Kiruba Shankar needed an Etihad ticket. So he pointed Claude Code at the airline’s website and asked it to find the best option across a month of dates, four routes, and every stopover combination. The agent returned 881,076 fare combinations after evaluating every permutation it could construct. Shankar posted the result on LinkedIn, and within days the story had travelled from a single founder’s curiosity to a feature in Skift, then into the kind of conversation that airline commercial teams have been quietly having for months.
The metric nobody outside distribution has heard of
The number is the headline, but the real story is what it does to a metric most travelers have never heard of: look-to-book ratio, the relationship between how many times a fare gets searched and how often that search turns into a sale. Airlines have always expected many searches to result in few bookings, but distribution pricing models were built around human behaviour rather than autonomous software capable of testing hundreds of thousands of possibilities. Under the old model, a human shopping for a flight runs ten or twenty searches before booking, and the eventual sale covers the cost of all that looking. An AI agent does not stop at twenty. It does not get tired, and it does not need a reason to keep checking one more date or one more stopover. When the searching scales into the hundreds of thousands and the booking stays at one, the unit economics that have quietly funded airline distribution for decades start to come apart.
Airlines selling through NDC are exposed in a particular way. The whole pitch behind NDC was richer, more dynamic content: live availability, bundled fares, personalised offers, all retrieved in real time rather than pulled from a static cache. That richness is exactly what makes each search expensive. A legacy GDS query is comparatively cheap to serve. An NDC query that checks live rates, fare rules, and ancillary bundles before returning an offer is not, and an AI agent that runs that query 800,000 times in an afternoon turns a feature into a liability.
The industry already saw this coming
What makes this more than a viral anecdote is that at least one major NDC aggregator was already building a defence before the story broke. In March, TPConnects added a Model Context Protocol layer to its Astra NDC Gateway, built specifically to manage what the company called search volumes moving toward “million-to-one look-to-book ratios” as AI adoption accelerates. The company says the accompanying tool, Astra ConvertEngine, is already cutting search volume by 60 percent in production while keeping accuracy at 97 percent. Whether or not that holds up at scale industry-wide, the timing matters. The fix arrived before the public conversation did, which says the people closest to NDC infrastructure saw this coming well ahead of a LinkedIn post making it visible to everyone else.
Why Africa and the Gulf carry more exposure
For carriers and aggregators operating in Africa and the Gulf, the exposure looks different from a US major’s. A large network carrier can often absorb a surge in shopping traffic through scale and infrastructure investment. A smaller airline relying on outsourced distribution infrastructure has far less room for error. The cost of search was already a quiet tax on smaller carriers before AI agents arrived. An agent that runs the same query loop intended for one traveler’s itinerary, multiplied across an unknown and unbilled volume, is a tax with no agreed-upon ceiling.
Who pays is still an open question
That is the part of this story still unresolved everywhere, not just in emerging markets. Nobody in the value chain has settled who pays when an AI agent, not a person, generates the search. The airline absorbs the infrastructure cost of serving the query. The aggregator or GDS sits in the middle, trying to filter or cap the volume before it reaches the airline. The traveler, who outsourced the job to an AI assistant in the first place, has no idea any of this is happening and would be surprised to learn their one ticket search showed up somewhere as 881,076 line items.
Search caps and smarter filtering are the first response, and they will not be the last. The harder conversation, the one that has not happened publicly yet, is whether the commercial terms between airlines and aggregators need to be rewritten for a world where the buyer on the other end of an API call might not be a person at all. The first decade of airline distribution transformation was about who owned the booking. The next may be about who pays for the search.
