Over 40% of enterprise agentic AI projects will be canceled by the end of 2027, according to Gartner, with escalating costs, unclear business value, and inadequate risk controls cited as the primary drivers. Reuters confirmed the prediction in June 2025. An Observer analysis published today argues the cancellation wave is already visible: the most consistent failure point across early deployments is not agent capability but the underlying IT architecture enterprises are deploying agents into.
The Architecture Problem
The Observer piece frames agentic AI deployment as an operating system problem. Before agents can function reliably, an organization needs four layers: data architecture the agents can navigate, a governance layer defining what agents are permitted to do, an orchestration layer sequencing agent activity, and a human interface layer determining where autonomous execution stops and human judgment begins. Without this fabric in place, agents get deployed into silos.
The most frequent constraint surfacing in early deployments is data readiness. Agentic systems execute multi-step tasks autonomously across enterprise systems and require high-quality, structured, accessible data. Fragmented pipelines do not merely slow implementation, Observer reports — they corrupt it. The cost items that most frequently surprise organizations are not headline technology spend but high-frequency API calls at scale, custom connectors to legacy systems never designed for autonomous interaction, and ongoing operational costs for agent monitoring and incident response.
Returns Are Real but Highly Use-Case Dependent
Early adopters who get the architecture right report an average 171% return on investment, reaching 192% in the US, largely driven by reductions in manual processing hours, according to PagerDuty survey data cited in the Observer analysis. But the Observer cautions those figures reflect the most favorable deployments, not the median. Customer service automation, where performance is measurable and failure is immediately visible, tends to yield faster returns than back-office process automation, where errors compound quietly before surfacing. Timelines to attributable returns range from two to four years for complex multi-system deployments and 12 months for narrower implementations with clean data.
Why This Matters Now
This week alone, Amazon launched three agentic products across DevOps, security, and observability. Permiso shipped the first agent skill sandbox. The vendor side is accelerating. The Observer analysis is a useful counterweight: the supply of agentic AI products is outpacing most enterprises’ readiness to deploy them. For builders and operators evaluating agent deployment, the 40% cancellation rate from Gartner suggests the gap between vendor pitch and operational reality remains the primary risk to manage.