Harvard Cloned Its Faculty and Charged $699 for the Privilege

A flight simulator does not teach you to fly. It teaches you to survive the first ten minutes of something going wrong, so that when the real thing happens, your hands already know where to go. Nobody confuses the simulator with the sky. The value is precisely in the fact that it is not the sky.

That framing is the most useful lens I have found for Harvard Business School’s new $699 bootcamp, part of its HBS Foundry accelerator, where AI avatars of the school’s own faculty give feedback while participants practice pitches and board meetings. The headline reads like a story about credential erosion. The actual engineering question is narrower and more interesting: what kind of agent architecture makes a simulated critic useful, and where does that architecture quietly fail?

Feedback is a scarce resource, and it does not scale

Start with the constraint the program is clearly built around. A faculty member’s attention is the least divisible asset in higher education. You can record a lecture once and serve it a million times. You cannot record a critique of a pitch nobody has given yet. Feedback is inherently reactive, which is why it has stayed expensive while content became free.

What an avatar-based agent attempts is to make the reactive part reproducible. The instructor’s judgment gets encoded well enough that it can respond to novel input — your specific slide, your specific hand-waving about unit economics — without the instructor being present. That is a genuinely different technical problem from content delivery, and it is worth taking seriously as an architecture rather than dismissing as a gimmick.

Where the hard part actually lives

The avatar is the least significant component here. Face and voice synthesis is well-trodden ground, and the visual likeness is closer to packaging than to capability. Strip it away and you still have the same system underneath.

The hard part sits in three places, none of which a rendered face solves.

• Evaluation criteria. An agent giving pitch feedback needs some representation of what a good pitch is — implicitly through training signal, explicitly through rubrics and prompting, or some hybrid. Whatever form it takes, that representation is a compression of the instructor’s judgment, and compression discards things. The question is what got discarded.
• Grounding in the user’s actual artifact. Generic advice is cheap. Feedback that references what you specifically said, in the order you said it, and identifies the moment the logic broke, is expensive. This is a context and retrieval problem before it is a generation problem.
• Calibration on unfamiliar cases. A human instructor who encounters a business model they do not understand will usually say so. A language model tends to produce fluent commentary regardless. The failure mode of a simulated critic is not silence, it is confident irrelevance.

The likeness question

The choice to render specific faculty rather than a generic coach is a design decision with real consequences, and I suspect it cuts both ways.

On the useful side, a named critic creates stakes. Practicing in front of something that resembles a particular person, with a particular reputation for asking hard questions, likely produces a different rehearsal quality than typing into a chat box. Simulators work partly because they induce the physiological conditions of the real event.

On the risky side, likeness borrows trust the system may not have earned. When feedback arrives from a face associated with genuine expertise, a user has fewer natural prompts to ask whether the reasoning holds up. The interface signals authority while the model underneath is doing statistical work on a compressed proxy of that authority. That gap is where I would focus scrutiny.

What I would want to know

The details of the implementation are not public, so I am reasoning about the class of system rather than this specific build. Still, the evaluation questions are the same ones I would ask of any agent deployed as a judge.

Does the agent’s feedback correlate with what the actual instructor would say on the same input? Does it degrade gracefully on domains outside its training distribution, or does it keep talking? Can a participant tell the difference between a well-grounded critique and a plausible-sounding one? Those are measurable properties, and they matter more than fidelity of the rendered mouth.

Simulators are good, on one condition

My reading is that this is a reasonable use of agent technology and a slightly unreasonable interface for it. Practice against a synthetic critic is a solid idea. Repetition against a consistent evaluator builds the reflexes that make real conversations survivable. That is real pedagogical value at a price point far below traditional access to the same faculty.

The condition is that everyone involved keeps the flight simulator distinction intact. The simulator is valuable because it is a model of the sky, cheap to enter and safe to fail in. The moment a rendered face convinces a participant it is the sky itself, the tool stops teaching and starts misleading. That boundary is maintained through design and framing, not through capability improvements, which is exactly why it deserves more attention than the avatar rendering ever will.

🕒 Published: August 24, 2026