A joint study by IDC and Lenovo reveals that 88% of AI proof-of-concept projects never make it to production, derailed by murky goals, unready data, and missing in-house expertise in agentic AI. The research highlights a recurring "build-vs.-run split" where the team that designs and deploys AI agents—often an external partner—hands off the system to an operating team that lacks the support to run it successfully. When operators try to manage agentic systems alone, errors multiply, and operational gaps can trigger cascading failures that turn into business disasters if a single agent strays beyond acceptable behavior.

The gap between proof-of-concept and production stems from a fundamental mismatch between controlled demos and real-world conditions. Serhii Semenchenko, CTO of Dedicatted, a Toronto-based AWS Premier Tier Partner for Agentic AI Specialization, explains that demos run on "clean, prepared data" with "no edge cases" and predefined inputs and outputs, while production environments are unpredictable and messy. Production-ready systems require logging, tracing, monitoring, circuit breakers, checks, and validation at every step to prepare for failure. Semenchenko's basic production checklist mandates that every agent have a timeout and that every output be logged for traceability. A separate survey from Modern Data Company found that 75.9% of enterprise data leaders and practitioners across 66 nations cite data quality and trust as barriers preventing AI systems from reaching production. In contrast, a PwC AI survey discovered that about two-thirds—66%—of AI agents that do make it to production deliver measurable value and boost productivity.

The report describes how agentic AI systems should mirror microservices architecture, with each agent granted its own permissions instead of static, long-term access credentials. Semenchenko emphasizes that agents should use short-lived keys granting access only to specific tools, whether those are external services or databases. Treating every agent as a separate service prevents data leaks, unintended use, and prompt injection attacks by stopping agents from calling tools they shouldn't touch. In multi-agent orchestration, a router collects inputs from the user, system, and platform, then decides how to handle the data—escalating to specialist agents for documentation, analysis, video, audio, or livestreams, each with its own set of third-party APIs. If the final composition step fails, the system doesn't invent missing pieces; it either retries or marks the failure. The report also warns of "memory poisoning," where corrupted agent memory is countered by guardrails—deterministic checks that ensure data validity and trace inconsistencies back to records, databases, and duplicates.

Real-world deployment shows these principles in action. Semenchenko recounts a multi-agent system built for CASSEN Testing Laboratories, a Canadian chemical environmental testing and occupational health lab facing multi-day turnaround times due to manual regulatory testing. While the project had clear input and output expectations, human-related edge cases emerged when chemists had personal preferences for how reports and chemical compounds from error-sample investigations should be presented, with report styles more in-house than domain-specific. The team adjusted the agent that composed reports so the final outputs looked identical to manually prepared versions. Another case involved Zivian Health, a U.S.-based compliance infrastructure for organizations managing nurse practitioner and physician assistant workforces across multiple states, which migrated from AWS to Azure under regulatory stakes. Because AWS and Azure use completely different access systems, everything had to be mapped, adjusted, and validated first, with cloud-agnostic tools deployed to monitor and trace all activity on the agentic AI platform. After migration, the team validated every agent, use case, and user flow through a data set covering all platform requirements, then checked whether they achieved the same success rate, latency, and token cost. Denys Yermakov, CEO of Dedicatted, believes the commercial reality of agent failure is often quieter, with the price of failure much higher compared to other platforms. He stresses that change management requires flexibility in the solution and adaptability for future change requests, and that enterprises must question the agentic AI partner about project scope and discuss ownership of code and platform, since the team will eventually handle incident response and maintenance. Yermakov advises verifying a partner's expertise by starting with a small piece of proof-of-concept or demo, delivered on real data with real workflows and real business challenges—not mock data or predefined scenarios irrelevant to the client's business case. He gives a thumbs up to partners who face challenges on the client data layer during demos and come with the right questions and approach. On security, Yermakov underscores that data must stay within the client's secure perimeter and not leave to train public models, a critical concern for traditional businesses with large-scale on-premises data. Semenchenko adds a technical layer from healthcare projects where personally identifiable information around health is critical: an additional layer cleans all PII data from agent outputs before traceable data goes to the observability platform, ensuring no PII appears in logs and monitoring. The report concludes that agentic AI project success hinges on staffing—always a team effort combining in-house and outsourcing resources. Outsourcing everything won't work because the partner can't understand the business in detailed depth, while building entirely in-house wastes the partner's experience shipping systems after overcoming challenges and knowing how to architect platforms efficiently with predictable scenarios. Subject matter experts and partners must engage from the start to achieve optimal knowledge transfer, with the in-house team owning and safeguarding the business process while the partner architects and operates the core runtime. For enterprises still asking basic questions—how to set up everything, what the platform should look like, what the benefits are, what the cost per user is—the report stresses it's critical to build the team and knowledge in-house for future platform maturity, and to architect platforms correctly from the very beginning to avoid redoing what's already been implemented into business processes. The lesson is clear: production readiness isn't a technical checklist but a deliberate handoff between those who build and those who run, with both sides fluent in the system's limits and prepared for the messy edge cases that demos never show. Organizations that treat agentic AI as a shared responsibility between partner expertise and internal ownership stand the best chance of crossing the chasm from proof-of-concept to sustained business value.