Two travel companies are quietly shipping agentic features. One will call it existential.
By Q4 2024 the travel category had spent eighteen months announcing agentic features. Booking, Expedia, Kayak, Tripadvisor, Hopper, and most of the larger OTAs had all shipped some version of an "AI travel concierge" or "intelligent search" feature, framed in press releases as the agentic shift the category had been promising.
Almost none of them worked in production for paying customers.
Two did. Airbnb shipped a support-side agentic deployment that, by Q4 2024, was handling 33% of North American support without human escalation. Booking shipped a planning-side agentic tool inside its Genius-tier app that was, by the same window, actually changing the booking flow for the cohort that used it. The other companies' agentic features were demos. They generated press cycles. They did not generate revenue.
The structural reasons the rest didn't are worth naming.
One: the data graph. An agentic feature in travel needs a data graph the agent can reason against. The agent that recommends the right hotel needs to know the customer's prior bookings, their preferences inferred from review behavior, their willingness-to-pay calibrated against past transactions, the inventory state at thousands of properties on dates relevant to their query. Airbnb has 500M reviews and 1M support-interaction transcripts to post-train on. Booking has the OTA's hotel-and-customer interaction graph at scale. The smaller OTAs and the legacy travel-sites have neither at production scale, and the agentic feature lands as a shallow wrapper over the model's training data rather than a feature that knows the customer.
Two: the integration surface. An agentic feature has to call into the booking flow, the payment system, the loyalty layer, the customer-service backend. The two companies that shipped working agentic features had spent five years building integration architecture that an agent could reason against. The companies that didn't were running on integrations built for human users, and the agent's calls into those integrations failed at rates the demo did not surface.
Three: the pricing model that pays for it. Agentic features are expensive to run. The model calls cost real money, the verification loops add tokens, the feedback infrastructure needs engineering investment. Airbnb's pricing model (host fee plus guest fee on every booking) absorbs the cost as a margin expansion. Booking's pricing model (commission on each booking) does the same. The OTA running on a thinner take-rate or on subscription cannot absorb the agentic feature's cost without margin compression that shareholders don't tolerate.
The two companies that shipped agentic features in 2024 are the two companies whose business model could absorb the cost while their data graph could feed the model. Both conditions had to be met. Neither was sufficient alone.
What happens next is the operator-class question.
One of the two will call this an existential category-shift first. _Airbnb already has._ Brian Chesky's late-2024 strategic reframing named AI as the company's fourth pillar and explicitly framed the agentic shift as a category-defining move. Booking's executive class is talking about agentic features in earnings calls but has not made the framing existential. Airbnb is, in operating terms, betting the brand on the category-shift; Booking is, in operating terms, hedging.
Both bets have merit. Airbnb's existential framing aligns the company's strategic capital and its hiring signal and its product roadmap behind the same direction. Booking's hedging preserves optionality. The right answer depends on whether the agentic shift is, as Airbnb is betting, a multi-year category-replatforming, or whether it is, as Booking is hedging, a five-year feature-cycle that compounds without replatforming the category.
The read that survives is that Airbnb is right about the direction and possibly early on the timeline. Booking is right about the timeline and probably miscalibrated on the conviction needed. The two companies will look very different in 2027-2028 depending on which framing was correct, and the gap between their stock prices will price the answer.
The thing that crosses pillars is sharper than the travel-press is reading.
Every category that has been announcing agentic features in 2024 is in the same shape. Healthcare, finance, retail, B2B SaaS — each has a small number of companies that shipped working agentic deployments and a long tail that shipped demos. The structural conditions (data graph at scale, integration architecture an agent can call, pricing model that absorbs the cost) recur across categories. The companies in each category that meet the three conditions ship working features. The companies that don't ship demos that get covered in the trade press and don't generate revenue.
The structural read for any company watching this play out is to ask which of the three structural conditions the company is missing, and to invest there before announcing the agentic feature. Most companies are doing the inverse: announcing the feature first, building the data graph and integration surface and pricing model retroactively, and discovering in production that the demo doesn't translate.
Airbnb and Booking shipped because they had built the conditions before they announced the feature. The rest of the category will catch up over the next 24-36 months, on a curve that compounds against companies that get the order right and against companies that don't.
One company will call this existential. One company will call this an iteration. The 2027 financials will price which one was correct, and the operator-tier observers who paid attention to the structural-conditions question in 2024 will know the answer before the financials confirm it.
—TJ