Why Scaling AI Agents Is Harder Than Building Them

This sponsored content was created in collaboration with a Skift partner. AI agents can look remarkably capable in a controlled demo. Putting that same technology in front of real customers, connected to live systems and exposed to unpredictable demand, is another matter. Skift Research found that 33% of travel operators are experimenting with agentic AI, while just 2% of travel organizations have scaled it. That gap raises a practical question: What changes when an agent moves from a controlled experiment into production?“A demo is designed to show what an agent can do,” said Erik Zahnlecker, travel industry lead at Sierra, a conversational AI platform that helps companies build and deploy AI agents. Sierra works with hotel groups, rental car companies, travel marketplaces, and airlines, including Philippine Airlines. In a demo, the agent typically runs in a controlled environment, without the live systems that affect latency and performance. “Scaling to real production traffic is where you’re testing the agent against real-world environments.” Moving into production introduces a much wider set of technical requirements. Enterprise AI agents have to connect with existing systems, respond fast enough to sustain a conversation, absorb demand spikes, and follow brand rules at scale. For chief technology and information officers, that turns AI deployment into a broader technology strategy question: What should the company own, and where does a specialized partner make more sense? The first article in this series explored how AI agents could maintain context across the travel customer lifecycle, connecting service, loyalty, and booking over time. This article looks at the infrastructure required to make that work at scale. The Path From Demo to Production Demos are an important first step in evaluating what an AI agent can do, typically in a controlled environment using mocked data. A proof of concept goes further by testing whether the agent can operate successfully in a limited environment before a company commits to broader deployment. Production then means connecting to real APIs, handling live traffic, and serving customers across what are often millions of conversations a year. Reaching that stage requires more than technical readiness. Companies also need internal buy-in and a business case for moving forward. Sierra uses proofs of concept to test the technology at a 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. Deployments then scale in stages. A deployment may begin by handling a narrow set of customer intents and 5% of traffic on a particular channel, then expand incrementally as confidence grows. On the phone, that path can eventually place the agent at the top of the interactive voice response (IVR) menu. The Model Is Only the Tip of the Iceberg Sierra uses what it calls the “agent iceberg” to illustrate how much of the work required for a production-ready agent sits beyond the model itself. Above the waterline are the parts most teams picture: a large language model (LLM), retrieval-augmented generation (RAG), which pulls information from external sources, and tools that allow the agent to interact with other systems. “We hear this a lot: ‘It’s just an LLM call with some tools and RAG,’” Zahnlecker said. “The reality is companies tend to underestimate what’s under the hood.” Below the surface are integrations with core systems, the runtime governing how the agent behaves, context management, testing, and optimization. The system has to guard against hallucinations, enforce brand rules, reliably call the right tools, and determine when a task needs a fast model versus one capable of deeper reasoning. None of that AI agent infrastructure stands still. Zahnlecker said Sierra ships an average of 17 meaningful platform updates a week, and he recalled a recent week when roughly eight new foundational models were released, illustrating the pace an in-house team would have to keep up with just to stay current with the frontier. The question, then, is less whether a large travel company can build an agent than whether maintaining that surrounding stack is where it wants to focus its technical resources. Why Travel Raises the Bar for Production Three realities are especially pronounced in travel. Each can strain traditional service operations, while also creating a strong use case for AI agents built to handle production conditions. The first is legacy technology: Some core airline systems have roots in decades-old mainframe infrastructure. AI agents can provide a modern, conversational layer for customers while integrating with those existing systems, but those connections still have to meet conversational latency standards. The second is volatility. For example, a nor’easter can ground hundreds of flights at once, creating a sudden surge in customer demand. Traditional support teams can struggle to absorb peak volumes, while AI agents built for production can provide additional capacity during demand spikes. “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. AI agents can help absorb that spike because they’re built to scale,” Zahnlecker said. Those moments also test how the agent responds to customers under stress. Zahnlecker described the goal as handling disruption “with EQ, not just IQ,” combining technical capability with the empathy customers expect when plans go wrong. The third challenge is language. An international airline may need to serve customers across 20 or more languages and locales, which can be difficult to staff around the clock. Philippine Airlines’ agent, built on Sierra, can switch between Tagalog, Taglish, and English in a single conversation. For global travel brands, that ability can make it easier to serve customers across languages and locales at scale. Where Travel’s AI Talent Matters Most Scaling these systems also depends on the people behind them. Data highlighted in Skift Research’s State of Travel 2026 report shows that travel has one of the lowest shares of AI-trained employees among major industries, with just 3% of the industry’s workforce considered AI-trained. Zahnlecker argues that the relevant question is what kind of AI expertise travel companies need to put the technology to work. Sierra regularly works with in-house data science teams, AI divisions, and product leaders at travel companies, he said. Those teams may not include specialists who pre-train foundation models, but they can still have the expertise needed to apply AI effectively. For technology leaders, that shifts the focus to where their best technical talent can create the most value, such as leveraging customer data, aligning AI with business needs, and improving the customer experience. Rethinking the Build-Versus-Partner Decision Skift Research’s State of Travel 2026 report identifies a key AI decision for travel operators: what to build internally and what to source from outside partners. Sierra’s view is that the answer lies in a division of labor. “We like to say that you should rent the intelligence but own the relationship with your customers,” Zahnlecker said. Travel companies, he added, should retain control of what differentiates them, including their data, customer relationships, operations, brand, and customer experience. Customer context can become part of that differentiation. An agent that remembers a previous conversation can pick up where it left off, while knowledge of a traveler’s preferences can help shape a more relevant offer. The open question is whether maintaining the infrastructure underneath creates a comparable advantage. Foundation-model fees may be one of the smaller costs of an internal build, Zahnlecker said. The higher costs can be people and time, particularly when a project takes months to reach production. Sierra occasionally sees companies spend six to 12 months pursuing an internal build before returning to the market for a partner, he said. What Technology Leaders Should Ask a Partner to Prove For companies that choose to partner, Zahnlecker said the evaluation should focus on whether the technology works in production, whether the company can operate it independently, and whether the commercial model aligns incentives. A partner should be able to demonstrate that its technology works with the company’s own systems, customers, and traffic patterns. “Prove that it works in my world, with my APIs, with my customers, with my spikiness,” Zahnlecker said. The travel company’s own team should also be able to use the platform to build and scale agents without relying on the vendor for every change. And Zahnlecker argues that the commercial model should align the provider’s incentives with the customer’s results. Sierra uses outcome-based pricing, meaning it is compensated when an agent delivers a defined result rather than based primarily on technology consumption. Together, those criteria help technology leaders separate the capabilities that create competitive advantage from the infrastructure that simply needs to work reliably at scale. Learn how Sierra helps travel and hospitality companies deploy enterprise-ready AI agents. This sponsored content was created collaboratively by Sierra and Skift Studio.