Implementation types

Institution-run, multi-model environments with institutional controls — one shared service offering a choice of models. Examples include U-M GPT, Amplify GenAI, TritonGPT, ZotGPT, and Harvard AI Sandbox.

A sandbox is best understood as one mode of such a platform — an experimentation space for trying models, prompts, and tools before production — governed by the same system card, but with its own use boundary (typically non-sensitive data only, with outputs reviewed before production use). It is not a separate implementation type.

Governance note. When a platform offers several selectable models, multimodal capabilities, APIs, and extensible tools, it needs two complementary lanes — platform governance and tool & use governance — kept under continuous review as models, modalities, and integrations change. The system card carries the model catalog, routing, and change management; the service card sets the boundaries on downstream tools and uses.

General-purpose consumer and enterprise productivity tools adopted institutionally — chat assistants, suite copilots, and creative AI. Examples include Microsoft 365 Copilot, Google Gemini, Claude, Adobe Firefly, NotebookLM, and ChatGPT Edu.

AI capabilities infused into core administrative systems — SIS/ERP, HR, finance, ITSM, security operations, and institutional data and analytics platforms. Examples include ServiceNow AI, Oracle and Workday AI, and Salesforce Einstein.

Recruitment, marketing, nudging, chatbots, and student-success and relationship tools. Examples include Element451, Ivy / Ocelot, Slate AI, and Salesforce Education Cloud with Einstein.

Course-level tools, tutoring, proctoring, plagiarism and gen-AI detection, and academic-integrity systems. Examples include Canvas AI features, Gemini LTI, tutoring assistants, and proctoring or detection tools.

High-governance sub-area. Assessment and academic-integrity AI — proctoring and gen-AI detection in particular — carries a distinct risk, equity, and shared-governance profile, and warrants closer review than other classroom tools.

The cross-cutting enabling layer: model routing, policy enforcement, logging and observability, red-teaming, and a centralized control plane that feeds the other categories. Examples include institutional AI gateways or control planes, cloud AI access layers, and AI toolkits.

Cross-cutting. Listed as its own category, but it often underlies the others — and it still earns its own system card and model references. In the demo, it is the gateway powering the managed platform.

Secure research workbenches, lab-specific tools, research data management, and domain-specific analysis environments. Examples include research clusters, secure local LLM deployments, and research-data workbenches.