Three Major Companies Reveal: The Bottleneck for AI Agent Adoption Isn't the Model—It's the Infrastructure

LinkedIn, Walmart, and Zendesk—three companies of different scales and industries—independently pursued AI agent deployment at scale and arrived at the same conclusion: the problem is not the performance of the AI model itself, but the underlying system infrastructure on which it runs. This insight was shared during a panel discussion at the 2026 conference "VB Transform 2026."

Most enterprise systems have traditionally been designed with the assumption that humans would operate them. As long as the system kept pace with how quickly humans click, approve, and complete tasks, that was sufficient. But AI agents process information at millisecond speeds that humans could never match. This "fundamental speed mismatch" caused various failures in production environments. The gap each company faced was precisely this disconnect.

At LinkedIn, the container orchestration system Kubernetes became the first barrier. The traditional method of starting containers on demand took several seconds, which was far too slow compared to agent processing speed. Consequently, they switched to a "pre-pooling approach" where containers are prepared in advance, enabling real-time task switching. An additional challenge that emerged was how to control the agents themselves. When AI evaluates the output of another AI, the evaluating AI is prone to the same failures. LinkedIn responded by redesigning their system to handle approximately 80% of workflows deterministically through scripts, limiting AI reasoning only to scenarios where it is actually necessary.

Walmart faced a challenge born from success. When they opened an internal AI agent development environment to the company, even non-engineers began creating their own agents independently, and these proliferated rapidly throughout the organization. As a result, numerous duplicate agents with the same objectives were created, making management difficult. To address this, they implemented a "governance" framework that detects duplicates and promotes superior agents to production. Rather than restricting development freedom, they created a system to organize and control the chaotically expanding agents.

Sami Goush, formerly at Zendesk, took his current position following Zendesk's acquisition of Forethought in March 2026. The bottleneck he identified was on the data side. While possessing large data assets is an advantage, the design of which data to expose to agents directly determines actual response quality. When data design and management are deprioritized, he explained from firsthand experience, even rapid AI processing speed cannot deliver quality results.

What these three cases demonstrate is that moving AI agents from a "pilot testing stage" to an "actual organizational deployment stage" requires more than model improvements alone. Container startup latency, evaluation logic design, governance frameworks, data management—these seemingly unglamorous infrastructure issues are becoming the practical constraints on agent production deployment.

At a time when AI investment tends to concentrate on model development, these cases support the view that "infrastructure investment is key to widespread adoption." Going forward, the extent to which each company's agent implementations can scale horizontally, and how governance and infrastructure maturity become standardized across the industry, will be critical points to watch.