Scaling AI Agents in Travel: From Demos to Production

Production is where the hard work starts

AI agents can impress in a demo. The bigger test is whether they can serve customers through live systems, sudden demand spikes, and the messy conditions that travel companies actually face. That gap helps explain why so many operators are experimenting while so few have made the jump to scale.

Skift Research found that 33% of travel operators are experimenting with agentic AI, but only 2% have scaled it. Those figures make the central issue plain: building a convincing prototype is a different job from running an agent in production. Erik Zahnlecker, Sierra’s travel industry lead, says demos typically use a controlled environment and leave out the live systems that shape latency and performance. Sierra is a conversational AI platform whose customers include hotel groups, rental car companies, travel marketplaces, and airlines such as Philippine Airlines. “A demo is designed to show what an agent can do,” he said. Production traffic tests whether it can do that work amid real-world conditions. The distinction matters for chief technology and information officers. A production agent must connect to existing systems, respond quickly enough to sustain conversation, handle surges, and follow brand rules. That makes deployment a question of operating strategy as much as software: which capabilities should a travel company build and maintain itself, and where does a specialist make more sense? A previous article in the series examined how AI agents could preserve context across the travel customer lifecycle, linking service, loyalty, and booking. The question here is what it takes to support that experience at scale.

From proof of concept to live traffic

A demo usually relies on mocked data in a controlled setting. A proof of concept tests whether the agent can work in a limited environment; production means connecting to real APIs and serving customers at scale, potentially across millions of conversations each year. That transition also needs internal support and a business case. Sierra uses proofs of concept to test its technology at limited production scale and demonstrate return on investment. “We’ve found that there’s no better way than to build together and prove it,” Zahnlecker said. Rollouts can expand gradually. An agent might start with a small set of customer requests and handle 5% of traffic on one channel. As the company gains confidence, its role can grow. For phone support, that could eventually mean placing the agent at the top of the interactive voice response (IVR) menu.

The model is only one layer

Sierra’s “agent iceberg” describes how much of a production system sits beneath the visible AI. The parts companies tend to recognize are the large language model (LLM), retrieval-augmented generation (RAG), which fetches information from external sources, and tools that let the agent interact with other systems. “‘It’s just an LLM call with some tools and RAG,’” Zahnlecker said, describing a common assumption. “The reality is companies tend to underestimate what’s under the hood.” That hidden work includes connecting core systems, managing context, setting runtime behavior, testing, and optimizing performance. The agent must avoid hallucinations, honor brand rules, select the right tools, and know when a quick model is sufficient or a more capable one needs to reason through a harder task. The infrastructure keeps changing, too. Zahnlecker said Sierra releases an average of 17 meaningful platform updates each week. He recalled one recent week in which roughly eight foundational models appeared. For an in-house team, keeping pace with that activity would be part of the ongoing cost of maintaining an agent stack. The strategic question is whether that work deserves the company’s technical attention.

Travel puts agents under pressure

Three features of the travel business make production especially demanding: older core technology, sharp swings in customer demand, and the need to serve many languages and locales. Some airline systems still rely on mainframe technology built decades ago. An agent can give customers a conversational way into those systems, but the connection must still be fast enough to support a natural exchange. Demand can change just as abruptly. A nor’easter grounding hundreds of flights may send thousands of customers looking for help at once. “When 10,000 calls come in at the same moment because hundreds of flights are grounded in the Northeast, traditional support teams can quickly become overwhelmed,” Zahnlecker said. Production agents, he argued, can absorb some of that surge because they’re built to scale. Capacity alone won’t satisfy a traveler whose plans have fallen apart. Zahnlecker described the goal as responding to disruption “with EQ, not just IQ.” That’s a tougher bar than answering routine questions: the system needs to combine operational competence with the empathy customers expect in a stressful moment. Language adds another constraint. A global airline may need to support customers in 20 or more languages and locales, around the clock. The Philippine Airlines agent built on Sierra can move between Tagalog, Taglish, and English within one conversation. For international brands, that kind of switching can help extend service across language communities.

Put internal talent where it can make a difference

Travel has a staffing gap to address. Skift Research’s State of Travel 2026 report says just 3% of the industry’s workforce is AI-trained, one of the lowest shares among major industries. Zahnlecker’s argument is that companies should define the expertise they actually need. Sierra works with travel companies’ data science groups, AI divisions, and product leaders. Those teams may not train foundation models from scratch, but they can still apply AI to business problems. For technology leaders, the work is to direct skilled people toward the company’s customer data, operating needs, and service experience. That’s where a travel business may have an advantage that a technology vendor can’t supply on its own.

Decide what to own

Skift Research’s State of Travel 2026 report frames the build-versus-partner choice as a central decision for travel operators. Sierra’s proposed division of labor is simple: “You should rent the intelligence but own the relationship with your customers,” Zahnlecker said. That means retaining control of the assets that distinguish the business: its data, customer relationships, operations, brand, and customer experience. An agent that remembers an earlier conversation or knows a traveler’s preferences could make future help and offers more relevant. Whether the infrastructure behind those capabilities creates a similar advantage is less clear. Model fees may not be the biggest expense in an internal build, Zahnlecker said. Staffing and time can cost more, especially when a project takes months to reach production. Sierra sometimes sees companies spend six to 12 months building internally before returning to the market to find a partner. That’s a meaningful cost of delay, even though the fragment offers no figures for the projects’ total expense or outcomes.

Make a partner prove it

If a travel company decides to work with a vendor, Zahnlecker says it should test three things: production performance, the company’s ability to operate the platform, and whether the provider’s pricing rewards the results the customer wants. A vendor should show that its technology works with the buyer’s APIs, customers, and traffic patterns. “Prove that it works in my world, with my APIs, with my customers, with my spikiness,” he said. The company’s own staff should be able to build and expand agents without asking the vendor to handle every change. The commercial terms deserve equal scrutiny. Sierra uses outcome-based pricing, which ties payment to a defined result rather than primarily to technology consumption. The useful dividing line is straightforward: keep the capabilities that give the travel company an edge, and expect the surrounding systems to prove they can work reliably at scale. Learn how Sierra helps travel and hospitality companiesdeploy enterprise-ready AI agents.This sponsored content was created collaboratively bySierraandSkift Studio.