The Foundry is Harvard Business School’s proprietary AI-native platform that gives all founders (not just HBS students) in the United States access to Harvard expertise, tools, and a community to build and pitch ventures.
Last year, the team behind Harvard Business School's Foundry noticed something strange in their usage data. Founders on the platform were producing their best work there, then copying it somewhere else.
It made little sense on its face. Founders build their companies from the ground up on Foundry, yet users were lifting their chats and memories out of the platform and pasting them into ChatGPT.
The reason wasn't quality. The reason was connection. ChatGPT was already wired into the rest of their tools through the Model Context Protocol, the open standard known as MCP. It held the rest of their lives.
"That was humbling and clarifying," says Shivesh Sood, Foundry's product lead. "Your most valuable context is your biggest strength. How much other context you can pull in is your next biggest. In 2026, connectors and MCP aren't a differentiator. They're a base expectation."
Sivesh Sood is HBS Foundry's product lead
Shivesh Sood
MORE FOR YOU
Stories like this one are why I sought Shivesh out. The AI industry has an adoption problem it rarely discusses: MIT researchers found that 95% of enterprise generative AI pilots deliver no measurable return, and the gap is rarely the model. Shivesh has spent two years on the other side of that gap, taking four versions of Foundry's AI product to market, each tested by thousands of founders before it shipped wide. With Harvard Business School professor Shikhar Ghosh, he co-authored the two-part "AI Development Guide: Assistants," published by Harvard Business Publishing.
6 Lessons
Our conversation kept returning to one theme: the choices that make busy, skeptical people adopt an AI product and keep coming back. Six lessons stood out.
Know Your User Twice Over
Ask Shivesh what protects an AI product over time and he doesn't mention models, data or patents. He points to how deeply the team understands its users' pain.
"Your real moat is user empathy," he says. "The best AI products are knee-deep in it, rooted in a pain no product out there currently solves."
Foundry earned that depth by testing at increasing scale, from a handful of founders in a room to thousands putting each version through its paces before launch. In-person testing with founders in India surfaced something no dashboard had: impatience. "You can't watch someone's face fall over a survey link," Shivesh says.
Knowing your user in the AI era also means placing them on two curves at once. Every team knows the classic adoption curve, from early adopters to laggards. Shivesh argues a second curve now sits on top of it, running from AI-native users to AI-traditional ones, and the two don't line up. "Not all early adopters are AI-native," he says. "If you design for one curve, you silently lose half the other." The teams that win segment users by AI fluency as deliberately as they segment by industry or role.
The Ask Is Rarely the Need
Foundry's largest research effort, a survey of more than 20,000 founders, found that over 60% named funding as their biggest pain point. The team didn't build a funding feature. They ran the answer through the Five Whys and found the deeper need was storytelling. Founders couldn't articulate their venture in a way that made anyone want to fund it.
"A lot of founders simply don't know what a VC is looking for," Shivesh says. So the team built a pitch simulation against a realistic AI investor that pushes back the way real VCs do. The feature sees 60% repeat usage, and multiple founders have reported winning $10,000 pitch competitions after practicing with it 20 times.
"The ask was funding. The need was storytelling," Shivesh says. "Users hand you their symptoms. The product has to treat the disease."
Nobody's First Instinct Is a Course Anymore
In test after test, the Foundry team saw the same pattern: founders don't want to learn about a business model before building one. They want to learn it through the process of building it with an AI.
A week after watching founders fidget through content in India, the team returned with a different prototype: an AI that worked through a business model Socratically, teaching as they built. Founders understood it far better than when they had read about it alone.
"The first thought isn't 'let me find a course,'" Shivesh says. "It's 'let me have an AI explain it to me step by step while I build.'" The same impatience, he says, is coming for every onboarding flow, help center and training program. Any product that requires users to learn before they can do will lose to one that lets them do both at once.
Build a Co-Founder, Not a Magic Wand
The obvious product for Foundry to build was one-click startup generation: press a button, receive a business plan. The team tested it. The output looked complete. The founders weren't. They had checked a box, and the AI had finished their work for them without teaching them anything.
So the team built the opposite: an AI that behaves like a Socratic co-founder and questions founders' assumptions the way a demanding partner would.
"A founder who outsources their judgment hasn't built a company. They've printed one," Shivesh says. "Our users unanimously tell us the pushback is what they love."
Getting an AI to push back is harder than it sounds, since models drift toward agreement by default. OpenAI rolled back a ChatGPT update last year after it grew so agreeable it praised obviously bad ideas, a reminder that sycophancy is a product failure, not just an annoyance.
The approach paid off. Founders came back saying the AI had changed how they think about their strategy, and in several cases the pushback led them to pivot entirely, with stronger results.
Pool the Context, Then Connect It
Inside teams, Shivesh keeps seeing the same ceiling: companies hand every employee an AI assistant and stop there.
"A personal AI knows your prompts. It doesn't know your business," he says. "Valuable adoption comes from pooling team context into something every person's AI can draw on."
Gartner reached a similar conclusion in 2025, declaring that context engineering has displaced prompt engineering as the determinant of AI quality. The ChatGPT episode that opened this story supplies the corollary: pooled context isn't enough if it's stranded.
AI Can Help Us Be More Human
The lesson Shivesh tells with the most feeling has nothing to do with retention curves.
One Foundry feature helps founders tell their own story around the questions investors actually ask: why this, why now, why me. Several founders, years into their ventures, carried real adversity that had quietly become the drive powering their companies.
"We watched founders cry," Shivesh says. "Not because the AI wrote something for them, but because it helped them see, all at once, how much they had grown." He offers it as a counterweight to the industry's automation reflex. "The products people love aren't the ones that replace them," he says. "They're the ones that reveal them."
Where This Goes
Two predictions came up as we closed. Chat will not kill the interface, because well-designed UI is what generates a user's next question, a point usability researchers call the articulation barrier. And model providers, from OpenAI to Anthropic, will converge into orchestrators that route work across many models rather than selling raw inference.
Strip away the campus setting and the six lessons read as a checklist for any AI product team. Segment users by AI fluency, not just role. Treat every feature request as a symptom. Teach inside the work. Strengthen judgment instead of replacing it. Pool the context you own, then connect to the context you don't. And look for the moments where your AI shows people something true about themselves.
Models are rented, and everyone rents the same ones. What can't be rented is a team's accumulated understanding of its users.
>> Home