The Deployment Without a Framework

OpenAI announced this week that Samsung Electronics will deploy ChatGPT Enterprise and Codex across its South Korean operations and globally to its Device eXperience division employees. This is reportedly one of OpenAI's largest enterprise deployments to date. Almost simultaneously, a separate report surfaced describing what employees across industries are calling "AI sprawl": workers now navigating dozens of AI tools at once, with the proliferation driving confusion rather than productivity gains. These two stories, arriving in the same news cycle, are not coincidental. They describe the same structural condition from opposite ends of the organizational hierarchy.

What Samsung Is Actually Deploying

When a company the size of Samsung rolls out ChatGPT Enterprise across coding, automation, and general operations, the instinct is to frame this as a capability acquisition. The organization is purchasing access to intelligence, as InstaLILY CEO Amit Shah recently put it when distinguishing between companies that "rent intelligence versus companies that own it." But this framing contains a buried assumption that deserves scrutiny. Deploying a tool and developing the organizational competence to coordinate around that tool are not the same event. Samsung now has access. Whether Samsung has the schema to use that access effectively is a separate and more consequential question.

The Sprawl Problem Is a Schema Problem

The AI sprawl reporting is instructive here. Employees report not that the tools are bad, but that the accumulation of tools without structural coherence is cognitively overwhelming. This maps precisely onto what Kellogg, Valentine, and Christin (2020) describe in their review of algorithmic work: organizations implement algorithmic systems without resolving the coordination mechanisms that would allow workers to respond adaptively. The result is not incompetence at the individual level. It is a structural mismatch between the complexity of the deployed environment and the schemas workers have available to interpret it.

Hatano and Inagaki (1986) drew the foundational distinction between routine expertise and adaptive expertise. Routine expertise handles known procedures in stable environments. Adaptive expertise handles novel conditions by reasoning from underlying principles. A Samsung engineer who learns the specific prompt syntax for Codex has acquired routine expertise. An engineer who understands why large language models respond differently to structural versus semantic cues, and can adjust accordingly across tools, has acquired adaptive expertise. The distinction matters enormously when the tool set is expanding and changing, which is the current condition in every enterprise deploying AI at scale.

The Inversion That Large Deployments Miss

Classical organizational theory, drawing on coordination mechanisms from markets and hierarchies, assumes that competence precedes deployment. You train workers, then you give them the system. Platform coordination inverts this sequence: competence develops endogenously through engagement with the system itself (Schor et al., 2020). The ALC framework I am developing in my dissertation takes this inversion seriously. It predicts that workers given general schema-level training about how algorithmically-mediated systems are structured will outperform workers given platform-specific procedural training, particularly when the environment shifts. Large-scale enterprise deployments like Samsung's are a natural test case for this prediction, because the environment will shift. ChatGPT Enterprise today will not be the same system in eighteen months.

Gagrain, Naab, and Grub (2024) find that algorithmic media use does not automatically produce algorithmic literacy. Exposure and competence remain decoupled. This decoupling is precisely what large deployment announcements tend to obscure. Access is legible and annountageable. Schema development is slow, organizational, and harder to put in a press release.

Renting Intelligence Versus Owning the Framework

Shah's framing about renting versus owning intelligence points at a real strategic distinction, but I think it misidentifies the scarce resource. The scarce resource is not the model. OpenAI will continue licensing ChatGPT Enterprise to any company that can pay for it. The scarce resource is the organizational schema that allows workers to coordinate effectively within algorithmically-mediated environments, transfer that competence across tools, and adapt when the tools change. Companies that develop this structural understanding will not be locked into any single vendor. Companies that only accumulate tool access will reproduce the AI sprawl problem at increasing scale, regardless of whether they are renting from OpenAI or have built proprietary systems internally.

The Practical Question Samsung's Announcement Raises

Samsung's deployment is large enough and public enough that it will generate observable outcomes over the next year or two. The question worth tracking is not whether employees use ChatGPT Enterprise. They will. The question is whether the deployment produces variance reduction across the workforce or variance amplification. If power-law distributions emerge from this deployment, where a small fraction of employees generate disproportionate value while the majority plateau, that is evidence that access alone is insufficient and that the schema problem is structural rather than individual. That outcome would be theoretically informative, and based on what the current literature on algorithmic work suggests, it is the more likely one.