Two years into the mainstream availability of large language models, travel companies have had enough time to move past initial experiments and assess where LLM-based tools are actually delivering value in production — and where the limitations are real rather than solvable with prompt engineering.
Content Production at Scale
The clearest and most consistent value delivery is in content production: destination descriptions, hotel write-ups, itinerary narratives, and marketing copy. Travel companies that produce large volumes of content have achieved measurable productivity improvements — content that previously required hours of consultant or copywriter time can be drafted in minutes, with human review and refinement rather than creation.
The important nuance is that LLM-generated travel content requires expert review, not just proofreading. Models hallucinate specific facts — hotel amenity details, operating hours, location specifics — with the same fluency as accurate content. An expert reviewer catching these errors is less work than an expert creating the content, but the expert cannot be removed from the workflow.
Customer Communication Drafting
LLMs are producing value in drafting customer communications — responding to complex inquiries, composing amendment confirmation emails, and writing complaint response drafts that a customer service agent then reviews and personalizes. The model handles the structural complexity of the response; the agent focuses on the tone and personalization. This hybrid produces faster, more consistent communications than agent-only composition.
Internal Knowledge Management
An underappreciated application is LLM-based retrieval over internal knowledge bases. Travel companies with large libraries of supplier contracts, product guides, and operational procedures have deployed LLM-based Q&A tools that let agents query this knowledge in natural language rather than navigating document libraries. The accuracy requirements are lower than for customer-facing applications, and the productivity gain for agents who need to quickly access policy or supplier information is meaningful.
Where the Limits Are Real
LLMs cannot reliably handle: booking transactions where accuracy is absolute (dates, routes, passenger names must be exact, and models make errors at a rate that is unacceptable for financial transactions); real-time data-dependent tasks (availability, pricing, and confirmation status require live data connections that current LLM architectures do not handle natively); and compliance-dependent decisions where the consequences of error are regulatory or financial. These are not temporary limitations of current models — they reflect structural characteristics of the LLM architecture that will require hybrid approaches regardless of how capable the models become.
