AI tutoring was supposed to revolutionize education, but a year into the reality, the promise has quietly collapsed. Sal Khan, founder of Khan Academy, admitted in April 2026 that Khanmigo, the flagship AI tutoring product he once called "the biggest positive transformation that education has ever seen," became "a non-event" for most students. The gap between the hype and what actually shipped reveals a deeper truth about AI in education: the problem was never the technology itself.

What Happened to the AI Tutoring Revolution?

When Khan Academy launched Khanmigo, the vision was clear and compelling. Students would have access to personalized AI tutors available 24/7, catching up on challenging material independently. The reality proved far different. Text-based chatbots, it turns out, lack the engagement factor needed to keep teenagers motivated through independent study sessions. The tech worked fine in classrooms where educators integrated it into instruction, but voluntary use by students working alone on homework? That never materialized.

This wasn't a surprise to everyone. Observers who studied the EdTech landscape predicted this outcome a year earlier, noting that the fundamental issue wasn't capability but task design. Students need motivation to use a tool repeatedly, and a text-based chatbot asking them to work through difficult content independently simply didn't provide it. Khan's public retreat signals something larger: the standard-bearer for AI tutoring is now openly conceding that the promise didn't land.

Why Does the Technology Keep Failing at Scale?

The technical architecture behind most AI tutoring products relies on three core components: system prompts that define the bot's personality, fine-tuning to customize its behavior, and RAG (Retrieval-Augmented Generation), a technique that lets the AI pull information from a knowledge base. Together, these tools inject static knowledge into the system, whether that's curriculum materials or district policies. But the limitations are severe and haven't improved.

The most critical flaw is what researchers call "anterograde amnesia." Every time a student starts a new session, the bot begins from scratch. It has no memory of previous conversations, no understanding of who the student is, and no ability to learn from the content fed to it. The bot must re-parse all the curriculum materials with fresh eyes, as if encountering them for the first time. This means the AI tutoring system cannot actually learn about the student it's meant to help, defeating the core promise of personalization.

Could New Standards Like MCP Fix the Problem?

In November 2024, Anthropic released Model Context Protocol (MCP), a new standard that allows large language models (LLMs), the AI systems powering tools like Claude, to use tools through application programming interfaces (APIs). While tool use itself wasn't new, MCP provided the first industry-wide standard for how LLMs connect to external systems. Think of it like USB-C for AI: it didn't invent charging, but it gave every device the same plug.

In theory, MCP could solve the memory problem. A tutoring bot could now plug directly into the systems that hold student data: the Student Information System (SIS) for student identity, the Learning Management System (LMS) like Google Classroom or Canvas for assignments and grades, and adaptive assessment platforms for what students actually know. In practice, however, the drawbridge stays closed.

How to Understand Why EdTech Companies Won't Share Data

• SIS Access Control: Student information systems are tightly gated by school districts, who view student data as sensitive and proprietary, limiting third-party access.
• LMS Competitive Moat: Learning management systems like Google Classroom and Canvas control the infrastructure where assignments are posted, submitted, graded, and discussed, giving them enormous leverage to resist external AI tools.
• Adaptive Platform Data Lock: Assessment platforms like Renaissance and IXL have spent decades building deterministic content curation algorithms powered by national-scale longitudinal data, and they have no incentive to open that data to competitors.

The result is predictable. Why would any of these systems voluntarily lower their drawbridge to let a wrapper AI tool take their seat at the teacher's desk? They wouldn't. Instead, the merger happening in real time is between adaptive learning and generative AI within existing platforms. Renaissance, for example, launched its new Education Intelligence System earlier in 2026, combining both capabilities under one roof. Gemini is already baked into Google Classroom. Everyone is pulling the blanket toward themselves, leaving the wrapper AI tools naked.

MCP is undeniably convenient for plugging LLMs into existing workflows, but it doesn't solve the fundamental memory problem. Having a spreadsheet of student data is not the same as knowing the student. All the content fed to the machine to customize it to a district is still examined with fresh eyes at every query. The bot can retrieve information, but it cannot become a better tutor through repeated interaction.

What Does This Mean for the Future of AI in Schools?

The EdTech wrapper business faces an uncertain future. MagicSchool, the leader in the wrapper space, raised $45 million in February 2026 out of a total of $60 million since launch, suggesting either a hidden competitive advantage or a cautionary tale in the making. The real innovation is happening inside the platforms that already own the data and the relationships with schools.

The broader lesson extends beyond education. The energy bottleneck, the limitations of RAG, and the challenge of serving AI capabilities cheaply at scale are now hitting mainstream discourse. The bottleneck for useful AI tutors has never been the technology. It's the scale. Until the economics of running large language models become sustainable, and until the data silos that protect EdTech platforms crack open, the promise of personalized AI tutoring will remain exactly what it is today: a promise.