The category leader said his own AI tutor is “a non-event” for most students. The bottleneck has a name now — and the layer that solves it already ships.
Three weeks ago, in an interview with Chalkbeat, Sal Khan said something that should reframe how every product team in EdTech AI thinks about the next 12 months.
“For many students, Khanmigo was a non-event. They just didn’t use it much.”
Source: Chalkbeat — “Why Sal Khan is rethinking how AI will change schools,” April 9, 2026.
This is the founder of the Khan Academy. The person who, three years ago, said AI was about to revolutionize learning. He just publicly described his own flagship AI product as a non-event.
It would be easy to read that as a Khan Academy problem. It isn’t. It’s a category problem. And the diagnosis matters more than the headline, because it tells every EdTech AI builder what’s worth investing in next — and what isn’t.
The way the category talks about itself today is in two layers: the AI and the platform. That framing is wrong. There are four layers.
Layer 1 is content. The AI that explains, drills, summarizes, and plans. Khanmigo, Duolingo, MagicSchool, Gizmo, Century Tech. This is mostly solved. The 2025 randomized trials show AI tutors deliver real instructional gains — when used.
Layer 2 is delivery. The cameras, the panels, the LMS pipes that put content in front of the learner. Promethean, SMART, Logitech, Owl Labs, Zoom, Google Classroom. This layer is commoditizing fast. Q1 2026 brought four AI cameras with nearly identical features.
Layer 3 is engagement. Whether the human in front of the camera or the screen is actually paying attention, learning, or about to give up. This is where Sal Khan just told us the gap is.
Layer 4 is adaptation. Routing the next prompt, the next break, the human nudge based on engagement signal. Layer 4 only works if Layer 3 exists.
Khan Academy released its own 2026 Khanmigo numbers earlier this year. 108 million total interactions since 2023. 269,000 on a typical weekday this school year. Students who engaged 30 minutes per week in the 2026 pilot gained 2–3 weeks of instruction.
100% of eligible students were given access to Khanmigo. 15% used it. 85% never engaged.
The content layer worked for the students who showed up. The other 85% never opened the door. That isn’t a content problem and it isn’t an “AI is too new” problem. It’s an engagement signal problem. Nobody told the platform — and more importantly, nobody told the teacher — which 85% needed a different intervention.
Khan Academy launched “Reimagined for Districts 2026” the same week Sal Khan was publicly admitting Khanmigo’s adoption ceiling. The features are real and useful: team games, classroom leaderboards, streak motivation, a refreshed teacher dashboard, KAD Enterprise Starter for sub-1,000-student districts, expansion to 180+ countries with Microsoft.
Notice what isn’t on that list. A real-time measure of who is disengaged in the moment they’re disengaging. Cognitive load detection. Group-level engagement across a classroom. The signal that flags a student before the click — not the report that explains it three days later.
That’s because the signal doesn’t come from telemetry. It comes from the human. And reading it requires a different kind of measurement than the LMS layer can produce.
Augment Me reads engagement directly from the camera that’s already on the laptop, the classroom panel, the conference room. Patent-pending rPPG (remote photoplethysmography) measures subtle blood-flow signals through the skin — heart rate, heart rate variability, respiration, focus, cognitive load. No wearable. No special hardware. No facial emotion recognition.
In the last six months, four peer-reviewed papers — Nature npj Digital Medicine, Springer Applied Intelligence, Frontiers in Digital Health, MDPI Biomimetics — moved rPPG from “emerging research” to “clinical-grade.” That threshold matters because it means features built on top can be defended to procurement, compliance, and legal.
The EU AI Act’s high-risk rules fully enforce on August 2, 2026. Emotion recognition in workplaces and schools is already prohibited as of February 2025. Fines: €35M or 7% of global turnover.
Emotion recognition reads facial expressions and infers emotion. That’s the prohibited path. rPPG reads physiological signal and measures cognitive engagement. That’s a different category entirely. For any platform shipping into European classrooms or training environments next fall, the distinction is the difference between launching and pulling the product.
We’ve been building this layer in production since 2023. Augment Me is integrated into Explain Everything (Promethean, NYSE: MYND). 2M+ classrooms. 126 countries. Promethean was named BETT Company of the Year 2026 in January. Two-year Title I study: students surpassed national growth benchmarks by 100–400 points.
May 2026: expansion to 1M+ additional classrooms.
The category leader has just publicly named the gap. The cameras are at parity. The compliance window is closing. The research is over the threshold. Procurement teams are starting to ask “engagement intelligence” by name in RFPs.
If you’re running product at a camera OEM — Owl Labs, Jabra, HP Poly, Cisco Webex, Crestron, Neat, Yealink, Huddly — the next differentiator that can’t be shipped from a hardware roadmap alone is engagement intelligence. Our API runs on the feed your cameras already capture.
If you’re running an EdTech or enterprise training platform with video — Cornerstone Immerse (formerly Talespin), Mursion, Virti, Allego, Gong, MindTickle, Discovery Education, Canvas — the engagement signal is the missing layer between your content and your outcomes. We have the proof point. We have the pipeline.
Sal Khan’s reflection isn’t a setback for AI in education. It’s the moment the category names the bottleneck out loud. That’s the moment to build.
Sameer Yami is the Founder and CEO of Augment Me. Augment Me’s patent-pending rPPG technology is live inside Promethean’s Explain Everything platform (NYSE: MYND).
Comment