ÌÇÐÄÊÓƵ

As a partner at , invests in the AI application layer and B2B SaaS.

She began on the consumer side at early-stage companies including and pioneered the early push of CPG e-commerce at long before the acquisition. After a pivot into tech M&A at and a partnership at GGV Capital (now ), she joined NEA roughly three years ago.

Today, Luck’s thesis centers on the AI application layer, betting on vertical AI and the “last mile” of automation to bridge the gap between horizontal model potential and tangible enterprise ROI.

We recently spoke with Luck about the increasing relevance of vertical AI, how startups can carve out durable advantages in a world dominated by platform giants and more.

The interview has been edited for brevity and clarity.

Crunchbase News: You’ve seen the evolution of commerce from the early days of Amazon Fresh to the current AI boom. How does that background influence how you view the “friction” of AI adoption today?

Luck: I see incredible parallels. At Amazon, I spent my days convincing CPG manufacturers that e-commerce was the inevitable future. Back then, there was immense friction — technological, logistical and mental. Today, 500 companies are in a similar spot with AI.

While the potential is obvious, most organizations are still struggling to integrate it into their daily workflows. We are moving from a world where AI is a “shiny object” to one where it must solve a mechanical problem, but getting there requires overcoming that initial resistance.

You’ve talked about the “ question” — the fear that frontier models will eventually swallow the application layer. How can vertical startups build a durable defense?

It’s the primary concern for founders right now. Most horizontal tools, like Claude, currently act as research co-pilots. They are excellent at taking a user from 0% to 80%, but they don’t handle the “last mile.” For 99% of people, AI isn’t yet running in the background while their hands are off the keyboard.

Moats are being built by solving the specific hardships of that last mile. Take financial planning and analysis: You can plug data into a general model and ask questions, but the model won’t automatically re-forecast, flag specific trade-offs between burn rate and growth, or create a unified data layer across disparate sources.

Startups that build these purpose-built product flywheels — and use forward-deployed engineers to sit alongside users and identify workflow holes — build a moat that general models can’t easily replicate through scale alone.

Why is owning the end-to-end workflow becoming more valuable than the underlying model differentiation?

Because it removes the mental friction of “What do I do with this?” If a company can deliver a finished work product — an artifact — the ROI is undeniable.

Our portfolio company does this for legal due diligence. Another, , does this for equity research reports. When the output is a discrete document that looks exactly like (or better than) what a team of analysts would produce, the enterprise doesn’t care which model is under the hood. They care about the hours saved and the accuracy of the result.

You mentioned the idea of the “operating system” changing. How should startups think about partnering versus competing with platforms like Claude or ?

We haven’t truly seen our way of working change yet; we’re still using the same UIs. But I expect a shift in which a model becomes your de facto operating system — a command center from which you “call” other specialized applications.

Think back to five years ago, when startups used as their primary interface. You might see a future where a specialized tool like Samaya is integrated directly into a horizontal model’s UI.

The specialized knowledge graph and proprietary data remain with the startup, but execution occurs within the user’s primary “operating system.” Interoperability will be the next big frontier at the application layer.

In regulated industries, what is the “make or break” factor for an enterprise buyer right now?

It’s a mix of accuracy, auditability and cybersecurity. Enterprises are terrified of “data provenance” issues — they need to be able to audit the trail of every number.

I’m closely following (Artificial Intelligence Underwriting Co.). They are essentially building a “Moody’s for AI agents.” They’ve assembled a group of over 100 CISOs to create a for-profit certification standard. If a company like or can certify its agents against this standard, it gives the enterprise a layer of trust that a standard SOC 2 just doesn’t cover yet.

With AI-fueled hackers creating new attack vectors, this kind of authentication is becoming a necessity.

What does the “next frontier” look like for you as an investor?

We are in the “pre-mobile-native” era. We’ve moved the web to the phone, but we haven’t seen the apps that only AI can enable. I’m waiting for the “ moment” — the transition from a co-pilot, where you’re still “driving,” to a truly autonomous, agentic workflow. Whether it’s through voice-first interfaces or agent-to-agent interaction, the next 12 months will likely reveal the first truly novel ways of working that feel fundamentally different from the laptop-and-keyboard era.

Related Crunchbase query:

Illustration: