# Good News For Startups: Enterprise Is Bad At AI Source: https://www.youtube.com/watch?v=DULfEcPR0Gc Recap page: https://rapidrecap.app/video/DULfEcPR0Gc Generated: 2025-11-11T14:40:48.878+00:00 --- ## Quick Overview The main takeaway from the discussion is that the high reported failure rate of 95% for generative AI pilots, as suggested by an MIT study, is misleading because many enterprises lack the in-house technical skill or the necessary business/human context to successfully implement AI, often leading to reliance on consultants or internal IT teams ill-equipped for advanced AI deployment. **Key Points:** - An MIT study suggesting 95% of generative AI pilots fail is widely circulated on social media, but the panelists argue this statistic is misleading. - The failure is attributed more to a "listening problem" and a lack of understanding of use cases, talent gaps, and internal processes rather than the technology itself. - Companies often fail because they try to build complex AI systems internally when they lack the necessary expertise, leading to reliance on external consultants or siloed IT teams. - Successful deployments, like those by Kastle and Greenlite, involve integrating AI deeply into business processes and having founders with strong technical and domain knowledge. - Garry Tan noted that companies like Apple, despite having infinite resources, cannot build a simple yet highly functional product like the Calendar app, illustrating that building great software is inherently hard. - The panelists suggest that enterprises often fail by trying to copy superficial aspects of successful AI companies (like dressing up or copying homepages) instead of focusing on core engineering and context. - The success rate is much higher when companies buy products from vendors who have already solved the implementation complexities, as seen with Kastle's success with a bank customer. ![Screenshot at 0:00: Harj Taggar, identified as Managing Partner at Y Combinator, begins the discussion by questioning the premise that AI tools built by engineering teams who don't fully believe in AI will succeed.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-00-00.png) **Context:** The video features a panel discussion from Y Combinator's 'Lightcone Podcast' involving Diana Hu, Harj Taggar, Garry Tan (President & CEO of Y Combinator), and Jared Friedman. They analyze the widely reported MIT finding that 95% of enterprise generative AI pilots fail, debating whether this points to a flaw in the AI technology or deeper implementation and contextual issues within large organizations. ## Detailed Analysis The discussion centers on why an MIT study found that 95% of generative AI pilots in enterprises fail, with the panelists largely concluding the problem lies with the enterprises, not the AI technology itself. Harj Taggar pointed out that if the engineering team building the AI doesn't truly believe in it, the product will fail, and if enterprises rely on external consultants who lack deep internal context, they also face difficulties. Garry Tan emphasized that great software engineering is inherently difficult, citing Apple's calendar app as an example of something even resource-rich companies struggle to perfect. Diana Hu contrasted this with startups like Kastle and Greenlite, which focus on integrating AI deeply into specific business processes (like mortgage lending compliance or financial crime detection) and thus achieve success. She argued that many large enterprises, accustomed to the 'plug-and-play' model of standard SaaS, are ill-equipped to handle the complexity of truly embedding AI, often leading to projects that are either too superficial (just slapping AI on top of old systems) or too reliant on external vendors who don't understand the internal workflow nuances. The consensus is that successful AI adoption requires deep domain expertise and a cultural shift, not just technical prowess. ### Critique of MIT Study - Harj Taggar questions the premise of AI pilot failure - Garry Tan emphasizes that copying superficial aspects of successful AI firms is misguided - The 95% failure rate suggests a fundamental disconnect between AI capability and enterprise readiness. ### Enterprise Implementation Challenges - Diana Hu notes that enterprises often lack the necessary human context and deep domain expertise to integrate AI effectively - Companies are used to plug-and-play SaaS, not deep system rewrites required for true AI impact. ### Success Stories (Kastle & Greenlite) - Kastle built a custom AI engine for mortgage lending and servicing, succeeding where internal builds failed - Greenlite uses AI agents for compliance (AML/KYC), showing success comes from deep process integration. ### The 'Mote' Analogy - Garry Tan argues that superior engineering talent and deep domain knowledge (like Apple's) create a moat, which is hard for others to replicate or simply buy off the shelf. ![Screenshot at 0:00: Harj Taggar and Diana Hu setting up the discussion about AI pilot failures at Y Combinator.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-00-00.png) ![Screenshot at 0:35: Garry Tan speaking, who later highlights that executives are being 'horribly triggered' by the study's findings.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-00-35.png) ![Screenshot at 1:11: Visual display of the MIT report cover: 'The GenAI Divide: STATE OF AI IN BUSINESS 2025'.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-01-11.png) ![Screenshot at 1:18: Jared Friedman discusses the viral tweets misinterpreting the study, suggesting the issue is a 'listening problem' \(human context\) not a tech problem.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-01-18.png) ![Screenshot at 2:35: Diana Hu elaborates on how startups like Kastle succeed by embedding AI deeply into business processes rather than superficial application.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-02-35.png) ![Screenshot at 3:54: Garry Tan humorously criticizes Apple's Calendar app as an example of how hard even simple, high-quality software is to build.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-03-54.png) ![Screenshot at 6:31: Screenshot of Reducto's website, an example of a company selling AI for document processing.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-06-31.png) ![Screenshot at 9:40: Screenshot of Kastle AI's website, showcasing 'Purpose-built AI agents for mortgage lending and servicing'.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-09-40.png) ![Screenshot at 10:54: Garry Tan encourages founders to apply to YC now, noting the acceptance rate is under 1%.](https://ss.rapidrecap.app/screens/DULfEcPR0Gc/00-10-54.png)