Remember when the big controversy in online education was whether a recorded lecture could ever match the real thing? MOOCs promised to democratize elite universities, we all signed up for courses we never finished, and the debate settled into a comfortable stalemate: video is fine for content, but feedback needs a human. That stalemate just got interesting. Harvard Business School is now selling an eight-week entrepreneur bootcamp, called HBS Foundry, for $699 — and part of what you get is AI avatars of its instructors, built by a startup called HeyGen, giving you feedback during practice sessions.
The program isn’t fully synthetic. There are live sessions with instructors every week, plus virtual simulations. But the avatars are the part that caught my attention, because the avatars are the part I know how to build. And as someone who spends his days wiring up bots, I have thoughts.
What This Actually Is, From a Builder’s Chair
Strip away the Harvard crest and what you have is a familiar architecture pattern: a conversational agent with a synthetic face and voice, dropped into a structured scenario. The avatars provide feedback during practice pitches and board meetings — which means the interesting engineering isn’t the face at all. It’s the evaluation layer behind it.
Anyone who has built a feedback bot knows the face is the easy sell and the hard part is everything upstream:
- Scenario design. A practice board meeting only works if the agent stays in character, tracks what the student said, and pushes back the way a real board member would. That’s prompt architecture and state management, not video rendering.
- Feedback quality. Generic praise is worthless. Useful critique requires the system to compare a student’s pitch against some model of what a good pitch looks like. Whoever tuned that rubric did the real teaching here.
- The persona layer. Making the avatar feel like a specific instructor — their framing, their standards, their style of pushing back — is where these projects live or die. A face without a consistent voice is just an expensive puppet.
HeyGen handles the avatar generation. The pedagogy wrapped around it is where Harvard is earning its $699.
Why the Avatar Layer Matters More Than You’d Think
My first instinct with any avatar product is skepticism — I’ve argued before that a plain chat interface often outperforms a talking head, because the face adds latency and cost without adding intelligence. But pitch practice is one of the rare use cases where the face earns its keep. Pitching is a performance skill. Practicing it against a text box teaches you nothing about holding attention, reading a room, or staying composed while someone looks unimpressed. A rendered board member staring back at you, even an artificial one, exercises muscles that text simply can’t.
That’s the design lesson for bot builders: match the interface to the skill being trained. If the skill is verbal and social, the bot should be too.
The Uncomfortable Question About Scale
There’s a second reason this launch matters, and it’s economic. Instructor feedback has always been the bottleneck of education. One professor, thirty students, limited hours — that math is why elite programs stay small and expensive. An avatar that delivers practice feedback breaks the bottleneck. The live weekly sessions stay human; the repetition — the tenth pitch run-through, the third mock board meeting — goes to the machine.
That’s a sensible split, and it’s the same one I recommend to clients building support or training bots: humans for judgment and connection, agents for reps and availability. Harvard putting its brand on that split is a signal to every institution watching. If the most prestigious business school in the world is comfortable letting synthetic versions of its faculty critique student work, the reputational barrier for everyone else just dropped considerably.
What I’d Watch For
The open question is whether the feedback is actually good — whether these avatars critique like their human counterparts or like a polite autocomplete wearing a professor’s face. That’s not something a press cycle can answer; only students grinding through their fifth practice pitch will know.
But as a proof of concept for avatar-based training agents, this is the highest-profile deployment I’ve seen. If you’re building bots for education, coaching, or any performance skill, study this one closely. The template — human-led live sessions, agent-led
🕒 Published: