Synopsis: The rush to automate is hitting a major speed bump as enterprises realize that bolting a high-performance AI engine onto a legacy horse-and-buggy process is a recipe for a pilot program disaster. Harshil Shah, Director of Engineering for R Systems, warns that the real challenge isn't building the agents themselves, but rather evolving our governance models from simple software checks to role-based oversight that treats AI like a new member of the engineering team. To survive this transition, organizations must move beyond the hype and prioritize an "observability-first" approach that builds evaluation and security guardrails into the foundation long before the first agent goes live.
Many organizations are rushing to insert AI agents into existing workflows without first evaluating whether those workflows are actually ready for autonomous systems. Shah compares the approach to bolting a high-performance engine onto a horse carriage. The technology may be powerful, but if the surrounding processes, data and governance structures are not designed for it, the result is often failed pilots and abandoned initiatives.
One of the first questions enterprises need to address is data readiness. In many organizations, critical information remains scattered across siloed systems or embedded as tribal knowledge within teams. Without a clear, unified view of where data lives and how it can be accessed, agents struggle to perform reliably.
Equally important is observability. Shah emphasizes that many organizations build agents first and worry about evaluation later. Instead, enterprises should establish the metrics, logging and tracing mechanisms needed to evaluate agent performance before deployment. Without that visibility, teams cannot determine why an agent failed, whether hallucinations occurred or how results can be improved.
Governance is another area that requires a shift in mindset. Shah argues that agents should not be treated as ordinary software components. In many ways they behave more like new team members that require defined responsibilities, oversight and clearly scoped permissions. Granting agents the same access rights as human users can create unnecessary risk.
Finally, organizations must address the cultural dimension. Employees may initially view agents as a threat to their roles, when in reality the shift is toward supervising and orchestrating automated processes rather than executing them directly.
The organizations that succeed with agentic AI will likely be those that treat it as an operational transformation, not simply another technology deployment.
