traveler who books a room, then keeps returning to see whether a cheaper rate has appeared. Each return can look like fresh interest if a system counts it carelessly. The person eventually goes to sleep or gets called into a meeting. An agent can run the same errand all night, never tiring of checking whether breakfast has become complimentary.
Airlines have been wrestling with this for years. I might spend an hour wondering whether flying home on Tuesday would save me twenty dollars. An agent can check Tuesday, Wednesday, the following weekend, and every connecting flight while I'm making tea. Airline people call the resulting headache look-to-book: an awful lot of fare shopping for every ticket sold. IATA has been examining better ways to measure it, while OAG has questioned what all this machine-generated searching tells airlines. Picture the morning numbers. Searches have shot up, ticket sales have barely moved, and somebody has to figure out whether thousands of travelers are planning holidays or a few hundred agents have been particularly busy overnight.
Extra shopping can also produce extra sales. A traveler who gave up after checking three expensive hotels might finally discover a fourth they can afford. Adobe found that visits referred from generative AI tools during the 2025 holiday period converted 31% more often than visits from other sources. That finding raises questions about shoppers clicking through from AI tools. It says something useful about assisted discovery, while leaving the autonomous-agent question open. More efficient shopping may help customers buy things they would otherwise miss.
Back at the hotel, bot spikes aren’t new for the analyst. She works with the website's traffic figures, where Google Analytics already filters known bots. Even then, the remaining searches still need an explanation. My agent checks refundable Friday rates on the hotel's website while another explores Sunday rates on a booking platform. On paper, they seem like different searches, but each is perfectly reasonable. Both the hotel and the booking platform that answer those requests have no clue that I'm having lunch and still arguing with myself about taking Monday off.
Link every request to a person, then? I could reserve three rooms for my family. That's three intended purchases by one person. My partner and I could each send an agent to find the same room. Two people, one intended reservation. Forcing everyone to solve more traffic-light CAPTCHAs would mainly improve our ability to recognize traffic lights.
Now let’s picture what the hotel manager would face on a Thursday. He is worried the weekend is approaching, and 12 of his rooms still aren't booked for Saturday. On top of that, the hotel across the street just put up a big discount. The manager’s report says searches are booming, and the owner wants to know whether they should discount too. What would the manager do? If I were in his shoes, I'd pull up the reservations first. If bookings are crawling along, that busy search chart suddenly feels like a poor excuse to hold prices.
Perhaps there's a wave of last-minute guests on the way. I'd still be nervous gambling twelve unsold rooms on traffic that could include the same handful of travelers sending agents around in circles.
Because the person using five agents is still planning a weekend in Istanbul. The hotel is still trying to sell rooms. The airline is still trying to fill seats. The machinery between the customer and the company has acquired the patience of a thousand interns and the ability to make one indecisive shopper look like a stampede. That's a dangerous optical illusion for anyone being paid to forecast demand.
If five AI agents spend all day hunting for my one hotel room, how will your business know whether it's found five customers or one exceptionally determined cheapskate?
Six flight options are usually enough to send me back to the baggage rules to figure out which "cheap" ticket is actually cheap. Shankar had code sift through hundreds of thousands of possibilities and hand him a usable list. The 881,076 figure counts fare options; the number of airline requests remains unknown. I kept picturing the view from the other side of the search. If one traveler's software explores route after route, date after date, what does that activity look like to a company using searches to estimate demand?
At 8 a.m., picture a hotel revenue manager with rooms to sell, a competitor discounting across the street, and a weekend creeping closer. Rates need setting before guests arrive. Bookings help. So do cancellations, searches, and rejected prices. Lighthouse describes using hotel and flight searches to spot demand ahead of reservations. At Outrigger Hospitality Group, Jenna Villalobos has described using data on shoppers who couldn't get their preferred stay or walked away from a price to inform rate decisions. These clues arrive while customers are still making up their minds.
