If you haven’t come across this piece of news (screenshot below), pay attention. I felt its a good place to start.

Now, take a second to absorb that.

Everyone talks about Chips/GPUs, compute and model APIs, but data infrastructure - the human-in-the-loop layer that teaches AI what “good” looks like has quietly become the next big revenue engine in the stack.

Companies like Scale AI and Surge AI are reportedly generating close to a billion dollars annually (each) by combining data labeling, model evaluation, and feedback loops for training and fine-tuning large models.

A friend of mine in fintech casually mentioned he’s doing a side gig for one of the relatively smaller AI data annotation companies helping them annotate data in the field of finance. And that’s the real story here. Domain expertise and data not model size is the chasm AI has to cross to actually become useful deep in the industries. It’s not enough for AI to “know finance”; it has to think like a banker, reason like an analyst, and speak in the language of Excel. who will make that happen?

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

Last year when we started gAI Ventures (alongwith Kushal - my co-founder/CTO, ex-AI researcher and AI architect) we had a hypothesis that Vertical AI is different and will require domain expertise and propriety data and more than just AI. We believed that Industry-specific AI startups that embed deeply into existing workflows, automate high-value processes, and drive measurable economic impact will win. So we assembled a team of AI engineers and partnered with Vijay (ex-500 Global and BBVA Ventures). Together we started building gAI Ventures brick by brick, putting together the systems and processes to identify ideas/opportunities and the right founders. Its been a lot of hard work and you have to stay focused and avoid distractions. Since then we have done a lot of R&D, gulped dozens of research papers (arXiv) and built a lot of AI native stuff including a product that went to #1 product on product hunt globally and another one that serves 13+ customers including a $1.5 B Aum RIA in 6-7 months from just an idea.

Here are some lessons we have learned on the way building and investing in Vertical AI:

1. Generative AI tech can do wonders. This is a generational opportunity to blend human intuition with AI efficiency

   • Thomas Dohmke (CEO of GitHub)

     “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 (CEO of Amazon)

     “Today, in virtually every corner of the company, we’re using Generative AI… We have over 1,000 Generative AI services and applications in progress or built.”

   • Goldman Sachs CEO David Solomon. He stated that AI can generate “95% of an IPO prospectus (S-1 document)” in minutes, leaving only the final 5% as the human-differentiator.

   We have seen it first-hand building AI native products and hearing that “wow” from customers. Imagine from gathering data via meeting bots and document processor AI, extracting information and updating CRM and financial planning software while allowing wealth advisors to spend more time with customers (investors) and letting AI doing the back-end heavy lifting.

   Another Example: Harvey AI - AI drafts contracts, analyzes clauses, or summarizes regulations. Lawyers then apply domain expertise, negotiation skill, and risk judgment. Impact:

   • Associates save 30–40% of drafting/review time.

   • Partners focus on strategy, not boilerplate.

2. And oh… “AGI is still a decade away”

   Andrej Karpathy in the excellent podcast (Dwarkesh) said AI can beat us at Go, schoolwork, analyzing medical images, and many well defined tasks. At the same time, it lacks the reward systems that humans use to improve including curiosity, empowerment, play, intrinsic motivation, and culture.

We now know that we’re years away from a “magical” superintelligence that will do anything and everything for us. Many insiders who are building in AI knew but the latest Karpathy podcast cleared the air. What we have right now, increasingly, are AI models that exhibit strong general reasoning capabilities with improving skills in specific domains. This will all get better and better over time, unquestionably. But they still need a ton to be useful in most domains.

Karpathy said in a tweet:

Podcast link if you want to go deeper.

3. AI is not good at many things today and autonomous agents fail all the time

   We’ve learned that AI progress isn’t linear. It’s loops of sweat, code, and small wins. Turns out, building AI products is 10% models and 90% messy human workflows. I read a post recently …

Two identical tax AI startups launched on the same day. Six months later, one has 10,000 users, the other has 12. They used the exact same model. Here’s what made the difference:

Founder A dumped everything into the context window:

→ Full transcripts
→ Entire IRS documents
→ Generic instructions

Results were inconsistent. Sometimes wrong.

Founder B built a tight pipeline:

→ User intent selector (”file,” “explain,” “plan”)
→ IRS content retrieved at section level only
→ Compact user profile summaries
→ Output constrained by formatting rules
→ Injected reasoning steps before answers

Same underlying model. Very different product.

