What exactly are you paying for when you buy a course from a professor who never shows up?
That question sits at the center of HBS Foundry, Harvard Business School’s $699 bootcamp for entrepreneurs. The program puts AI-generated replicas of its instructors in front of participants, and those avatars give feedback while people practice pitches and run mock board meetings. Real faculty, digitally duplicated, available on demand.
I build bots for a living, so my first reaction was not outrage. It was curiosity about the architecture. Because whatever you think of the price tag, someone had to actually make this thing work, and the design choices behind it are more interesting than the headline.
The problem this actually solves
Practice feedback is the least scalable thing in education. A pitch coach can sit through maybe eight or ten practice runs in a day before their judgment turns to mush. Meanwhile the thing that makes founders better at pitching is repetition with correction, ideally dozens of reps, ideally at 11pm the night before the real meeting.
That gap is genuinely bot-shaped. You are not asking the system to invent new knowledge. You are asking it to apply a known evaluation framework to a fresh input, over and over, without getting tired or bored. That is the sweet spot for this kind of tool, and it explains why Harvard reached for avatars instead of just recording more video lectures.
A recorded lecture is broadcast. An avatar that reacts to your specific pitch is a loop. The loop is the product.
What I would be building under the hood
Strip away the face and the voice, and a practice-pitch avatar is a fairly recognizable system. From my own work on similar setups, the pieces that matter are:
- A capture layer that takes your spoken pitch and turns it into text plus signal you cannot get from text alone — pacing, filler words, how long you talked before saying what the company does.
- An evaluation rubric that encodes what the instructor actually cares about. This is the part people underestimate. Without an explicit rubric, a language model will happily tell you your pitch was “compelling and well-structured” forever.
- A persona layer that shapes tone and priorities so the feedback sounds like a specific person with specific opinions, not a committee.
- A presentation layer — the video and voice synthesis that makes it feel like a conversation rather than a form submission.
The presentation layer gets all the attention and does the least work. The rubric does the heavy lifting. If you are building anything in this shape, spend your time there.
The persona question nobody wants to answer
Here is where I get uneasy, and it has nothing to do with the technology being bad.
When you put a named professor’s face on a feedback engine, you are making an implicit promise: this response reflects how that person thinks. That promise is very hard to keep. A persona layer can capture someone’s vocabulary, their pet peeves, their preferred framing. It cannot capture the thing that makes a good mentor valuable, which is the ability to notice the one weird detail in your business that breaks the standard advice.
Good advisors know when their own framework does not apply. A replica trained on that framework will apply it confidently to every case.
So the honest framing is that these avatars are rubric-driven practice partners wearing a familiar face. That is a useful product. It is not the same product as access to the person, and the face makes it easy to blur the two.
What builders should steal from this
If you run any kind of training, coaching, or onboarding, the pattern here is worth copying at a much smaller scale.
Pick one repetitive feedback task where a human currently reviews the same kind of work again and again. Write down the rubric that human uses, in painful detail, including the things they never say out loud. Build a bot that applies it. Skip the video avatar entirely on your first pass — text feedback with a clear persona gets you most of the value for a fraction of the effort.
Then watch for the failure mode: the moment your bot starts giving advice that is technically consistent with the rubric and obviously wrong for the situation. That boundary is where you learn what actually needs a person.
The price is the real story
$699 for practice reps with a synthetic instructor is not expensive by executive education standards. It is also not cheap for a system whose marginal cost per session is close to nothing. That spread is the business model, and I expect a lot of institutions to notice it.
Which means the interesting competition ahead is not between universities. It is between anyone who can encode a genuinely good rubric and everyone who just bought a video synthesis license. The rubric is the moat. The face is packaging.
🕒 Published: