Good News For Startups: Enterprise Is Bad At AI
YC partners Garry Tan, Harj Taggar, Diana Hu, and Jared Friedman argue the MIT ‘95% enterprise AI failure’ stat proves startups have an unprecedented opening, not that AI doesn’t work.
- MIT study shows 2/3 of enterprise AI projects were built in-house or with consultants; outside-vendor projects had materially higher success rates.
- Enterprise engineering teams often don’t use AI tools themselves and actively want the ‘it’s all hype’ narrative to be true, which is why they can’t ship.
- Reducto closed a FAANG company 154 days after YC batch by outcompeting the company’s own internal team that had tried open-source and AWS Textract for years.
- Tactile built a real-time KYC/AML REST API for banks in months; Citi and JPMorgan each spent 3–5 years and tens of millions building equivalent internal systems.
- Castle AI won bank deals by beating incumbent vendors who added AI as a bolt-on rather than building AI-native from the start.
- A $5B financial services CIO said directly: once they invest in training a system, switching costs become prohibitive — that’s the moat for AI startups.
- Best enterprise champion archetype: a big-company employee who dreamed of doing a startup but never did — they’ll fight for you internally to live vicariously.
- Karpathy’s ‘agents are overhyped’ interview was misread; his actual point was agents need proper data, context, and evals — which is a large build opportunity for startups.
2025-10-30 · Watch on YouTube
