A quick recap for anyone joining late. For thirty years travel ran on one bargain: looking was free because booking paid for it. The search got cross-subsidised, buried inside the commission on the one trip in a few hundred that someone actually bought. AI broke that bargain.
Agentic search runs thousands of tireless queries per trip and moves the cost off the rare booking and onto the infinite look, where for now it sits on nobody’s invoice.
That is where part one left our distribution analyst, watching the compute bill climb and refreshing a screen that would not change. Now we find out whose screen it really is.
Who eats it
So, walk the chain, because the pain is not evenly distributed and the challenge is knowing where it lands.
Start with the suppliers, the airlines and hotels. For years airlines fought to escape GDS fees by pushing New Distribution Capability. Two-thirds of airlines have it now. Feels like a win. Except in winning, they took the search compute onto their own servers. Under the old model the GDS ate the cost of generating offers.
Under NDC the airline eats it, on every query, whether or not anyone buys. Air Canada saw this coming back in 2023 and slapped a Distribution Cost Recovery surcharge of $20 to $30 on GDS bookings while exempting its own NDC and direct channels. That was an airline trying to shove distribution cost back down the pipe. But a surcharge on bookings does nothing about a cost that now lives on the looks. You cannot toll a conversion that is one in 200,000.
Hotels face the same squeeze from a different angle: every AI agent rate-shopping their inventory in real time, around the clock, with no intent to sleep there.
Then the intermediaries, the GDS players themselves. Here is an uncomfortable truth. The product the GDS actually sold for 30 years was not booking. It was free search at scale, subsidised and hidden inside the segment fee.
When the look-to-book ratio inverts, a per-segment toll collected on the rare booking simply cannot fund infinite looks. And metasearch, the Kayaks and the Skyscanners whose entire business is monetising the act of looking and handing off the buying, is the most exposed creature in the whole ecosystem. Their product is the look. The look just became the cost centre.
Then the TMCs, the travel management companies your programme actually pays. Their transaction-fee model assumes a human makes a countable number of bookings: $8 online, $35 with an agent, fine, because Karen in accounts payable books maybe 40 trips a quarter.
When the booker is software running thousands of speculative queries, the fee-per-transaction logic wobbles and the value proposition moves. The booking keystroke is worth nothing when a machine does it for a fraction of a cent.
What a TMC can still charge for is orchestration, duty of care, the messy exception when your traveller is stranded in Frankfurt at midnight and no agent wants to touch a booking that an AI made and that an AI broke. That is real value. It is just not the value they have been invoicing for, at least so far.
Then your seat, the corporate travel programme. This is the part that should make you put your coffee down. The metric you have optimised for a decade, cost per transaction, was built on the assumption of human search. AI changes the denominator and the hidden cost base underneath it.
The booking tools are pitching you genuine capability. Navan says its engine analyses more than 35 data points per search and says 80 per cent of travellers book one of its top 10 results, and it captures over 130 data points per expense. That is useful.
But every one of those data points is compute, and right now that compute is bundled into a SaaS subscription where you cannot see it. Bundled costs do not stay bundled. They get unbundled the moment they hurt the vendor's margin, and the search-cost line is going to hurt.
Add the implementation reality, where 46 per cent of operators told PhocusWire the technological complexity of wiring AI into legacy systems is the main barrier, and legacy itinerary engines that ran fine under human pace fall over when AI agents start firing requests at machine speed. The savings are real. They are also not free, and the bill has a delay on it.
And finally the traveller, who gets sold the dream and immediately gets rationed. The pitch is a frictionless agent that shops the whole world and hands you the perfect trip. The reality forming underneath is search caps, throttling and what one vendor cheerfully calls “intelligent caching,” which is a polite phrase for showing you the answer that was cheap to serve rather than the answer that was best. When looking costs money, infinite looking quietly becomes finite looking with better marketing. The traveller will not see the cap. They will just see slightly worse options and never know why.
The objection, and why it does not save you
The smart pushback is that tokens keep getting cheaper, so this is a temporary spike that Moore's Law washes out. I would love that to be true. It is not, and the reason is the one we already covered: Volume is outrunning price, by Gartner's own numbers, straight through 2030.
The other pushback is that agents will eventually book at higher intent, so the look-to-book ratio recovers and the cross-subsidy heals. Maybe. Someday. But the cost lands now and the conversion benefit, if it ever shows, comes later. You do not get to pay next year's improved economics with this year's compute bill.
There is a better objection, and it comes from the engineers, not the optimists. They will say that the look-to-book explosion is not a law of nature, it is a design failure. A well-built agent thinks inside its own head, evaluates 10,000 options for a fraction of a penny, and only touches an airline system to validate the five that survive. Thinking is cheap and invisible. Requests are expensive and loud. Conflate them and you mistake a badly built robot for an inevitability. They are right about the mechanism.
But watch what the engineers’ argument rests on: the industry choosing to redesign. The obstacle is not technical but commercial, because the incumbents who could fix this can also bill you for not fixing it. That is the whole problem. The cost is landing now, on the legacy workflows we actually have, while the elegant architecture of optimisation is still a slide in a deck. You do not get to run your 2026 travel programme on 2030's reference design.
There is plenty written about the look-to-book explosion as a supply-side headache, and plenty written about the shiny AI booking experience. What almost nobody has actually published is the head-to-head unit economics. A clean model that says a human search costs this much in commission terms and an AI search costs that much in tokens, set side by side across all five stakeholders. It does not really exist yet.
Even Skift's own framing leans on a phrase about the industry lacking transparency around who bears the search cost, which is consultant-speak for “nobody built the model.” When an entire industry keeps describing a cost as opaque, that is not modesty. That is the tell. It means the spreadsheet that would answer the question would also indict the business model, so the spreadsheet does not get built.
What actually changes for you
So here is what you do with this, the next time a TMC or a booking-tool vendor is across the table. Stop opening with the transaction fee. The transaction fee is a fossil from the human-search era and arguing it to the dollar is fighting the last war.
Open instead with one question that nobody has a clean answer to yet: What is your search cost model, and who eats the compute when my AI assistant runs 4,000 queries to book one trip? Watch the body language.
The vendor who has thought about it will talk to you about caps, caching, and who absorbs what. The vendor who has not thought about it will say something about innovation and efficiency and the future of travel, which means they are going to put it on your invoice later and call it a platform fee.
The deeper pattern is that every distribution system humans have ever built assumed that human attention was a natural rate limiter. Fees, free search, the whole cross-subsidy, all of it rested on the fact that people eventually stop looking and go to dinner. AI removed the rate limiter. And when you remove a constraint that an entire economy was unknowingly built on top of, the economy does not adapt gracefully. It finds out which costs it was never really paying and starts paying them all at once.
But here is the part the panic gets wrong: The number is not fixed by physics. It falls for the system that redesigns the search, and it stays broken for the one that waits for its GDS to hand the savings back out of kindness.
So the signal, the next time a vendor is across the table, is not whether they will charge you for search. A vendor who wants to charge you per search is signalling they intend to profit from the inefficiency rather than fix it. The vendor worth hiring is the one building the agent that never needed brute-force in the first place. One is selling you a meter. The other is selling you an engine. Everyone in this market is about to claim they sell the engine while quietly installing the meter. Learn to hear the difference. Those who are not aware are still refreshing the dashboard, waiting for the number to change.
It is not going to change.
______________________________________________________________________
Steve Clagg is the founder of Clagghaus Consulting, where he works with corporate travel buyers and procurement leaders on technology strategy, supplier evaluation, and programme architecture. If your organisation is navigating enterprise AI integration and trying to figure out where your travel programme fits, that is exactly the kind of work Clagghaus does.