Enterprise has spent 85% of its AI budget building pilots that will never see production, and the problem isn't what anyone thought it was.
The Summary
- 85% of enterprises pilot AI agents, but only 5% ship to production — Amazon's AGI director says the blocker is reliability measurement, not model capability
- Agentic queries at Meta grew 30x in six months, breaking infrastructure assumptions built over two decades — agents now outnumber humans online
- Real agent failures look like this: a vision model reads serial numbers perfectly for two months, then quietly starts reading wrong numbers because a human-imperceptible UI change altered pixel positioning
- Companies need four-dimensional reliability frameworks (consistency, robustness, predictability, safety) and sovereign control over the full stack to make agents production-ready
The Signal
The gap between pilot and production is a measurement problem masquerading as a capability problem. Bryan Silverthorn, who leads multimodal agent training at Amazon's AGI lab after joining through the Adept acquisition, described a customer deploying an agent for software QA. Serial number extraction from screens. Worked flawlessly for two months. Then it didn't.
The vision encoder behaved differently depending on where serial numbers appeared on screen. A software update humans couldn't detect triggered cascading failures. The agent didn't get dumber. The evaluation framework never caught what mattered. Silverthorn argues reliability has four distinct dimensions: consistency, robustness, predictability, and safety. Most internal evals tangle these together, which is why agents ace benchmarks and collapse in production.
"The models have to be better. Obviously, we're working hard on making the models better. But the deeper takeaway is about measurement, not just models."
The infrastructure side is equally broken. Meta's VP of Engineering Barak Yagour said agentic queries hitting the company's data systems grew 30x in a single half, inverting assumptions the company spent 20 years building. Three things are breaking simultaneously:
- Capacity: Agent workloads don't behave like human workloads. They spike, sustain, and scale differently.
- Identity: Systems built to authenticate individual users don't know what to do when the requester is an agent acting on behalf of 10,000 users.
- Velocity: Automated traffic overtook human traffic last year at 51% and is growing eight times faster, according to HUMAN Security's 2026 report.
Yagour opened his VB Transform talk wearing Ray-Ban Meta AI glasses. The agents aren't coming. They're already here, and your infrastructure was designed for a species that no longer represents the majority of your traffic.
Cohere's VP of product engineering Rachad Alao added another constraint: sovereignty. For banks, hospitals, and governments, deploying agents means controlling the full stack, from GPUs and private cloud through governance systems, connectors, search tools, and agent frameworks. "You want to have control on the entire stack," Alao said. Sovereignty isn't downloading an open model. It's jurisdiction over every layer an agent touches when it acts on your data.
The economic assumption is also inverting. Token prices are falling, but Cohere's Marshall challenged whether agent workloads might outrun those cost declines. If agents make 100x more requests than humans, cheaper tokens don't save you. Volume multiplies faster than price declines.
The Implication
If you're running an enterprise AI pilot, ask this: what breaks when the agent does something imperceptibly different from what it did yesterday, and how will you know before your customers do. The 17:1 ratio between pilot and production isn't a maturity gap. It's a measurement gap. Companies that ship agents to production will be the ones that build reliability frameworks that separate consistency from robustness from predictability from safety, and treat each as a distinct engineering problem.
For infrastructure teams, Yagour's timeline matters. 20 months is not a suggestion. It's a clock. Your capacity models, identity systems, and rate limiters were designed for human behavior patterns that no longer describe the majority of your traffic. The agent economy doesn't wait for your infrastructure roadmap to catch up.