Delta's AI pricing drew a Senate letter, Expedia's Romie can plan a trip but not book one, and hotel revenue management has quietly run on forecasting models for over a decade — a look at what's actually automated in travel versus what's still consumer-facing theater.
In July 2025, three U.S. senators sent Delta Air Lines a letter accusing it of planning to charge every passenger a different fare based on what its AI model could infer about them — "predatory pricing," in Senator Ruben Gallego's words. Delta's response, published on its newsroom site, was blunt: there is no fare product that targets individual customers based on personal data, and the AI system (built with an Israeli startup called Fetcherr) only adjusts pricing using aggregated market signals — demand, competitor fares, booking pace — not browsing history or loyalty status used as a billing lever. By the company's own account, AI was already setting prices for about 3% of domestic fares in mid-2025, with a plan to reach 20% by the end of the year.
Whichever side of that argument you land on, the episode is a good entry point into what AI is actually doing in travel and hospitality right now, because it surfaces the real tension in the industry: pricing algorithms have used aggregate data for decades (that's classic revenue management), and what's changed is the speed, scale, and opacity of the systems making the calls — plus a public that's primed to assume the worst about "personalized AI pricing" whether or not it's actually happening.
Travel is a useful vertical to look at because it has four semi-independent fronts where AI adoption is real but uneven: fare pricing, trip-planning assistants, hotel revenue management, and airport/security processing. Each is at a different point on the maturity curve, and each has its own version of the "is this actually new" question.
Airlines have run algorithmic, demand-based pricing since at least the 1980s with systems like Sabre's fare management tools — that's not new. What Delta's Fetcherr deployment changes is the method: instead of rule-based yield management (if load factor crosses X, raise price band by Y), a machine-learning model ingests "thousands of variables simultaneously" — per Delta's own description — and outputs a price recommendation in something closer to real time, continuously, rather than through scheduled fare-file updates pushed via ATPCO a few times a day.
The distinction regulators and critics are drawing is between:
Delta says it does the first, not the second, and that fares remain visible and identical to any shopper regardless of personal data. Harvard researcher Noah Giansiracusa's point in the aftermath is the more durable one: the technology to do individualized pricing exists and has been tried before (Orbitz once showed Mac users pricier hotel results), and public backlash — not law — is mostly what's stopped it so far. That's a thin reed to hang consumer protection on, which is why the Senate letter exists at all and why this is worth watching rather than treating as settled.
Adoption reality: this is real, in production, at scale — not a pilot deck. The friction is entirely regulatory and reputational, not technical.
Expedia's Romie and Booking.com's AI Trip Planner are the most visible consumer-facing AI products in travel, and both work roughly the same way: a conversational layer (Booking.com's is built on top of ChatGPT; Expedia built Romie as an in-house assistant) takes natural-language trip constraints and returns itineraries, hotel shortlists, and activity suggestions, with some ability to track group chats or trip disruptions.
A side-by-side test by USA Today in 2024 is instructive about where these tools actually stand. Romie produced a full day-by-day itinerary once given interest categories (food, history, art) and proactively suggested nearby restaurants — but it couldn't complete a booking, sent at least one broken (non-clickable) ticket-purchase link, and lost context when the tester switched from the app to SMS. Booking.com's planner returned reasonable hotel shortlists and generic "greatest hits" attraction lists quickly, but needed explicit prompting for anything specific (dietary needs, trip length) rather than inferring it. Neither tool, as of that test, could close the loop from "plan this" to "book this" without handing the user back to a traditional checkout flow — which both companies describe as intentional: Expedia's product lead told USA Today the experience of letting an agent charge your card automatically "isn't quite ready yet."
That gap — planning assistance without transaction authority — is the honest state of agentic travel booking industry-wide in 2026, not just at these two companies. The models are good at synthesis and recommendation; they're deliberately kept away from payment execution, both for liability reasons and because travel inventory (fares, room availability) changes fast enough that an agent acting on stale context is a real failure mode, not a theoretical one.
Adoption reality: live, used by millions, but functioning as a smarter search-and-shortlist layer, not a replacement for the booking funnel.
