SchoolAI hits 1M classrooms by ensuring AI tutors don't cheat

Caleb Hicks remembers the blind spot. Back in 2013, teaching nearly 300 students a day, he could easily spot the kids who were soaring and the ones crashing. But the middle 80 percent? They were a blur. A decade later, his startup SchoolAI is now in a million classrooms across more than 80 countries, trying to fix exactly that problem with a platform that pointedly refuses to give students answers.

While school districts spent 2023 panicking about ChatGPT-enabled cheating, Hicks saw an opening. His pitch isn’t about AI replacing the teacher but about building a “teacher-in-the-loop” system that surfaces student struggles before they become report card nightmares. The architecture relies heavily on OpenAI’s latest models. GPT-4.1 handles deep reasoning for multi-step math problems, GPT-4o powers a conversational assistant called Dot for lesson planning, and text-to-speech provides spoken feedback in over 60 languages. That last feature proved critical for one newly arrived student who spoke only Dari. Using the platform’s Sidekick tutor for real-time translation, the kid was participating in group work within weeks.

But the real engineering flex is the guardrails. Instead of a simple prompt-and-response, every student input passes through an agent graph with dozens of nodes running specialized checks before anyone sees a reply. It’s an expensive way to build a chatbot, so the team routes only the heavy lifting to the big models and offloads quick verification checks to lighter models like GPT-4o-mini to keep costs from spiraling.

“If AI just gives the student the answer, we’ve failed,” Hicks says. “The point of teaching is to coach and to keep them engaged in the work.” That philosophy seems to be landing. Teachers report saving over 10 hours a week, and the company recently drew 10,000 educators to a product showcase. When a startup’s big risk is scale stress right before a keynote, you know the growth is real.