Valentyn Kropov, CTO at N-iX, published a Forbes Technology Council framework on July 1 arguing that enterprise agentic AI projects fail not because of model selection or agent framework shortcomings, but because the integration work gets deprioritized after the demo.
His thesis is direct: “If your agentic AI project is failing, your problem is almost certainly not the model, and it is probably not even the agent. Your problem is that you treated the integration work as somebody else’s problem to solve after the demo.”
The Integration Gap
Kropov identifies a consistent pattern across enterprise deployments. Teams start by choosing a model, wrapping it in an agentic framework, and pointing it at their data. Model selection becomes the strategic decision. The systems architecture underneath, where the agent’s output flows back into operational workflows without breaking downstream dependencies, gets deferred.
Current frontier models, Kropov argues, can handle most enterprise tasks. The bottleneck is the unified access layer across systems that were never designed to interoperate: reconciling conflicting identifiers, engineering acceptable response times, and building enough logging and traceability for decisions to survive audit.
Compliance as Architecture, Not Afterthought
The second failure mode Kropov flags is treating compliance as a layer added after a working system exists. The version of agentic AI promoted most heavily over the past 18 months, fully autonomous end-to-end process execution with no human review gates, is incompatible with regulated industries.
“Legal and compliance teams will not sign off on a system where an algorithm makes a binding operational decision without human review,” Kropov writes. “Even in less regulated environments, one wrong automated decision multiplied by the volume an agent operates at costs more than skipping the review step.”
His recommendation: treat the agent as a senior specialist, not a replacement. It prepares recommendations and flags low-confidence outputs. A human reviews and makes the final call. Accountability stays with a person, which is where it needs to be for legal reasons in finance, healthcare, and insurance.
The Fleet Optimization Case Study
Kropov cites a deployment for a fleet operator managing more than 80,000 vehicles transporting millions of students daily. Before the agent deployment, an operations analyst spent days each week pulling data from 20+ disconnected enterprise systems, consolidating it for a mixed integer programming model, and configuring and interpreting results.
The agent layer automated system access, input preparation, output interpretation, and recommendation surfacing through conversational and web interfaces. But the visible interface, according to Kropov, was not where the effort went. The harder work was creating unified access across systems never designed to interoperate.
Results: optimization processing time fell 68%. More significantly, optimization savings increased 29%. The distinction matters. The first number means faster execution. The second means analysts shifted from data assembly to reviewing recommendations and investigating higher-value edge cases that previously went untouched.
Why This Keeps Happening
Kropov’s framework aligns with a pattern visible across recent enterprise deployments. AWS committed $1 billion to Forward Deployed Engineers specifically to embed integration expertise inside customer teams. Gartner declared 2026 an inflection year for enterprise agentic AI. Yet the gap between demo and production remains the primary failure mode.
The reason is structural. Model capability improves on a research timeline. Integration architecture improves on an engineering timeline. The second is slower, less visible, and harder to showcase in a pitch. Kropov calls the discipline “pragmatic AI software engineering,” framing it as an engineering problem rather than an AI problem.
The Regulated Industry Constraint
For teams deploying agents in finance, healthcare, or insurance, full autonomy is not a feature request deferred to a later sprint. It is architecturally impossible under current regulatory frameworks. Human-in-the-loop review gates are the only path through legal review, compliance review, and change management.
The teams that ship agent deployments in regulated environments are the ones that design for these constraints from day one rather than retrofitting them onto systems built for full autonomy. That distinction, between designing for review and adding review later, determines whether the deployment survives contact with the compliance team.