We believe in using the right tool for the right purpose. Every major AI provider now lets you build a custom assistant by giving a model instructions and some reference material: OpenAI’s ChatGPT (via Projects, and formerly GPTs), Microsoft’s Copilot agents and Copilot Studio, Google’s Gems in Gemini for Education, and Anthropic’s Projects in Claude for Education. Education-focused platforms such as Playlab and Mindjoy offer similar capabilities.
We are glad the field has converged on this approach; it tells us that giving people plain-language control over AI is the right idea. For some educators, one of these tools will be the right choice, particularly where an institution has already licensed it and the use case sits within that vendor’s ecosystem.
The landscape also shifts quickly. OpenAI announced in 2026 that custom GPTs will be retired, with existing GPTs ceasing to run in December. This is not a criticism of OpenAI, but it illustrates a real consideration for institutions: an agent built inside a vendor’s consumer product exists at that vendor’s discretion.
Cogniti was built by educators for a narrower problem: deploying AI to students, at scale, in a way that keeps the educator in control and gives them visibility over what happens. That focus is where the differences lie.
Where Cogniti differs #
Students use it without a vendor account. Cogniti is provisioned by the institution and reached through the LMS via LTI 1.3 (Canvas, Moodle, D2L Brightspace, and others) or a link. Students do not sign up with OpenAI, Microsoft, Google or Anthropic, and access does not depend on who can afford an individual subscription.
Educators can see how their agents are being used. Conversation histories are available (de-identified) to the educator who built the agent, along with AI-generated summaries of common questions and misconceptions. This is routinely reported as one of the most valuable features of the platform, and it is largely absent from general-purpose tools, where an agent’s conversations belong to the individual user.
Students can flag and rate AI responses. Feedback goes to the educator and administrators, closing the loop on quality and safety.
It is model-agnostic. Cogniti runs on Azure, so educators can choose between models from OpenAI, Anthropic and others for a given agent, and institutions are not tied to one vendor. User data is never used to train models.
Agents are shareable and clonable across institutions. Pedagogically designed templates and a growing library of shared agents (for example, the ‘designing for diversity’ coach and rubric assistants) mean educators can start from something that already works.
Students can build their own agents in educator-controlled sandboxes, supporting AI literacy without unsupervised access.
Interactive mini apps. Beyond conversational agents, educators can describe a quiz, simulation, game or visualisation in plain language and have it built, deployed via the LMS, and instrumented with telemetry they can query in natural language. General-purpose tools can generate interactive content, but not deploy it to a cohort with analytics attached.
Institution-level guardrails. Administrators can set organisation-wide instructions that every agent adheres to, alongside content safety filtering and audit logging.
Continuity. Because Cogniti sits above the model providers rather than inside one of their products, an agent’s instructions, resources and conversation history persist when a vendor changes course. When a model is retired, the educator switches the agent to another model; the agent itself does not disappear.
Comparison at a glance #
The table reflects our understanding. All of these products change quickly; please check the vendors’ current documentation, and let us know if we have something wrong.
| Cogniti | ChatGPT | Microsoft Copilot | Google Gemini | Claude | |
|---|---|---|---|---|---|
| Build an agent with plain-language instructions and uploaded resources | Yes, Cogniti agents | No, custom GPTs are being retired. ChatGPT Projects are shared workspaces with shared context and chats. | Yes, Copilot Studio agents | Yes, Gemini Gems | Partly, Claude Projects |
| Interactive learning activities deployable via the LMS to a cohort with telemetry | Yes (mini apps) | No | No | Partly, no telemetry or LMS embedding | Partly, Claude Artefacts, no telemetry or LMS embedding |
| Access provisioned by the institution, independent of student subscriptions | Yes | With ChatGPT Edu licence | With M365 licence; agents are pay-per-use | Free with Workspace for Education (limited quota) | With Claude for Education licence |
| LMS integration | LTI 1.3 and API; agents embed directly in Canvas, Moodle and others | No native LTI | Via Microsoft 365 and Teams; LMS support varies | Via Google Classroom | Canvas LTI for the Claude app; not per-agent |
| Educator can view student–agent conversations (de-identified) and AI summaries | Yes | No | Summaries available | No | No |
| Students can flag and rate AI messages, with feedback routed to educators | Yes | Rating goes to vendor | Rating goes to vendor | Rating goes to vendor | Rating goes to vendor |
| Choice of underlying model and vendor per agent | Yes (OpenAI, Anthropic, Gemini, Azure Foundry, etc) | OpenAI only | Models via Microsoft Azure | Google only | Anthropic only |
| User data not used to train models | Yes | Edu/Enterprise tiers | Enterprise data protection | Workspace for Education terms | Edu/Enterprise tiers |
| Share and clone agents across institutions, with pedagogical templates | Yes | Shared Projects within workspace | Within tenant | Within organisation | Within organisation |
| Students build their own agents and mini apps in educator-controlled sandboxes | Yes | Depends on licence and admin settings | Depends on licence | Depends on admin settings | Depends on licence |
| Organisation-wide guardrail instructions applied to every agent | Yes | Admin controls | Admin controls | Admin controls | Admin controls |
