McKinsey survey finds only 10% of AI agent experiments scale beyond pilot stage

The survey underscores that agentic AI’s strengths—probabilistic reasoning, tool use, and autonomy—also create compounded risks. Errors in context, tool misuse, or policy violations can escalate quickly, especially when agents rely on outdated or flawed data. Non-determinism further complicates incident response, as failures depend on unpredictable variables like model versions, retrieved documents, and intermediate decisions. Cost and latency spikes may signal inefficiency or loops, but without centralized visibility, teams struggle to detect drift or reconstruct workflows after incidents.

McKinsey notes that organizations achieving real AI impact are nearly three times more likely to have redesigned workflows rather than retrofitting AI. Reliability requires continuous monitoring, guardrails on autonomous actions, and human oversight for high-risk tasks. The report introduces AI SRE (Site Reliability Engineering) as a framework to connect decentralized innovation with centralized oversight, ensuring agents operate safely and predictably at scale.