AI and the New Power Map of Travel
Summary
This ~30-minute ITB Berlin panel, moderated by Lea Jordan, examined how AI is redrawing influence and power across the travel industry. The session opened with a brief executive impulse from Matias Vela, CEO of Muse (a hospitality enterprise platform), who framed a fundamental shift from scale-based discovery to context-driven intelligence. Vela argued that OTA filter-based search — where travelers select 'gym: yes/no' — is being replaced by LLM conversations where travelers describe nuanced preferences (e.
g., 'a gym with real dumbbell variety for free-weights training' and 'a hotel that warmly welcomes my dog, not just accepts dogs'). The key implication: hotels that feed rich, contextually detailed content to the internet will be found by AI agents, while those relying on sparse OTA attribute data will be invisible.
Simultaneously, he argued that internal proprietary data — operational records, guest history, reservation notes — is a 'moat' that should NOT be shared publicly. He previewed a specific Muse agent deployment that reads previous guest bills and reservation notes to auto-generate pre-arrival tasks: removing dairy from the fridge for lactose-intolerant guests, assigning rooms away from the elevator, etc. The panel then deepened these themes.
Dr. Patrick Andrae (co-founder & CEO, HomeToGo) noted that LLM-driven traffic to HomeToGo is 'going up and up and up' but remains far below Google-driven volumes — emphasizing that the shift is real but early. He flagged the speed of AI evolution: 'what we thought was true in December is already completely different now.
' He cited Claude Code as a concrete example: when launched roughly a year ago Anthropic reported ~20-30% of their own code was AI-generated; now they claim 100%. He cautioned against abandoning ideas after a single failed test, stressing that capabilities improve so fast that something that didn't work three months ago may work today. Andrae also pushed back on stock market overreactions to AI news — arguing that software complexity determines actual disruption risk (a simple to-do list tool is genuinely threatened; SAP-level financial systems are not).
Bhanu Chopra (co-founder & CEO, RateGain — India's first publicly listed SaaS company, with 1,400 employees) offered a two-category framework for understanding AI companies: foundational LLMs (the infrastructure/ammunition layer) versus 'wrappers' (companies building on top of LLMs). He predicted LLMs will build some capabilities themselves — as evidenced by Claude's own product expansion — but argued that deep workflow domain knowledge and complex ecosystem integrations will preserve category-leading wrapper companies. His analogy: 'electricity companies don't go and build hotels and airlines; they just power them.
' Chopra shared that RateGain has deployed multiple internal agents: 'Remo,' an HR agent that operates 24/7, never gets sick, and speaks 100 languages — serving their 1,400-person global workforce. They have also launched a customer-facing AI agent for hotel front desks that handles reservations, answers questions, and makes changes in 100+ languages. He predicted the industry's fragmented intermediary stack — GDS, metasearch, OTAs — will be compressed by AI consumer agents.
The panel converged on a diagnosis of the travel industry's core problem as a 'mindset issue': technology to recognize returning guests and personalize stays has existed for years but is not deployed. A telling moment: when Vela asked the audience 'who has actually built an agent themselves?', only a few hands went up in a room full of travel industry professionals.
He challenged this directly — 'if your company is a chatbot, you're not going to survive' — and argued that vibe-coding (using tools like Claude Code) has removed the technical barrier entirely. He described personally building a working chatbot in 4-5 hours on a Sunday, including setting up a server, email connection, and database. For CEO agendas in 2026: Vela recommended deploying at least one AI application immediately ('just push it ahead — if it doesn't work, replace it, because it moves fast'); Chopra recommended building a company-wide culture of curiosity and learning as the strategic priority; Andrae highlighted that legacy travel companies with decades of unique data and exclusive offerings (e.
g., package tour operators) have a hidden moat that should be leveraged rather than ignored in favor of surface-level chatbot deployments.
Key Takeaways
- •The fundamental shift in hotel discovery: OTAs offer binary attribute filters ('gym: yes/no'), while LLMs parse nuanced guest context ('a gym with real dumbbell variety for free-weights training'). Hotels that publish rich contextual content to the public internet will be surfaced by AI agents; those relying on sparse OTA data fields will be invisible. Vela's framing: 'the most meaningful, usable data is the data that will win to get bookings into your property.'
- •A critical data duality for hotels: external-facing data should be maximally rich and publicly accessible for LLM discovery; internal operational data (guest history, reservation notes, billing records) is a proprietary moat and should NOT be shared externally. These two imperatives require different strategies.
- •Muse is actively deploying an agent that reads previous guest bills and all reservation notes to auto-generate pre-arrival task lists — for example, instructing housekeeping to remove dairy and replace with oat milk for lactose-intolerant guests, or allocating rooms away from the elevator. The human remains a 'co-pilot' who confirms actions with a click before execution.
- •LLM-driven traffic to HomeToGo is growing rapidly but remains 'far from' Google-level volumes — meaning the channel is real and must be prepared for, but is not yet a primary revenue source for most travel businesses. Andrae's caution: don't over-index now, but don't ignore the trajectory.
- •Patrick Andrae's AI evolution warning: 'What we thought was true in December is already completely different now.' He used Claude Code as a concrete benchmark — when launched ~1 year ago Anthropic reported 20-30% of their own code was AI-assisted; they now claim 100%. Leaders must continuously re-test AI capabilities every few months rather than writing off failed experiments as permanent limitations.
- •Bhanu Chopra's two-category AI taxonomy: (1) foundational LLMs — the infrastructure layer providing 'ammunition' to build on top of; (2) 'wrappers' — companies building domain-specific applications on LLMs. His prediction: LLMs will expand into some wrapper territory, but deep domain knowledge, workflow complexity, and ecosystem integrations will protect category-leading wrapper companies. Analogy: 'electricity companies don't build hotels — they power them.'
Transcript
[music] Hello, hello, hello. Thank you for still being here. Thank you for coming in and joining us. I love that. And I'm repeating myself, but I know you're all spoiled for choices. So much is going on and you chose us, which we love, of course. not smart decision because the next 40 minutes I'd say there is super interesting insightful I'm totally excited for that and what we're going to talk about obviously is AI artificial intelligent intelligence and it's not longer a side conversation as w...
