Harvard Business School’s startup bootcamp now lets students practice their pitch on an AI avatar of a venture capitalist. The program, called HBS Foundry, costs $699, runs eight weeks, and uses avatar technology from HeyGen to deliver feedback during practice pitches and simulated board meetings. It’s one of the more visible bets yet on ai career coaching as a scalable stand-in for human mentorship, and it won’t be the last. Career apps, LinkedIn’s own AI features, and a wave of interview-prep startups are all racing to put an AI coach in front of anyone with a job search and a subscription. HBS Foundry is just the version with a Harvard name attached, which makes it a useful case study in what that substitution actually gets you, and what it doesn’t.

According to TechCrunch, HBS Foundry still runs live weekly sessions with real instructors. The AI avatars handle something narrower: the one-on-one feedback work, standing in for professors and mentors during rehearsed pitches and mock board meetings. Katharina Rings, the program’s project director, told the New York Times she originally imagined the AI layer as something closer to a chatbot. Students wanted more than that, so HBS built out a more guided, avatar-based experience instead. The fact that Harvard still pays for live human instruction every week, on top of the AI layer, is itself a hint about how much weight the avatars are actually meant to carry.

What Harvard’s AI avatars actually do

The clearest test case comes from New York Times reporter Sarah Kessler, who tried the program herself and pitched an AI-generated copy of Flybridge Capital co-founder Jeff Bussgang. Her business idea was “Uber for bananas.” Both the real Bussgang and his digital copy were unimpressed. Kessler reported that the avatar version delivered its skepticism with a frozen smile the whole time it talked.

Bussgang is candid about the tradeoff. He told Kessler his AI double is “a little creepy,” then added the line that explains why HBS built this in the first place: “My students love it.” That reaction makes sense. An AI avatar is available at 11pm before a pitch deadline. It doesn’t get tired of hearing the same 90-second pitch for the fifth time. It scales in a way that Bussgang’s actual calendar cannot, and for founders who just need reps, that’s a genuinely useful feature.

None of that is nothing. Repetition helps. Getting comfortable saying your pitch out loud, again and again, until it stops sounding memorized, is a real skill, and a bot that never gets bored is a decent tool for building it. But notice what the practice actually produces: a better-rehearsed founder, not a founder anyone new has met. Kessler walked away from her session with sharper delivery. She didn’t walk away with a warmer relationship to Flybridge Capital, because there wasn’t a person on the other end to build one with.

Why ai career coaching feels like coaching but isn’t

Here’s the distinction that’s easy to miss if you only look at what the avatar does in the room: practicing a pitch with an AI version of Jeff Bussgang teaches you to pitch. It does not put you in a relationship with Jeff Bussgang.

The real Bussgang has a network, a reputation, and the standing to make an introduction or write a check. His avatar has none of that. It can mimic his feedback style and maybe even his tone, but it can’t vouch for a founder to another investor. It can’t remember a specific student six months later when a portfolio company needs a hire. It can’t pick up the phone. That’s not a limitation of this particular product. It’s structural. A simulation of a mentor is a rehearsal tool. A real conversation with a real professional is the beginning of a relationship, and relationships are what move careers, hiring decisions, and funding checks.

This is the same gap that shows up everywhere ai career coaching gets marketed as a stand-in for human contact, not just at HBS. AI mock interviewers, AI networking chatbots, an ai mentor for career advice that lives in an app, they can all sharpen your delivery. None of them can introduce you to anyone. They rehearse the performance without creating the connection the performance is supposed to lead to.

Picture two job seekers preparing for the same interview. One spends the week running practice questions through an AI coach until the answers are polished and quick. The other spends the same week having a real fifteen-minute call with someone who works at the company, one of the field’s classic weak ties, and walks away with the name of the actual hiring manager. Both show up to the interview prepared. Only one of them has already been talked about inside the building before they walk in the door.

The data on what actually moves careers

If the case against AI avatars sounds like an argument for old-fashioned networking, there’s real research behind it, and it’s more specific than “networking matters.”

