Safely manage your Zendesk from the AI assistant you already use, via the Deltastring MCP. Beacon configuration platform
← Back to news
ai

What billions of AI predictions taught Expedia before the age of AI agents

Expedia's experience with billions of AI predictions reveals a critical gap between proof-of-concept deployments and production-grade systems. The company's core insight—that velocity without discipline becomes a liability rather than an asset—directly challenges how most CX teams approach AI implementation. Rather than chasing the latest agent capabilities, Expedia learned that sustainable AI requires architectural discipline, rigorous measurement frameworks, and strategic alignment with business outcomes. This distinction matters acutely for teams evaluating tools like Agentforce or building custom solutions: the question isn't whether an AI agent can handle a customer interaction today, but whether your infrastructure can sustain thousands of daily predictions with consistent quality, cost control, and explainability. Many organisations optimise for initial deployment speed, only to discover their systems degrade under scale or become prohibitively expensive to maintain.

The implications for CX operations are substantial. Expedia's approach suggests that teams should audit their current AI implementations against three dimensions: whether predictions remain reliable as volume increases, whether the cost-per-resolution model actually improves unit economics at scale, and whether the system can be debugged and refined when performance drifts. This reframes the conversation around Salesforce's outcome-based pricing model, which aligns vendor incentives with actual resolution quality rather than deployment volume. For support leaders, this means demanding transparency from vendors about how their agents perform under load, not just in controlled pilots. The broader risk is that teams deploying agents without this discipline will face either escalating costs, declining customer satisfaction, or both—particularly as consumer frustration with AI in customer service grows.

The strategic takeaway is that AI maturity in CX isn't measured by feature adoption but by operational discipline. Teams should establish baseline metrics before deployment—resolution rates, cost per interaction, customer satisfaction—and treat agent performance as a continuous optimisation problem rather than a one-time implementation. This approach separates organisations building lasting competitive advantage from those simply riding the current wave of AI hype.