A joint IDC and Lenovo study reports that 88% of AI proof-of-concept (POC) projects fail to reach production because of unclear objectives, unprepared data, and a lack of in-house agentic AI expertise. These result in a build-vs.-run split, where the team that designed and deployed AI agents, often an agentic AI partner, doesn’t get to run them over the business processes.

When the operating team, often experienced on demo, uses the agentic system without any external help, errors arise. Operational gaps can cause cascading failures and business catastrophes if a single agent deviates from normal behavior beyond the permissible range.

When asked about estimating whether an agentic AI system is production-ready or limited to demos, Serhii Semenchenko, the CTO of Dedicatted, a Toronto-based AWS Premier Tier Partner for Agentic AI Specialization, tells The New Stack, “The demo is a clean environment where we have clean, prepared data. No edge cases, so everything is defined — input and output. “

He means the demo data is scripted with clean prompts, while the production environment is unpredictable. “Every single step should include the logging, tracing, monitoring, breakers, all the checks, and validation,” he adds. All these steps ensure failure preparedness.

Semenchenko shares a basic production checklist: “every single agent should have a timeout and every single step that agents, like, generate some output should be logged for traceability.” Broadly, he believes agentic AI systems should self-assess in production, decide whether to continue after a particular event or reschedule tasks.

Semenchenko then compares agentic AI systems to microservices architecture, where each agent has its own permissions rather than static or long-term broad access keys/tokens/credentials.

Source link