EO StudioEveryone Builds, Ships, and Sells. Winners Do It Differently. | Kimberly Tan, a16z Investing Partner
At a glance
WHAT IT’S REALLY ABOUT
Enterprise AI winners win on customer proximity, ROI, and deployment rigor
- Kimberly Tan explains why early “GPT wrapper” critiques underestimate the engineering, workflow, and deployment work required to turn base models into enterprise-ready AI products.
- She argues that vertical/applied AI companies build durable moats by learning industry-specific rules and tacit operational knowledge directly from customers, often through forward-deployed, onsite implementation.
- Tan highlights that AI’s non-determinism makes production performance and guardrails more important than impressive demos, shifting how products must be built and evaluated.
- She stresses that enterprise adoption depends on selling quantifiable ROI via well-designed pilots, citing customer support automation (e.g., Decagon) as a category with unusually clear metrics.
- The conversation also covers how automation is constrained by cultural and regulatory realities, and how patient, action-oriented investor support helps founders navigate inevitable volatility.
IDEAS WORTH REMEMBERING
5 ideas“GPT wrapper” misses the real work: the last mile is the product.
Tan argues that enterprise value requires substantial work between the foundation model and the customer environment—data integration, workflow mapping, guardrails, and change management—so founders should treat “implementation” as core product, not a service afterthought.
Build your moat at the customer’s desk with deep onsite discovery.
Because enterprises hold critical tacit knowledge in people’s heads, early teams gain disproportionate leverage by sitting with users, learning the real workflow, and encoding industry-specific logic into the product.
In AI, production proof matters more than polished demos.
AI behavior is probabilistic; a demo can hide tail-risk failures that show up in production, so investors and buyers should overweight live deployments, monitoring, and real operating metrics.
Winners orchestrate many models; they don’t just call one.
Rather than defaulting to the “best” model, teams should route tasks across multiple models to balance accuracy, latency, and cost, using orchestration/chaining as a competitive differentiator.
Don’t sell AI—sell a measurable ROI outcome from a fast pilot.
To get bought, enterprise AI pilots must be designed to reach production quickly and culminate in a crisp metric (e.g., CSAT, resolution rate, ticket volume, cost), not vague “AI transformation.”
WORDS WORTH SAVING
5 quotesI think what people don't understand is that it's very difficult to work with these models. The capabilities out of the box are not the same as the capabilities needed to actually show value to an end enterprise buyer.
— Kimberly Tan
I highly encourage early-stage founders who are building an enterprise AI to fly to their customer, to sit next to them, to really understand to the depth that you can.
— Kimberly Tan
One of my, like, very strong beliefs about investing in AI is that there's a huge gap between a fancy demo and real production application.
— Kimberly Tan
AI is non-deterministic by nature, and so knowing something that ninety-five percent of the time might do this thing, but five percent of the time might do something totally different, it's just a totally different paradigm of building than enterprise software, which is a hundred percent deterministic.
— Kimberly Tan
You can really build an excellent solution, but you still need to find someone who will buy it, and I think a lot of people who are technologists, they understand the value of the technology, and they understand why it's amazing and why there's gonna be ROI. But that doesn't matter from a business standpoint if you can't explain it to a customer, and they can't see the value, and you can't implement it in a way where they are happy about the product.
— Kimberly Tan
High quality AI-generated summary created from speaker-labeled transcript.
