Jared says viral posts claiming that 95% of AI projects fail misrepresent the study. He argues that the findings reflect familiar problems with enterprise software and help explain which approaches can work.
Good News For Startups: Enterprise Is Bad At AI
Speakers
Jared, a Y Combinator guest, discusses why enterprise AI projects often fail and how startups can win enterprise customers with the Lightcone hosts and Nana.
Jared points to weak internal systems, siloed teams, and consulting projects that gather requirements but may lack the software expertise to build a good product. He compares the result to a committee-designed “camel” and notes that even Apple’s Calendar app has recurring bugs.
The hosts describe Taktile building a real-time decision API for banks after similar internal efforts reportedly took three to five years and tens of millions of dollars. Greenlight won work after Ernst & Young spent a year on an AI system that did not work, while Castle AI competed against established bank vendors with older systems.
Nana describes Reduct winning a document-processing deal after a large company’s internal tools and OCR options fell short. She says the founders built a close relationship with their champion and navigated internal politics; the hosts also recommend finding champions through founders whose companies were acquired by large firms.
The hosts say enterprise demand is strong, but some engineering teams remain skeptical of AI and do not use code-generation tools, limiting their ability to build products. They urge engineers to try the tools on real projects and argue that AI-native systems and trained workflows can create opportunities and switching costs.
Takeaways
- Test AI tools on a real project and give yourself time to learn their limits.
- Build around a specific business process and integrate deeply with the systems people already use.
- Pair technical skill with product judgment and a close understanding of users’ work.
- Find a committed internal champion and learn how the organization makes decisions.
- Treat enterprise implementation effort and trained workflows as potential sources of customer retention.
Mentioned in this talk
Search inside this talk
What was claimed 11
- featureEnterprise integration: Deep integration into business processes and systems of record can lead to large enterprise deals, but takes time.
“when you do succeed and plug into the systems of record, the pot of gold is actually quite big, though it takes a long time.”
YouTube Unverified1 source· Talk
- metricTaktile: The software handles millions of decisions per day at scale.
“It does that millions of times per day at scale.”
YouTube Unverified1 source· Talk
- featureTaktile: Built a REST API for real-time decisions that can use the latest AI models.
“They did all that for a fraction of the budget and in much less time.”
YouTube Unverified1 source· Talk
- featureTaktile: Provides a real-time business decision engine for banks, including KYC and AML checks.
“In real time, it can help them handle KYC and AML checks”
YouTube Unverified1 source· Talk
- metricMIT study on AI projects: Two-thirds of surveyed projects were built internally or with consulting help; one-third involved buying from an outside company.
“two-thirds were projects where the enterprise built software internally or did it with help from a consulting agency”
YouTube Unverified1 source· Talk
- featureGreenlight: Provides an AI system to a bank; the system was fully deployed and working after Ernst & Young’s attempt failed.
“Greenlight now has its system fully deployed at the bank, and it’s working.”
YouTube Unverified1 source· Talk
- featureCastle AI: Builds an AI mortgage servicing system.
“They’re basically building an AI mortgage servicer.”
YouTube Unverified1 source· Talk
- metric
Reducto: Closed a large fan company 154 days after the batch.“They just announced their Series B recently, and they actually closed a fan company 154 days after the batch”
YouTube Unverified1 source· Talk