Hotel pricing is the sleeper category here, because revenue management systems (RMS) from vendors like IDeaS and Duetto have used forecasting models for over a decade — this isn't an LLM-era arrival. What's changed recently is the addition of generative and agentic layers on top of the existing forecasting core: Duetto's "Advance" product now runs dynamic optimization passes every 30 minutes rather than once or twice a day, continuously re-pricing within guardrails a human revenue manager sets, and logs every automated rate change for audit.
The more interesting adoption story is cultural, not technical. A ZS/HSMAI industry study cited in trade coverage found that hotel revenue managers spend roughly half their time on tasks that aren't actually about setting rates — competitive shops, reporting, internal alignment meetings — which is the opposite of the airline world, where pricing has been close to fully automated for years. Hotels, unlike airlines, still mostly have a human in the loop reviewing AI-suggested rates before they go live, and pricing strategy often varies property-by-property within the same brand rather than being centrally automated. The AI's real contribution so far is less "set the optimal price" (airlines already largely solved that problem procedurally) and more "free the revenue manager from manual shops and let them cover more properties."
Adoption reality: mature on the backend (forecasting), still maturing on the autonomy side (humans still gate most rate pushes).
The fourth front is the one most travelers physically touch: facial recognition at security checkpoints. TSA's Credential Authentication Technology units at many major U.S. airports now pair ID verification with a live facial match, and CLEAR — a private biometric verification service used at both airport security lines and stadium/venue entrances — runs a similar model commercially, matching a traveler's face or fingerprint against a pre-enrolled profile to skip the ID-check step entirely.
This is the one application area where the "limitation" isn't really about model capability — facial matching accuracy at this point is high for the specific controlled, well-lit, cooperative-subject setting an airport checkpoint provides — but about consent and false-negative handling: TSA's program is framed as opt-out-able, but travelers report inconsistent staff awareness of that option, and any identity-verification system that's wrong even a fraction of a percent of the time at airport scale produces real daily friction for a non-trivial number of travelers who get flagged for manual review. Unlike the pricing and planning use cases above, failure here isn't "the itinerary was generic" — it's a missed flight, which raises the bar for what "acceptable error rate" means even when the underlying number sounds good in a vendor deck.
Adoption reality: widely deployed at major hub airports and venues, the most physically consequential of the four, and the one facing the most organized privacy pushback (state biometric privacy laws like Illinois' BIPA have already produced litigation against similar systems in other industries).
| Application area | Core technique | Adoption stage | What actually limits it |
|---|---|---|---|
| Airline fare pricing | ML-based dynamic pricing (Fetcherr/Delta) | Production, scaling (3% → 20% of fares) | Regulatory scrutiny, "surveillance pricing" perception |
| AI trip planning | LLM-based conversational assistants (Romie, Booking.com AI Trip Planner) | Live, consumer-facing | No transaction authority; breaks on edge cases, link handling |
| Hotel revenue management | Forecasting + gen-AI layer (IDeaS, Duetto) | Mature core, maturing autonomy | Human sign-off culture, per-property inconsistency |
| Airport biometrics | Facial recognition matching (TSA CAT, CLEAR) | Widely deployed at major hubs | Consent clarity, false-negative cost, state privacy law |
The pattern across all four: the oldest, most constrained problem (airline pricing, hotel RMS) is where AI is most deeply embedded, because those systems have been iterating on structured, numeric optimization for 20+ years and LLMs just made the models better, not new. The newest, most open-ended problem (conversational trip planning) is where AI is most visible to consumers but least autonomous, because natural-language itinerary generation is a much easier problem than reliably executing a non-reversible purchase on someone's behalf.
If you're building or evaluating AI in travel, the useful question isn't "does this company use AI" — virtually all of them do and have for years — it's "what decision is the model actually allowed to make unsupervised." Fare-pricing AI gets real autonomy because a wrong price just sits in the market and gets corrected by demand. Trip-planning AI gets none, because a wrong booking is expensive and hard to undo. Biometric AI gets autonomy over identity verification but under tight regulatory watch, because the cost of being wrong is a missed flight, not a bad recommendation. That boundary — not the presence of AI itself — is what's actually different across these four parts of the same industry.