Field notes

They called it a wrapper. The application layer was the business.

By Liam Webb · July 6, 2026

Harvey got dismissed as just a wrapper on GPT. Its CEO just walked Sequoia through why the opposite is true: the value in AI lives at the application layer, in the messy real work, the trust, and keeping the models swappable underneath. That is the exact bet we made building SmartAlex.

They called it a wrapper. The application layer was the business.

I watched a Sequoia interview this week with Winston Weinberg, the co-founder and CEO of Harvey. If you have not come across it, Harvey is the company that put AI to work inside some of the biggest law firms in the world. I watched it because Winston spent an hour describing, almost word for word, the bet we made when we built SmartAlex. He just made it in law instead of on the phone.

The insult that turned out to be wrong

When companies like Harvey first appeared, the smart money had a name for them: wrappers. Just a thin layer of prompts on top of someone else's model, no real business, no moat, gone the moment the model got better. It was the fashionable take of 2023, and it was wrong. Winston's answer is the whole point. These industries are messy. The gap between what a raw model does and what a law firm actually needs is enormous, and closing that gap is the business. If all Harvey ever did was sit citations on top of a model, then yes, it is a wrapper and it goes away. The ambition to actually transform an industry is what makes it something else.

The moat was never the model

Here is the line that stuck with me. The hard part, Winston said, is that the process data does not exist on the internet. How to book a flight is online. How to run a complex merger is not. So Harvey hires domain experts to sit down and encode the actual steps, then builds systems around them, and every prestigious firm they win makes that knowledge deeper. The model is a commodity you rent. The workflow, the data, and the judgement about whether an answer is actually right, that is the part that compounds. And it lives in the application layer, not in the model everyone else can call too.

He was just as clear about the models themselves. Harvey does not marry one provider. They build so that a model can be pulled out and swapped for another, because who is in the lead changes constantly and different models are good at different things. If that sounds familiar, it is the same conclusion I wrote about a couple of weeks ago. Own the layer that is genuinely yours. Keep the intelligence underneath it swappable. Never hand your whole stack to the one company that might wake up tomorrow and decide to compete with you.

Humans became the premium, not the casualty

A person smiling on a phone call at home
The low-end work gets automated. The moments that need a person become worth more, not less.

The back half of the interview was about jobs, and it was far more optimistic than the doom you usually hear. His view: the low-end, commoditised work gets automated, and the high-end judgement gets more valuable, not less. A young lawyer stops spending years in a data room and gets to do the real work sooner. The human in the loop becomes the premium. That is exactly how we think about voice. AI should take the boring nine calls. Your best person should take the tenth, the one that was always worth taking, with the context already gathered.

This is the bet we made with SmartAlex

Swap law for the phone and Winston is describing SmartAlex. Answering a business's phone is the messy, trust-dependent, real-world work that a raw model cannot simply do. The value is not the model. It is the workflow around it, the data that stays yours, the judgement about when to hand a call to a person, and an engine that keeps the intelligence swappable underneath. We did not build a thin layer on top of someone else's AI and hope. We built the application layer for voice, and we made sure the part that is yours stays yours.

What Harvey got right, and what we copied

The moat
It is the workflow and the data, not the model everyone can rent.
The models
Swappable underneath. No single provider owns your stack.
The humans
AI takes the volume, a person takes the calls that matter.
The trust
In anything that actually matters, reliability is the product.

The people who called Harvey a wrapper were not looking at the right layer. The value in AI was never going to sit in the model that everyone can call. It sits one level up, in the messy, specific, trusted work that a business actually runs on. For law, that turned out to be Harvey. For the phone, we intend it to be us.

Sources