ATPCO has developed an approach to improve how artificial intelligence (AI)-powered flight shopping connects traveler intent to airline products.
The company ran a two-week proof of concept to explore how a request expressed in everyday language could be matched with the structured airline data needed to identify relevant products.
AI assistants are already good at understanding what a traveler is asking for, Anand Mishra, VP of technology at ATPCO, told PhocusWire. The challenge is linking that traveler intent to structured airline data.
If a traveler asks AI for a comfortable seat, for example, the company is then thinking about how it links the “semantic meaning of comfortable seat" with the seat map in its data, Mishra said.
ATPCO also tested how it might return relevant content if a traveler provided the AI with an image of what they wanted instead of just text, drawing from sources including Routehappy, which provides visuals and product attribute detail, he said.
“The experiment became, how do we get intent, match it with our data so that the human travelers can be better served, and, secondly, how do I get that intent to the airline so that the airline that can serve the customer’s intent best can present an offer?”
Translating intent into product matches
ATPCO approached the problem in two stages, the company explained in a blog post.
For the first step, ATPCO used Gemini to create vector embeddings from both the traveler’s request and a subset of its own data. That allowed the system to identify ATPCO data that was semantically closest to the request, rather than relying on exact wording.
Mishra likened this process to placing pins on a map.
“The description in our product catalog of what those cabins mean, what those seat dimensions mean, all that text, all those descriptions, the image and video content, all that gets mapped to a vector space,” he said. “Then the request comes in from Gemini. It also places a pin on the map and says, ‘What are all the pins closest to me?’ And then we surfaced those records back to the AI assistant.”
Next, ATPCO connected records across its fare data, product catalog and Routehappy that referred to the same product.
“The hard part, from a technology engineering perspective, was actually stitching disparate systems together into what we call the canonical record,” he explained.
The solution, drawing from auto manufacturing for inspiration, was to combine carrier, brand and aircraft type into a single key associated with a single product, the company explained in its blog post.
“The human intent doesn’t know that this is how we’re structuring the data,” Mishra said. “They just have a semantic goal.”
The prototype helped to bridge the gap between the traveler’s goal and that structured data.
Testing the limits
ATPCO tested the approach across 12 scenarios, including some in which no suitable data match existed, Mishra said.
The company measured whether the system returned the expected canonical record and how close the vector search came to the intended result.
In some cases, the system still returned a record even though none should have matched.
“Even if that other point is really, really far away, it’s going to pick it,” Mishra said.
The test also raised questions about how the system should handle low-confidence results. Mishra said ATPCO is considering whether exposing confidence levels could help indicate how closely a result matches the traveler’s intent, adding that the underlying issue is one the company expects to be able to address.
What comes next
Moving beyond proof of concept will require ATPCO to scale the approach.
“We took a very small set of data. We created the canonical record; that is actually not easy to do,” he said. “And as the numbers of data and data sets increase, that becomes more and more complicated.”
ATPCO also needs to make the rules in its data standards machine-readable. Mishra said those rules are currently documented for human experts.
The next step would be to provide the canonical records and machine-readable information through its distribution systems to airlines and other customers, he said.
“If they have it, and then the LLMs can find it, then you’re getting into this really nice state,” he said. “The LLM will call the airline if it makes the LLM’s job easier.”
The company could also publish its knowledge graph so LLMs could identify airlines whose data meets ATPCO standards, he said.
Longer term, Mishra said the company is exploring what he described as an “airline context protocol” that could define a standard handshake between large language models, airlines and online travel agencies.
For Mishra, the proof of concept showed the potential for ATPCO’s canonical records and semantic data graph to increase trust and provide airlines with more specific information from AI shopping requests.
“That’s what it unlocked,” he said.