Enterprises are failing at AI not because their models are broken, but because their organizations are not built to use them, according to an analysis by Adi Polak, director for advocacy at Confluent.
The pattern repeats across dozens of initiatives: engineering teams build models that product managers cannot operate, data scientists produce prototypes that operations teams cannot maintain, and finished AI tools sit unused because the people they were built for had no say in defining what useful meant. The announcement says organizations that do extract value from AI share one common trait — they have built real collaboration across departments and established shared accountability for outcomes.
Spreading AI Literacy Past the Engineering Team
Polak’s first prescription is expanding AI literacy across every function, not by turning everyone into a data scientist, but by ensuring each role understands how AI applies to their specific work. Product managers need a realistic sense of what predictions or recommendations are achievable given available data. Designers need enough understanding to build features users will actually find useful. Analysts need to know which outputs require human validation and which can be trusted without review.
When only engineers hold that knowledge, collaboration collapses. Trade-offs go unevaluated. Interfaces get designed for capabilities nobody can articulate. Outputs go unvalidated.
The second fix targets a structural problem most companies handle badly: defining where AI acts on its own and where a human must approve first. Many organizations default to one of two extremes — routing every AI decision through manual review, or running systems with no guardrails at all. Neither works.
Rules, Playbooks, and Who Decides What
Polak recommends a framework built around three requirements: auditability, meaning the decision path can be traced; reproducibility, meaning that path can be recreated; and observability, meaning teams can monitor AI behavior as it unfolds. Practical boundaries need to be set explicitly — whether AI can approve routine configuration changes, recommend schema updates without implementing them, or deploy code to staging but not to production.
Without those rules, the report argues, organizations face a binary outcome: AI so constrained it offers no advantage, or AI making decisions no one can explain or reverse.
The third change is creating cross-functional playbooks — shared documentation developed by teams together, not handed down from above. These documents address operational specifics: how to test AI recommendations before production deployment, what the fallback procedure is when an automated deployment fails, who gets involved when a human overrides an AI decision, and how feedback loops back into the system.
The intent is not bureaucratic. It is to eliminate the inconsistent results and duplicated effort that emerge when every department improvises its own process.
Polak is direct about where the real gap sits. Technical performance in AI matters, but enterprises that focus entirely on model quality while leaving organizational structure unchanged are, in her framing, creating avoidable problems for themselves. The deployments that work, she says, treat cultural and workflow changes with the same seriousness as the technical build.
Photo by Yan Krukau on Pexels
This article is a curated summary based on third-party sources. Source: Read the original article
