The Foundry is Harvard Business School’s unique AI-native platform that provides HBS student and alumni founders around the world with access to Harvard’s expertise, tools, and community to build and market their ventures.
Last year, a team at Harvard Business School’s Foundry noticed something odd about their usage data. Founders on the platform were creating their best work there and copying it elsewhere.
His face meant little. While the founders were building their company from scratch on Foundry, users were pulling their chats and memories off the platform and pasting them into ChatGPT.
The reason was not quality. The reason was connection. ChatGPT was already integrated into the rest of the tools through the Model Context Protocol, an open standard known as MCP. It sustained them for the rest of their lives.
“It was humbling and clarifying at the same time,” says Shivesh Sood, head of product at Foundry. “The most valuable context is the greatest strength. How much other context can be included is the next biggest thing. In 2026, connectors and MCPs are not differentiators; they are fundamental expectations.”
Sivesh Sood is a Product Leader at HBS Foundry
Shivesh Sood
Stories like this are why I sought out Shivesh. The AI industry has implementation issues that are rarely discussed. MIT researchers found that 95% of companies’ generative AI pilots don’t deliver measurable benefits, and that gap is rarely modeled. Shivesh spent two years on the other side of that gap, bringing four versions of Foundry’s AI product to market, each tested by thousands of founders before being widely shipped. He co-authored the two-part AI Development Guide: The Assistant with Harvard Business School professor Shikhar Ghosh, published by Harvard Business Publishing.
6 lessons
Our conversations always returned to one theme. That means busy, skeptical people choosing to adopt AI products and keep coming back. The 6 lessons were impressive.
Understand your users twice
When I asked Shivesh what protects AI products over the long term, he didn’t mention models, data, or patents. He points out how deeply the team understands users’ pain.
“The real moat is user empathy,” he says. “The best AI products are knee-deep and rooted in problems that no product that currently exists can solve.”
Foundry gained its depth by testing at scale, from a few founders in a room to thousands of people testing each version before release. During face-to-face testing with Indian founders, a sense of impatience that was not visible on the dashboard surfaced. “You can’t see someone’s face fall on top of a survey link,” Shivesh says.
Knowing your users in the age of AI also means placing them on two curves at the same time. Every team, from early adopters to laggards, knows the classic adoption curve. Shivesh argues that a second curve, running from AI-native users to AI-traditional users, is currently on top of that, and the two are not aligned. “Not all early adopters are AI natives,” he says. “When you design around one curve, you quietly lose the other half of the curve.” Winning teams segment users by AI fluency as much as they segment by industry or role.
Questions are rarely necessary
Foundry’s largest research effort, a survey of more than 20,000 founders, found that more than 60% cited funding as their biggest pain point. The team did not build a fundraising function. They looked at the five whys to find the answer and found that the deeper need was storytelling. The founders couldn’t articulate their venture in a way that would make anyone want to fund it.
“Many founders have no idea what VCs are looking for,” Shivesh says. So the team built a realistic AI investor pitch simulation that pushes back against real-world VC practices. 60% use this feature repeatedly, and multiple founders report winning a $10,000 pitch contest after practicing this feature 20 times.
“The need was for funding. The need was for storytelling,” Shivesh said. “Users tell you their symptoms. The product must treat the disease.”
No longer is anyone’s first instinct the course
In test after test, the Foundry team saw the same pattern. Founders don’t want to learn about business models before building them. They want to learn it through the process of building with AI.
A week after watching the founders tinker with content in India, the team returned with another prototype. It was an AI that works Socratically with business models and teaches as you build. The founders understood it much better than when they read it alone.
“My first thought was not, ‘Let’s let them find their way,'” Shivesh says. “It’s, ‘Let’s let the AI walk us through it as we build it.'” The same impatience is occurring in every onboarding flow, help center, and training program, he says. Products that require users to learn before learning will lose out to products that allow them to do both at the same time.
Grow co-founders, not magic wands
The obvious product for Foundry to build was one-click startup generation, where you press a button and receive a business plan. The team tested it. The output appeared to be complete. The founders did not. They simply checked a box, and the AI completed the task for them without being taught anything.
So the team built the opposite. It’s an AI that acts like a Socratic co-founder and questions founder assumptions like a demanding partner.
“Founders who outsource their decisions didn’t create the company; they just printed it,” Shivesh said. “Our users unanimously say they love pushback.”
Pushing back against an AI is harder than you might think, since the model defaults to consensus. OpenAI last year rolled back a ChatGPT update that became so positive that it praised an obviously bad idea, but it was a reminder that flattery isn’t just an annoyance, it’s a product failure.
This approach paid off. Founders are coming back to say that AI has changed the way they think about strategy, and in some cases, they say they have completely pivoted in the face of backlash and delivered stronger results.
Pool and connect contexts
Shivesh continues to see the same ceiling within the team. Companies give every employee an AI assistant and stop there.
“Personal AI knows your prompts, but it doesn’t know your business,” he says. “Valuable adoption comes from pooling the context of teams into something that everyone’s AI can leverage.”
Gartner reached a similar conclusion in 2025, declaring that context engineering has replaced prompt engineering as the determinant of AI quality. The ChatGPT episode that opened this article brings home the corollary that pooled context is not enough when you get stuck.
AI helps make us more human
Shivesh’s most emotional lesson has nothing to do with retention curves.
One feature of Foundry is helping founders tell their stories around the questions investors actually ask: why this, why now, why me? Several founders have experienced adversity many years after launching their ventures, but it secretly became the driving force behind the company.
“We saw founders cry,” Shivesh says. “Not because AI wrote anything for them, but because they can now see how much they have grown all at once.” He offers it as an antidote to the industry’s automation reflex. “Products that people love don’t replace people,” he says. “They’re the ones that reveal it.”
where is this going
Two predictions came out at the end. Chat doesn’t take away from the interface. A well-designed UI creates the user’s next question, a point usability researchers refer to as the articulation barrier. And model providers from OpenAI to Anthropic converge on orchestrators that route work across many models rather than selling raw inference.
Strip away the campus settings and read the six lessons as a checklist for AI product teams. Segment users by AI fluency, not just role. Treat all feature requests as symptoms. I teach at work. Enhance your judgment, not replace it. Pool the contexts you own and connect to contexts you don’t own. And look for moments when AI shows people the truth about themselves.
The models are rented and everyone rents the same model. What you can’t rent is the understanding your team has built up about your users.
