Enterprise AI Agents at Scale: Infrastructure Lessons from LinkedIn, Walmart and Zendesk

What happened

At VB Transform 2026, senior technology leaders from LinkedIn, Walmart and Zendesk reached a shared verdict on what is genuinely blocking enterprise AI agents at scale: the problem is not the models, it is the infrastructure built around them. Speaking on a panel, Animesh Singh (Senior Director of AI Platform and Infrastructure, LinkedIn), Desiree Gosby (SVP of Corporate Technology Services and Technology Strategy, Walmart) and Sami Ghoche (VP of Applied AI, Zendesk) each described what broke when their organisations moved AI agents from controlled pilots into live production environments.

The unifying insight across all three accounts was that most enterprise infrastructure was designed around human working speeds — batch processes, scheduled jobs, synchronous workflows — and AI agents operate on an entirely different tempo, reasoning and acting in milliseconds. Closing that gap required deliberate re-engineering, not simply plugging a new model into an existing stack. Gosby, reflecting on Walmart's experience scaling agents within its own workforce operations, was direct: the lessons came from what failed in production, not from what looked promising in the lab.

Why it matters

For customer experience practitioners, this panel surfaces a critical and frequently underestimated risk: organisations that invest heavily in AI agent capabilities but leave legacy infrastructure untouched are, in effect, fitting a high-performance engine into a vehicle whose chassis cannot handle the load. The customer-facing consequence is latency, inconsistency and failure at precisely the moments that matter most — high-volume service interactions, real-time personalisation and autonomous resolution of complex queries. Behavioral economics tells us that customers are acutely sensitive to response speed and reliability; a hesitant or erratic agent erodes trust faster than no agent at all.

From a service-design perspective, the shift these three companies describe is architectural as much as it is technological. Designing for agent-speed means rethinking data availability, memory, orchestration and fallback logic from first principles — not retrofitting them. The organisations that get this right will be able to deliver genuinely autonomous, low-friction service experiences; those that do not will find their AI investments producing marginal gains at best and customer frustration at worst.

The Renascence take

The conversation at VB Transform 2026 will be read by many as a technical briefing for engineers. It should be read by CX leaders as a procurement and governance warning. The decision about which infrastructure to modernise — and in what sequence — is a customer experience decision, not merely an IT one.

Most organisations are measuring AI agent success by containment rates and deflection metrics, when the more revealing signal is infrastructure-induced latency and its downstream effect on customer trust. The behavioral principle at work is straightforward: humans calibrate confidence in an agent — human or artificial — by its fluency and speed, not its accuracy alone. A technically correct answer delivered haltingly feels wrong. What Walmart, LinkedIn and Zendesk are really describing is the engineering cost of perceived competence. Customer-obsessed operators should audit their agent journeys not for model quality, but for the invisible delays their own legacy systems are inserting — because that is where the experience is actually breaking.