Platform Boundary Resources Are the New IT Artifact

When OpenAI changed what GPT-4 could and could not do through its function calling interface, it did something that most people treated as a product update. It was not. It was a governance decision. The function calling API, the system instructions, the context window limits, the fine-tuning permissions, these are boundary resources. And when a platform owner changes a boundary resource, they are not just updating software. They are reshaping who gets to participate in the ecosystem, under what terms, and with how much autonomy.

I kept running into this same pattern while reading Mayer et al. (2025). Their JMIS paper on how generative AI reshapes platform governance through boundary resources made visible something I had been sensing but could not name precisely. The thing that matters most in a platform ecosystem is not the algorithm or the model. It is the interface between the platform owner and the complementors, the boundary resource, and whoever controls that interface controls the ecosystem. Mayer et al. studied Dugga, an EdTech platform that introduced an AI Exam Engine as a boundary resource for teachers creating assessment content. What they found was a cycle of resistance and accommodation. Teachers resisted the GenAI tool because it threatened their expertise and produced unreliable content. Dugga responded by shifting from what Mayer et al. call a logic of "GenAI hegemony," where the AI tool was positioned as superior, to a logic of "GenAI assemblage," where teachers were repositioned as custodians of the AI's output. Each shift required Dugga to reconfigure the boundary resource itself, adding guardrails, allowing teachers to upload their own materials, adding disclaimers about AI errors. The boundary resource was not static. It was the arena where governance happened.

Orlikowski and Iacono (2001) asked what the IT artifact is in IS research and found that 88% of the papers they reviewed did not theorize it properly. Most papers treated IT as a setting, a background variable, something present but not doing theoretical work. Five views: Nominal, Computational, Tool, Proxy, Ensemble. Only the Ensemble view, where the artifact is inseparable from the social practices around it, treated IT as a genuine theoretical actor. I wrote about this before. The 88% number still bothers me. But here is what I think has shifted since 2001. The IT artifact that matters most now is not the hardware, the software, or even the user interface. It is the boundary resource. The API, the SDK, the prompt structure, the system instruction, the content policy enforced through model behavior. These are the artifacts that determine who can participate in a platform, what they can build, and how much control they retain over their own output.

Mayer et al. define boundary resources as "software tools and regulations that serve as the interface for the arm's-length relationship between the platform owner and the application developer," drawing on Eaton et al. The definition is straightforward, but what they show in the Dugga case is that boundary resources are not just interfaces. They are contested. When Dugga introduced the AI Exam Engine, teachers perceived material resistance from the tool, it contradicted their goals of producing reliable content and maintaining expertise. They disengaged. Dugga then reconfigured the boundary resource to give teachers more control: the ability to turn the AI off, upload their own content to guide it, and see disclaimers about AI errors. In distributed tuning terms, the boundary resource became a site of negotiation. The platform owner and the complementors took turns asserting their goals, and every reconfiguration of the boundary resource encoded a new settlement of power.

Heimburg et al. (2025) push this further. They studied complementors in OpenAI's platform ecosystem and identified two facets that make GenAI platform ecosystems fundamentally different from traditional ones. Open-endedness means the interfaces are not confined to a specific format or syntax. You can prompt a model in natural language and get almost anything back. Inscrutability means the relationship between input and output is indeterminate and probabilistic. These are not bugs. They are structural features that change what a boundary resource even is. In a traditional platform, the boundary resource is a well-defined API with documentation, predictable inputs and outputs, and reusable functionality. In a GenAI platform, the boundary resource is an opaque model that might do anything, and the complementor has to build wrappers around it, system instructions, curated user inputs, context data, output revision, just to make it usable. Heimburg et al. identify four value co-creation mechanisms: utilizing system instructions, providing context data, curating user inputs, and revising AI model outputs. These are not enhancements. They are the mechanisms through which complementors create value at all. Without them, the boundary resource is too unpredictable to be useful.

Think about what this means for governance. In a traditional platform, the boundary resource constrains what complementors can do. The API defines the possible actions. The documentation tells you what inputs produce what outputs. Platform governance through boundary resources is about controlling the shape of participation. But in a GenAI platform, the boundary resource does not constrain in the same way. It enables a vast range of possible outputs, most of them unpredictable. So platform governance has to shift from constraining the possible to shaping the probable. This is exactly what Mayer et al. found. Dugga shifted from promoting the AI's superior capabilities to emphasizing that teachers were still in charge, adding guardrails, disclaimers, opt-out options. The governance moved from controlling what the boundary resource could do to managing how complementors should relate to what the boundary resource might do.

Mohlmann et al. (2025) approached governance from a different angle, studying YouTube and what they call algorithmic stakeholder governance. Their argument is that algorithmic governance alone is insufficient because fully automated systems struggle to address real human concerns, especially when interests conflict and cannot be reduced to binary decisions. Platforms need stakeholders, creators, consumers, advertisers, to participate in governance through critical reporting and feedback. Mohlmann et al. use the lead role perspective from stakeholder governance theory, where the platform acts as an arbiter in conflict resolution rather than a top-down controller. What struck me is the parallel with Mayer et al. Both papers find that centralized platform control breaks down when you introduce AI. The algorithm cannot do it alone. And the boundary resource cannot govern itself.

Platform governance in multi-sided markets has always been about balancing autonomy and control for complementors. Network effects create the pull, but boundary resources define the terms. What the 2025 papers show is that GenAI boundary resources are categorically different from APIs, SDKs, and plug-ins. They lack consistency and standardization. They exacerbate validation issues. They require new skills from complementors. And because they are opaque and unpredictable, the platform owner cannot simply write better documentation or tighten the API specs. The governance has to be relational, distributed, ongoing.

I think Orlikowski and Iacono would recognize what is happening here. When they wrote about the Ensemble view in 2001, they argued that the IT artifact is embedded in social practice and constitutes part of the causal story. Boundary resources in GenAI platforms are exactly this. They are not neutral interfaces. They encode power, they shift through resistance and accommodation, and they are shaped by the very people they are supposed to govern. The difference is that in 2001, most researchers were ignoring the artifact. In 2025, the artifact is impossible to ignore, because it has become the thing everyone is arguing about.

When OpenAI decides which functions GPT-4 can call, what system instructions are allowed, what content policies are enforced through model behavior, and what fine-tuning is available to developers, it is making boundary resource governance decisions. Every one of those choices shapes what complementors can build, how reliable their products are, and who gets to participate. AI vendor concentration creates resource dependency, and that dependency is mediated through boundary resources. The model is the platform core. The boundary resource is where the leverage lives.

Mayer et al. showed that GenAI boundary resources shift governance from centralized control to what they call polyphonic governance, where complementors' expertise becomes essential for validating AI content. Heimburg et al. showed that complementors in GenAI ecosystems have to build new mechanisms, system instructions, context windows, curated inputs, output revision, just to make the boundary resource usable. Mohlmann et al. showed that algorithmic governance fails without stakeholder participation, because fully automated systems cannot resolve conflicts that require human judgment. Together, these papers tell a story: the boundary resource is the IT artifact that matters, and governing it is the core challenge of platform ecosystems in the GenAI era.

I think this reframes the Orlikowski and Iacono question. They asked where the IT artifact was in IS research. The answer in 2026 is that it is sitting right there at the platform boundary, in the API terms, the system instructions, the content policies, the context window limits, and the function calling constraints. Every time a platform adjusts these, it is not a technical update. It is an act of governance. And every time IS research treats the boundary resource as a neutral conduit rather than a political artifact, it repeats the mistake that 88% of the papers Orlikowski and Iacono reviewed made twenty-five years ago.