Lessons from building and backing vertical AI companies over past year

If you haven’t come across this piece of news yet, pause for a moment.

Companies like Scale AI and Surge AI are reportedly generating close to a billion dollars in annual revenue each. Not by building new foundation models. Not by selling GPUs. But by doing something far less glamorous and far more essential.

They are teaching AI what “good” looks like.

Now take a second to absorb that.

Everyone talks about chips, compute, and model APIs. Very few talk about data infrastructure and the human-in-the-loop layer. Yet that layer has quietly become one of the most powerful revenue engines in the AI stack.

And it’s only getting bigger.

Models are general. Value is specific.

A friend of mine in fintech recently mentioned, almost casually, that he’s doing a side gig for a data annotation company focused on financial data.

Not generic labeling. Real work.

Annotating how financial documents should be interpreted. What matters. What doesn’t. Where nuance lives.

That’s the real story.

AI doesn’t fail because it lacks intelligence. It fails because it lacks context, judgment, and domain intuition.

It’s not enough for AI to “know finance.”
It has to think like a banker, reason like an analyst, and speak the language of Excel.

That translation layer doesn’t come from bigger models.
It comes from people who have lived inside those workflows.

That’s where the next trillion dollars of enterprise AI value will be unlocked.

Why we started gAI Ventures

Amit Goel (sold his previous company, Medici, NY based Unicorn, Prove), started gAI Ventures last year with his co-founder, Kushal Prakash (ex-AI researcher and Forbes 30 Under 30). They had a clear hypothesis:

Vertical AI is fundamentally different.

Winning here would require:

• Deep domain expertise

• Proprietary, high-quality data

• Workflow embedding

• Strong engineering discipline

• And far more than just “using AI”

We believed the companies that win will not feel like AI startups. They will feel like industry infrastructure.

So we started building deliberately.

We assembled a team of AI engineers. We partnered with Vijay Rajendran (ex-500 Global, BBVA Ventures). We invested years of operator experience into systems for idea selection, founder matching, and company building.

It was slow. Unsexy. Brick by brick.

Since then:

• We’ve consumed dozens of research papers

• Built multiple AI-native vertical AI products in financial services, enterprise productivity, and commerce. (Including FastTrackr AI, Swik AI, Stella AI, and a few other being actively built)

• Shipped a product that hit #1 on Product Hunt globally

• And launched another that now serves 13+ customers, including a $1.5B AUM RIA, within 6–7 months of inception

The pattern has been remarkably consistent.

AI works best when humans stay in the loop

This isn’t speculation. We’ve seen it firsthand.

Thomas Dohmke, CEO of GitHub, put it succinctly:

“Sooner than later, 80% of the code is going to be written by Copilot. And that doesn’t mean the developer is going to be replaced.”

Andy Jassy has said Amazon has over 1,000 generative AI services in progress.

Goldman Sachs’ CEO David Solomon noted that AI can generate 95% of an IPO prospectus (S-1) in minutes. The last 5% is where humans earn their keep.

The pattern is clear.

AI accelerates execution. Humans provide judgment.

What this looks like in practice

In wealth management, take the example of our portfolio company FastTrackr.AI:

AI can:

• Join meetings

• Process documents

• Extract structured data

• Update CRMs

• Draft follow-ups

• Help with onboarding new clients, and also transitioning.

The advisor:

• Builds trust

• Understands client nuance

• Makes judgment calls

• Navigates risk

The magic is not replacement.
It’s a reallocation of human attention.

We’ve heard the “wow” from customers when the backend friction disappears, and advisors get time back with clients.

That’s real value.

Another example: Harvey AI

Harvey AI doesn’t replace lawyers.

It:

• Drafts contracts

• Analyzes clauses

• Summarizes regulations

Lawyers:

• Apply judgment

• Negotiate

• Assess risk

The impact:

• Associates save 30–40% of drafting time

• Partners focus on strategy, not boilerplate

This is vertical AI done right.

AGI is not around the corner

Andrej Karpathy has been clear about this.

AI can beat humans at Go. It can analyze images. It can solve well-defined tasks. But it lacks curiosity, intrinsic motivation, culture, and reward systems.

Autonomous agents fail. A lot.

What works today is chunked autonomy with human supervision.

Which brings us to an uncomfortable truth.

Building AI products is 10% models, 90% messy workflows

Here’s a story that captures it perfectly.

Two tax AI startups launched on the same day. Same model.

Six months later:

• One had 10,000 users

• The other had 12

The difference?

Founder A dumped everything into the context window.
Founder B built a pipeline.

Intent detection. Retrieval at section level. Structured outputs. Reasoning scaffolds. Guardrails.

Same model. Radically different product.

That’s not AI magic. That’s product engineering.

Why vibe-coding won’t build serious companies

Tools like Cursor and Lovable are fantastic. We use them.

But they don’t build enterprise-grade products.

Chamath put it bluntly:

“Coding agents are slopware app-crappers.”

Output quantity is not output quality.

Without domain depth, rigorous engineering, and workflow embedding, you don’t get companies. You get demos.

Vertical AI needs more upfront work

More humans. More integration. More patience.

Forward-deployed engineers. Security reviews. Data privacy. Trust building.

You can’t just rename SaaS and call it AI.

Some incumbents will adapt. Many won’t.

The winners will rebuild from first principles, tightly mapping model capabilities to real problems.

The real revolution

This isn’t about GPUs or benchmarks.

It’s about knowledge transfer.

Messy, human, domain-specific wisdom turning general intelligence into something useful inside a spreadsheet, a hospital system, or a logistics dashboard.

That’s the quiet layer powering the AI boom.

And that’s where we’re building.

If you want to read further in depth, here's the link to the original Substack article written by Amit Goel (Founder and CEO of gAI.ventures)