Harvard Business School built its brand on scarcity — limited seats, limited access, limited faculty time. Its new entrepreneur bootcamp sells something close to the opposite: AI avatars of its instructors, available for practice sessions, bundled into a $699 program. The institution that rations human attention is now, in a sense, mass-producing it.
The program in question is HBS Foundry, an eight-week bootcamp for entrepreneurs. It includes live weekly sessions with actual instructors, alongside virtual simulations in which AI avatars — built by the startup HeyGen — provide feedback during practice pitches and board meetings. As someone who studies agent architecture for a living, I find the avatar less interesting than what has to sit behind it. So let’s take that apart.
An Avatar Is a Rendering Layer, Not a Mind
The public conversation about this launch will fixate on the face — the uncanny professor lookalike nodding at your pitch. But from a systems perspective, the avatar is the least consequential component. A photorealistic talking head is a presentation layer. The hard problem is everything upstream of it.
For an avatar to give feedback on a practice pitch or a simulated board meeting, some system has to do at least three things:
- Perceive — parse what the student actually said, including structure, claims, and gaps in the argument.
- Evaluate — compare that input against some model of what good looks like, presumably shaped by the instructor whose likeness is on screen.
- Respond — generate feedback that is specific, pedagogically useful, and consistent with that instructor’s actual judgment.
Each of those steps is a genuine open problem in agent design. The rendered face is just the last mile.
The Fidelity Question Nobody Can Answer From Outside
Harvard has not published — at least not in what’s publicly available — how faithfully these avatars reproduce the judgment of the humans they resemble. And that is the question that matters. There is an enormous difference between an avatar that speaks in an instructor’s voice and an agent that evaluates like that instructor thinks.
A general-purpose model wearing a professor’s face gives you generic feedback with prestigious branding. A system genuinely trained on or steered by an individual instructor’s evaluative criteria would be something more interesting: a distilled, queryable version of professional judgment. From the outside, we cannot tell which of these HBS Foundry is delivering. Students probably can’t either — which is precisely the concern. When feedback arrives wrapped in an authoritative face, learners are less likely to interrogate its quality.
Why the Hybrid Structure Is the Smart Part
Credit where due: the program’s architecture hedges intelligently. The live weekly sessions with real instructors remain the backbone. The avatars occupy the practice loop — pitches, simulated board meetings — where the cost of an imperfect response is low and the value of repetition is high.
This is, structurally, the right place to deploy current-generation agents. Practice environments are forgiving. A student rehearsing a pitch at midnight benefits from any reasonably coherent feedback, and the human instructor remains in the loop to correct course during live sessions. It’s a sensible division of labor: humans for judgment that carries stakes, agents for volume that carries none.
What Gets Commoditized, and What Doesn’t
The deeper story is economic. A professor’s time is the scarcest resource in education, and this program treats a simulated version of it as effectively free at the margin. If avatar-based feedback works even moderately well, the implication extends far beyond one bootcamp: the rehearsal portion of professional education — the repetitions, the practice reps, the low-stakes feedback cycles — becomes cheap and infinitely available.
What doesn’t get comm
🕒 Published: