This insight summarises Why Scaling AI Agents Is Harder Than Building Them, sponsored content created by Skift Studio with Sierra, a conversational AI platform, published on Skift on 4 November 2026. Sierra's claims are vendor-sourced, and its examples lean towards airlines rather than hotels.
Demo versus production. Skift Research found that 33% of travel operators are experimenting with agentic AI, but only 2% have scaled it. Erik Zahnlecker, travel industry lead at Sierra, says demos run in controlled environments, while production means real APIs, live traffic and often millions of conversations a year. A typical rollout starts with a narrow set of intents and about 5% of traffic on one channel.
The model is the tip of the iceberg. Beneath the LLM, retrieval and tool integrations sit core-system integrations, runtime behaviour rules, context management, testing and optimisation. These guard against hallucinations and enforce brand rules. Zahnlecker says companies "tend to underestimate what's under the hood." Sierra says it ships about 17 meaningful platform updates a week.
Why travel raises the bar. Agents must connect to decades-old systems at conversational speed, absorb disruption surges and handle stressed customers "with EQ, not just IQ." Language adds complexity: Philippine Airlines' Sierra-built agent switches among Tagalog, Taglish and English within one conversation. Talent is short too, with Skift Research putting the share of AI-trained employees in travel at 3%.
Build or partner. Zahnlecker's principle is to "rent the intelligence but own the relationship with your customers." Companies should keep what differentiates them: data, customer relationships, operations, brand and experience. Model fees are a small part of an internal build; people and time cost more. Sierra says it sometimes sees companies spend 6 to 12 months building before seeking a partner. It advises requiring proof in your own environment, operational independence from the vendor and aligned incentives, and notes that it uses outcome-based pricing.
AIHA practical takeaway. Few hotels will build agents from scratch, but the same vendor questions apply. Ask for a pilot against your live PMS and booking engine rather than a demo, limited to a narrow set of intents. Get in writing who owns conversation data and guest memory, whether your team can change agent behaviour without a vendor ticket and how any outcome-based charge is defined. Read the build-versus-partner conclusion as a vendor's view, since a partner benefits from it.
Related AIHA insight: When the Hotel's AI Acts, Who Is Accountable?