Enterprise

'Built for humans, not agents': Meta, LinkedIn, Walmart say enterprise infrastructure is breaking under AI

Post-VB Transform 2026 reporting crystallized a consensus verdict, and Cisco data puts the pilot-to-production chasm at 85% vs. 5%.

By Asher Glick · Enterprise software desk · 2026-07-18 · 3 min read

Photo: Unsplash / Taylor Vick — Rows of network cables and servers inside a data center

Agentic queries into Meta’s data systems grew 30x in a single half, and Barak Yagour, the VP of engineering who runs that infrastructure, told VB Transform 2026 the company has maybe 20 months to rebuild for a workload it wasn’t designed to serve. That’s the frame that stuck when reporters from the Hotel Nia stage in Menlo Park filed post-event pieces this week. The consensus verdict: the stack was built for humans, not agents, and the retrofit window is short.

The numbers underneath the panic are less rhetorical than they sound. Automated traffic crossed 51% of internet volume last year, per data cited from Imperva and HUMAN Security, and it’s growing roughly eight times faster than human traffic. Cisco’s evaluation-gap survey, run through VB Pulse in June 2026 across 157 enterprise respondents (38% final decision-makers, all at 100+ employee companies), puts 85% of enterprises piloting AI agents and only 5% shipping to production.

That gap is where the operational damage lives. Half of enterprises have deployed an agent or LLM feature that passed internal evaluations and still produced a customer-facing failure; one in four had it happen more than once. Only 5% fully trust the automated evals gating release. And yet 66% already permit some production deployment without human review, or plan to within 12 months. Monitoring is thinner still: 51% watch only whether the agent is running, and 23% check whether its answers are right.

The practitioners on stage described what a rebuild actually looks like. Animesh Singh of LinkedIn walked through a five-point eval system and noted that roughly 80% of LinkedIn’s agent workflow is now scripted, deterministic code. Desiree Gosby, Walmart’s SVP of corporate technology services, and Sami Ghoche, the Zendesk applied-AI VP who arrived via the March 2026 Forethought acquisition, described similar scaffolding around Zendesk’s 20 billion customer conversations. At Meta, Yagour said 63% of dashboards published across the company within three months were built via new agentic data tooling, with GitHub Copilot already writing 46% of the average developer’s code.

The compute side of the ledger sharpens the irony. A separate 107-enterprise survey found 86% running GPUs at half utilization or less, fewer than half rigorously tracking compute cost, 54% reporting an agent security incident or near-miss in the past year, and 27% exercising only reactive control over agent spend. Between 57% and 64% of respondents across five control layers plan to switch or add vendors within 12 months. Into that churn Cohere is pitching Command A+, a 218-billion-parameter mixture-of-experts model with 25 billion active parameters, with VP Rachad Alao arguing sovereignty requires owning the full stack.

The 2008 capacity-planning literature assumed traffic that browsed. This traffic reasons in milliseconds and doesn’t stop to read.

Sources