Build or buy? Scaling Your Enterprise AI Pipeline in 2025

This article is part of VentureBeat’s special issue, “AI at Scale: From Vision to Viability.” Read more from the issue here.

Accept scaling of generative tools it has always been a challenge to balance ambition with practicality and in 2025. the stakes are higher than ever. Enterprises racing to adopt large language models (LLM) face a new reality: Scaling isn’t just about deploying larger models or investing in cutting-edge tools—it’s about integrating AI in ways that transform operations, empower teams and optimize costs. Success depends on more than technology; this requires cultural and operational change that aligns AI capabilities with business goals.

The Scaling Imperative: Why 2025 is different

As generative AI evolves from experimentation to enterprise-scale deployment, businesses are facing an inflection point. The excitement of early adoption has given way to the practical challenges of maintaining efficiency, managing costs and ensuring relevance in competitive markets. The Scaling of AI in 2025 is about answering tough questions: How can businesses make generative tools impactful across departments? What infrastructure will support AI growth without crippling resources? And perhaps most importantly, how are teams adapting to AI-driven workflows?

Success depends on three critical principles: identifying clear, high-value use cases; maintaining technological flexibility; and fostering a workforce prepared to adapt. Enterprises that succeed don’t just embrace next-generation AI—they craft strategies that align the technology with business needs, constantly reassessing the cost, productivity, and cultural changes needed for sustainable impact. This approach isn’t just about implementing cutting-edge tools; it’s about building operational resilience and scalability in an environment where technology and markets are evolving at breakneck speed.

Companies like Wayfair and Expedia embody these lessons, demonstrating how hybrid approaches to LLM adoption can transform operations. By combining external platforms with bespoke solutions, these businesses exemplify the power of balancing flexibility with precision, setting a model for others.

Combining customization with flexibility

The decision to build or buy AI tools is often described as binary, but Wayfair and Expedia illustrate the benefits of a nuanced strategy. Fiona Tan, Wayfair’s CTO, emphasizes the value of balancing flexibility with specificity. Wayfair uses Google Vertex AI for general applications while developing proprietary tools for niche requirements. Tan shared the company’s iterative approach, sharing how smaller, cost-effective models often outperform larger, more expensive options when marking product attributes such as fabric and furniture colors.

Similarly, Expedia uses a cross-vendor LLM proxy layer that enables seamless integration of different models. Rajesh Naidu, senior vice president of Expedia, describes their strategy as a way to remain flexible while optimizing costs. “We’re always opportunistic, we look at best-in-class (models) where it makes sense, but we’re also willing to build for our own domain,” Naidoo explains. This flexibility ensures that the team can adapt to changing business needs without being locked into a single vendor.

Such hybrid approaches recall the evolution of enterprise resource planning (ERP) from the 1990s, when enterprises had to choose between adopting rigid, off-the-shelf solutions and highly customizing systems to suit their workflows. Then, as now, successful companies recognized the value of blending external tools with custom developments to address specific operational challenges.

Operational efficiency for core business functions

Both Wayfair and Expedia demonstrate that the true power of the LLM lies in targeted applications that deliver measurable impact. Wayfair uses generative AI to enrich its product catalog by enhancing metadata with autonomous accuracy. This not only streamlines work processes, but improves customer search and referrals. Tan highlights another transformative application: using LLM to analyze legacy database structures. With the original system designers no longer available, the AI ​​generation allows Wayfair to mitigate technical debt and unlock new efficiencies in legacy systems.

Expedia has been successful in integrating Gen AI into customer service and developer workflows. Naidu shares that a custom AI generation tool designed to summarize conversations ensures that “90% of travelers can reach an agent within 30 seconds,” contributing to a significant improvement in customer satisfaction. In addition, GitHub Copilot is deployed across the enterprise, speeding up code generation and debugging. These operational gains highlight the importance of aligning genetic AI capabilities with clear, high-value business use cases.

The role of hardware in the AI ​​generation

Hardware considerations for LLM scaling are often overlooked, but they play a critical role in long-term sustainability. Both Wayfair and Expedia currently rely on cloud infrastructure to manage their AI workloads. Tan notes that Wayfair continues to evaluate the scalability of cloud service providers like Google, while keeping an eye on the potential need for localized infrastructure to handle real-time applications more efficiently.

Expedia’s approach also emphasizes flexibility. Hosted mostly on AWSthe company uses a proxy layer to dynamically route tasks to the most appropriate computing environment. This system balances performance with cost efficiency, ensuring that inference costs do not spiral out of control. Naidu emphasizes the importance of this adaptability as enterprise-generation AI applications become increasingly complex and require higher processing power.

This focus on infrastructure reflects broader trends in enterprise computing reminiscent of the shift from monolithic data centers to microservices architectures. As companies like Wayfair and Expedia scale their LLM capabilities, they demonstrate the importance of balancing cloud scalability with emerging options like edge computing and custom chips.

Training, management and change management

Implementing LLMs is not just a technological challenge – it is a cultural one. Both Wayfair and Expedia emphasize the importance of fostering organizational readiness to adopt and integrate AI generation tools. At Wayfair, comprehensive training ensures departmental employees can adapt to new workflows, especially in areas like customer service, where AI-generated responses require human oversight to match the company’s voice and tone.

Expedia has taken governance a step further by creating a Council for Responsible AI to oversee all major AI-related decisions. This board ensures that implementation is aligned with ethical guidelines and business objectives, fostering trust across the organization. Naidoo emphasizes the importance of rethinking metrics to measure the performance of the AI ​​generation. Traditional KPIs often fall short, prompting Expedia to adopt precision and recall metrics that better align with business goals.

These cultural adaptations are critical to the long-term success of Gen AI in enterprise settings. Technology alone cannot bring about transformation; the transformation requires a workforce equipped to leverage Gen AI capabilities and a governance structure that ensures responsible implementation.

Lessons for scaling success

The experiences of Wayfair and Expedia offer valuable lessons for any organization looking to effectively scale LLM. Both companies demonstrate that success depends on identifying clear business use cases, maintaining flexibility in technology selection and fostering a culture of adaptation. Their hybrid approaches provide a model for balancing innovation with efficiency, ensuring that generational AI investments deliver tangible results.

What makes AI scaling in 2025 unprecedented challenge is the speed of technological and cultural change. The hybrid strategies, agile infrastructures, and strong data cultures that define successful AI deployments today will lay the foundation for the next wave of innovation. Enterprises building these foundations now won’t just scale AI; they will increase resilience, adaptability and competitive advantage.

Looking ahead, the challenges of inference costs, real-time capabilities, and evolving infrastructure needs will continue to shape the AI ​​landscape of the enterprise generation. As Naidu aptly says, “Gen AI and LLMs will be a long-term investment for us and that sets us apart in the travel space. We have to keep in mind that this will require some investment prioritization and understanding of use cases.”