A team from LinkedIn, MIT, Stanford, and Harvard Business School ran a five-year, large-scale experiment on LinkedIn’s “People You May Know” algorithm, covering 20 million users and roughly 600,000 new jobs. The study, published in the journal Science and detailed in a 2022 MIT Sloan write-up, randomly assigned some users to receive more weak-tie connection recommendations and others more strong-tie recommendations, then tracked who actually changed jobs. Weak ties, meaning the loose, infrequent connections you don’t talk to often, drove more job mobility than close contacts did. MIT Sloan professor Sinan Aral, one of the study’s authors, put it plainly: “Weak ties on social networks can be an extremely useful part of managing your career, promotions, advancement, and even wages.”

The researchers found the relationship wasn’t a straight line. It wasn’t “the weaker, the better.” Moderately weak ties, people you know but aren’t close with, produced the most job mobility, more than either your closest contacts or near-strangers. And the effect was strongest in digital, high-tech, and AI-adjacent industries, exactly the fields where you’d expect passive tools to have already replaced human contact, and exactly where the study found they hadn’t.

What makes a weak tie useful isn’t rehearsal. It’s that the person is real, remembers you, and can pass along information or an introduction you’d never find on your own. That’s also the logic behind asking a near-stranger for a coffee chat: the value isn’t in how polished your ask is, it’s in reaching an actual person who has context you don’t. An AI avatar, no matter how well it mimics a real mentor’s feedback, isn’t a weak tie. It isn’t a tie at all. It doesn’t know anyone. It can’t be surprised into thinking of you when a job opens up three months from now, because it isn’t out in the world having its own conversations.

Practicing a pitch is not the same as building a relationship

Put the HBS Foundry story next to the weak-ties research and the pattern comes into focus. Does AI networking work? As a rehearsal tool, sure, within limits. As a replacement for a real professional relationship, no, because rehearsal and relationship-building solve different problems.

Human networking vs AI tools isn’t really a fair fight, because they’re not doing the same job. An AI avatar can help you tighten your pitch, catch a weak transition, or practice answering a hard question under simulated pressure. That’s useful prep, and it’s worth doing. But prep isn’t the goal. The goal is getting in front of the right person, and only a real person can decide to introduce you to someone else, refer you internally, or remember your name when a role opens up. Bussgang’s avatar can critique a pitch about Uber for bananas all day. Only the real Bussgang can decide whether to fund it, or point the founder toward someone who might.

This is exactly why the HBS Foundry story works as a case study, not just a curiosity. The program pairs AI avatars with the thing the marketing pitch tends to underplay: real, live weekly sessions with actual instructors. Harvard isn’t betting the whole program on simulation. The avatars handle volume and repetition; the humans still show up for the parts that require judgment, memory, and standing. Even the people building these tools seem to know the simulation only goes so far.

The lesson generalizes past pitch practice. Passive tools of every kind, whether that’s an AI mock interviewer, a job board application, or an AI-generated networking message blasted to a hundred strangers, produce the feeling of progress without the substance of it. You did something. You didn’t build anything. Any ai avatars career development tool can help you rehearse what to say. None of them can say it to the person who’s hiring.

Direct outreach works precisely because it skips the simulation and goes straight to the relationship. Reaching out to an actual hiring manager, with something specific to say about their team or their work, does what an AI avatar structurally cannot: it puts a real person in a position to remember you, refer you, or decide to hire you. That’s not a soft-skills argument. It’s the same mechanism the weak-ties research describes, just applied one conversation at a time instead of across a five-year LinkedIn experiment. The same logic is why asking for an informational interview tends to outperform sending another application: fifteen minutes with an actual person beats a form nobody reads.

Simulated practice still has a place. Rehearse your pitch on an avatar, run a mock interview with a chatbot, workshop your talking points however you want. Just don’t confuse the rehearsal for the thing it’s supposed to lead to. The weak-ties research and the HBS Foundry rollout point the same direction: information and opportunity move through real people, not through anyone’s digital copy.

Finding the right person to reach out to, and figuring out what to say to them, is the part most job seekers get stuck on. That’s the gap angld.AI is built to close: paste in a job posting, and it identifies the actual hiring manager, researches their background, and drafts a personalized outreach message in under a minute. It doesn’t rehearse the conversation for you. It gets you into the real one. Start at angld.AI.