How to build vertical AI

Lesson 1

Define world-class before you build anything

The first job isn't choosing a model. It's writing down what the best practitioners in the domain actually know and do. His team distilled theirs from the best salespeople they could hire: knowledge (which customers fit, who decides, what matters to them), judgment (who to contact, with what angle, and when), procedure (the repeatable steps), and adaptation (changing course as the customer changes).

That definition is the specification for everything after it, and it's where the value is: when software is a commodity, the scarce thing is domain expertise.

“Go to market is one of the few remaining sources of alpha”00:05:43

The reading: each of the four maps to a layer of the product. Knowledge becomes data, judgment and adaptation become memory, procedure becomes workflows, and “world-class” becomes the evaluation.

Evidence: transcript 00:03:35, 00:05:43

Lesson 2

The chassis is generic; the effort goes into the domain

Most vertical AI systems share the same layers: the domain's data, a workspace where agents run, memory, domain workflows guided by experts, and evaluation running through all of it. That structure is reusable across industries, so don't spend your originality there.

“I could have pulled this and said this was a talk about insurance.”00:08:33

The work, and the advantage, is in what you put in each layer for your domain.

From the notes: “a GTM harness, or how to build a product in a domain-specific area”: the harness is the reusable part; the domain is the product.

Evidence: transcript 00:08:33

Lesson 3

Own the domain data nobody can buy

Separate the data anyone can get from the data only your customers generate. In the sales example, records about companies and people can be bought or scraped. What matters is each customer's own activity, their meetings and emails, flowing in continuously so the system's picture of their work is always current. Expose all of it to agents through one tool interface that can also plug in the customer's own systems. He calls this the most strategic piece, and more art than science.

The reading: the same move as George Fraser's talk that day: agents should read data that's already been brought together and kept current, not reach into every source live.

Evidence: transcript 00:10:23

Lesson 4

Encode the experts' workflows, and leave room for the customer's

Turn each expert practice into an opinionated workflow an agent follows: meeting prep, coaching, pipeline review. His examples show why. The expert move for a prospect who went quiet (follow up, at the right moment, framed the right way) becomes an agent with reply rates above 30 percent. The expert move after a win (go after the companies most like that customer) has led to deals as large as $15 million.

You can't know every niche, so also let customers encode their own playbooks with an agent builder.

“you know better than we do how to sell to dentists in the Midwest”00:15:00

From the notes: “vertical AI is new lingo, capturing workflows that are specific to domains”.

Evidence: transcript 00:15:00, 00:18:47

Lesson 5

Capture judgment as it emerges

Expertise isn't static, so the system has to keep learning it. Seed memory from what each customer brings at onboarding: their context, preferences, and definition of a good customer. Then add a learning loop his team calls reflections: when something meaningful happens, like a meeting booked or a deal closed, agents look back at what led to it, write candidate lessons, rank them, and promote the best into shared memory.

“Right, this is where you start capturing things like judgment.”00:08:33

Evidence: transcript 00:08:33, 00:10:23, 00:26:46

Lesson 6

Trace everything, and let experts define correct

You can't improve what you can't see, so trace every agent run. But traces only show what happened. In a vertical, “good” is a judgment with no universal answer, so the domain experts, not the engineers, define it and set the evaluations.

“Absolute correctness is only correctness informed by subject matter experts.”00:17:41

From the notes: “tracing and observability is key for everyone to see”: the trace is the evidence, the expert is the judge.

Evidence: transcript 00:15:00, 00:17:41

Lesson 7

Put the experts inside the build

The hardest gap is between domain experts, who often aren't technical, and the engineers. His team's answers are all organizational:

Use your own product, so you feel every failure first.
Put your best domain expert in charge of deploying agents, on the product team, rather than an engineer or product manager.
Give experts an equal vote in product design, and make them the first users of what ships.
Have experts and agents share the same tools, and make every trace visible to everyone.
Run hackathons that put experts and engineers in one room.

“This is literally the best salesperson at Monaco.”00:20:57

He closes on empathy: understanding the person across the table is what makes the translation from expertise to product work.

The reading, for The Focus: the process for a client is the same seven steps. (1) Write down what their best people know. (2) Reuse a generic chassis. (3) Bring their own data together and keep it current. (4) Turn expert practice into workflows, with room for their own. (5) Capture judgment as it emerges. (6) Trace everything and let their experts judge. (7) Keep those experts inside the build. Only step 2 carries from client to client.

Read alongsideHis Latent Space episode two days before the talk, How Stanford Teaches AI-Native Engineering, on teaching AI-native engineering.

Evidence: transcript 00:20:57, 00:25:21