Let’s say I plan a weekend in Istanbul. One agent suggested a hotel near the tram, which seems fine, but there’s a problem. I notice the booking is non-refundable, which sends me back to the search results. Meanwhile, the agent I forgot to shut off has located a room that is $12 cheaper. And it is actually better than the previous one. Forty minutes outside town. So no hustle and bustle of city life. Brilliant. I try to remember why I've got five of these things running in the first place. Wasn't one supposed to check another weekend? I haven't even bought a plane ticket, and hotel sites are already handling inquiries for dates and rooms I may never book.
Suppose a demand model sees those repeated checks and assigns too much weight to them. A manager might decide that the weekend is filling up and raise rates. Or interpret all those unaccepted offers as resistance to the current price. Either error is possible. In both cases, the manager gets a distorted picture of the customer.
Five tireless assistants can create five trails of commercial activity while pursuing one purchase. Treat those trails as five independent buying intentions, and the forecast has gone off the rails. The customer who prompted all of this could be sitting at home eating cereal, blissfully unaware that a piece of software is trying to interpret their collective enthusiasm.
The funny part? Hotels were dealing with an earlier version of this problem when our biggest shopping innovation was keeping eighteen tabs open, then forgetting which one had the decent rate.
IDeaS once described the traveler who books a room, then keeps returning to see whether a cheaper rate has appeared. Each return can look like fresh interest if a system counts it carelessly. The person eventually goes to sleep or gets called into a meeting. An agent can run the same errand all night, never tiring of checking whether breakfast has become complimentary.
Airlines have been wrestling with this for years. I might spend an hour wondering whether flying home on Tuesday would save me twenty dollars. An agent can check Tuesday, Wednesday, the following weekend, and every connecting flight while I'm making tea. Airline people call the resulting headache look-to-book: an awful lot of fare shopping for every ticket sold. IATA has been examining better ways to measure it, while OAG has questioned what all this machine-generated searching tells airlines. Picture the morning numbers. Searches have shot up, ticket sales have barely moved, and somebody has to figure out whether thousands of travelers are planning holidays or a few hundred agents have been particularly busy overnight.
Extra shopping can also produce extra sales. A traveler who gave up after checking three expensive hotels might finally discover a fourth they can afford. Adobe found that visits referred from generative AI tools during the 2025 holiday period converted 31% more often than visits from other sources. That finding raises questions about shoppers clicking through from AI tools. It says something useful about assisted discovery, while leaving the autonomous-agent question open. More efficient shopping may help customers buy things they would otherwise miss.
Back at the hotel, bot spikes aren’t new for the analyst. She works with the website's traffic figures, where Google Analytics already filters known bots. Even then, the remaining searches still need an explanation. My agent checks refundable Friday rates on the hotel's website while another explores Sunday rates on a booking platform. On paper, they seem like different searches, but each is perfectly reasonable. Both the hotel and the booking platform that answer those requests have no clue that I'm having lunch and still arguing with myself about taking Monday off.
Link every request to a person, then? I could reserve three rooms for my family. That's three intended purchases by one person. My partner and I could each send an agent to find the same room. Two people, one intended reservation. Forcing everyone to solve more traffic-light CAPTCHAs would mainly improve our ability to recognize traffic lights.
Now let’s picture what the hotel manager would face on a Thursday. He is worried the weekend is approaching, and 12 of his rooms still aren't booked for Saturday. On top of that, the hotel across the street just put up a big discount. The manager’s report says searches are booming, and the owner wants to know whether they should discount too. What would the manager do? If I were in his shoes, I'd pull up the reservations first. If bookings are crawling along, that busy search chart suddenly feels like a poor excuse to hold prices.
Perhaps there's a wave of last-minute guests on the way. I'd still be nervous gambling twelve unsold rooms on traffic that could include the same handful of travelers sending agents around in circles.
Because the person using five agents is still planning a weekend in Istanbul. The hotel is still trying to sell rooms. The airline is still trying to fill seats. The machinery between the customer and the company has acquired the patience of a thousand interns and the ability to make one indecisive shopper look like a stampede. That's a dangerous optical illusion for anyone being paid to forecast demand.
If five AI agents spend all day hunting for my one hotel room, how will your business know whether it's found five customers or one exceptionally determined cheapskate?