DataCamp scales AI tutor to one million learning hours annually

DataCamp has scaled its AI tutor from near-zero usage nine months ago to a run rate of over one million hours of learning per year, representing a portion of the platform’s total 10 million learning hours. The company recently acquired Optima, a learning firm founded by Chief AI Officer Yusuf Sabar, to bolster this effort. In A/B tests, the AI tutor demonstrated higher willingness to pay, engagement, and satisfaction compared to the classic DataCamp experience. The system adapts lesson flow based on learner background and knowledge gaps, using a multi-agent system to deliver personalized feedback. DataCamp is also developing an "AI creator" tool to ingest existing documents and transform them into teachable content.

To manage costs, which were initially 10 to 20 times higher than static content, DataCamp employs a multi-model architecture. A frontier model, Sonnet 5.5, handles about 70% of the work to maintain coherence over long lessons, while smaller models like Haiku and Gemini Flash handle 5-7% of simpler tasks. The company is experimenting with open-source models like Gemma 4, which outperformed previous performers in evaluations, though challenges remain regarding hosting and latency. CEO Jonathan Cornelison argues that the current tutor is the "worst" version the company will ever have, indicating they are at the start of an improvement curve. He also contends that reducing junior hiring due to AI automation is a mistake, as young, "AI native" talent will adapt faster than senior staff, creating a temporary but significant labor market imbalance.