Founder B’s assistant felt smarter, more reliable, more professional. That’s why his product exploded over the past few months.

and then I saw this tweet as well:

and the other day I got this email and look at it carefully (after gazillion emails being sent by AI, email campaigns are at their lowest performance levels):

writing is on the wall.

4. No, vibe-coding cannot help you build enterprise and/or production ready products. I will start with a deeply concerning chart…

Cursor and Lovable are awesome products. we use them all the time to build smaller modules. but you still can’t build a production ready enterprise ready product with them.

» “Coding Agents are Slopware App-crappers” said Chamath Palihapitiya

• Basically Chamath is saying: these “coding agents” (AI code bots, generative tools) are churning out low-quality software (“slopware”), producing many apps or features that are shallow, half-baked, maybe hype-driven rather than value-driven.

• In other words: Using coding agents often leads to a volume play (lots of app/features) but with weak discipline, resulting in many “app-crappers” rather than high-quality, well-engineered products.

• The implication: Just because you have coding agents doesn’t mean you build something good. you risk building more junk/mediocre apps.

• don’t confuse tool → output quantity with output quality. Without rigorous engineering + product design + domain embedding you risk building “slopware app-crappers”.

• And on top of that - In the vertical AI domain, what matters is domain depth, data, embedding into workflows, strong engineering, proper product-market fit and… not just “generate lots of features with an agent”.

5. AI’s capabilities are not superhuman (yet) and have limitations - If you watch how the best work gets done with AI, it happens in chunks where a human supervises the output and gives iterative feedback to the AI. Large scale autonomous agents are brittle and fail quickly. Watch anyone vibe code an application with any level of novel complexity.

   For vertical AI you need to have the prop data, understand the domain, the workflows, data privacy, security and did I say “data”.

   AI Models are great but what they need, to be most effective, in real world situations is context and usually an applied set of features to make the AI useful. They need the data that you’re working with, they need to show up in user interfaces that are relevant to the task, they need to have a sense of the domain that they’re being applied to.

6. Vertical AI needs more in the beginning

   AI also faces integration barriers into existing organizations. AI is not capable of pulling a lot of the levers you need to be effective like coordinating with multiple stakeholders, building trust, authenticity, and interacting with different modalities across time and space. You could argue that the average human doesn’t either, but people know when they’re interacting with an AI and don’t allow it the same agency as they do to people. The most successful AI B2B companies in many cases actually need more humans to integrate what they’ve built (forward deployed engineers) than traditional B2B SaaS. something builders and insiders know

7. You can’t just rename a legacy software or saas company and make it AI

   Some companies are sleeping. Some legacy software cos are complacent (one guy told me we have been doing AI for 20 years). While 90% of incumbents haven’t adapted, it is possible. Figma, Notion, Vercel, Box, and Intercom have done a great job of tearing down what they have and rebuilding AI native products. They have teams who are close to the current capabilities of the models. They also understand their problem domain and as new capabilities come out know what capabilities will map well to what problems in what way. They are able to deliver on AI’s promise to their customers. Whereas the majority of existing companies will die.

Okay….to wrap this up, I’ll leave you with a tweet that captures the spirit of what’s really happening beneath all the AI hype. That’s the quiet revolution. The real story is more that just about GPUs, model sizes, or even benchmarks. It’s about the knowledge transfer. The messy, human, domain-specific wisdom that turns general intelligence into something useful inside a spreadsheet, hospital, or logistics dashboard.

I will tell you what is the most important balance but first what are Vertical AI startups?

They are tech (AI) businesses that combine strong industry expertise, moats built on unique data, and AI at the heart of their value prop.

These startups are most impactful in industries where human effort still dominates and segments are data-heavy and labor-dependent. a good example is Harvey AI for law firms, etc. and FastTrackr AI for RIAs and financial advisors, that we are building

Ok now that we have some background what is that balance I spoke about??

In a Vertical AI company you need industry insiders early on. Trust, distribution, and deep context matters a lot.

Tech-first alone = lost in the wild (pivots after pivots; pilots not going to prod). Domain-first alone = trapped in the box (trying to build everything all at once; incremental vs. disruption).

You need both. but you see this in less than 5% of the startups. Its hard to get.

Vertical AI Teams Need Both Domain & Tech DNA from Day One

If you’re building a vertical AI company and you don’t bring in industry veterans early, you’re setting yourself for probable failure.

Full-stack AI products can’t be built in a vacuum. they require intimate knowledge of the domain: how workflows actually run, what pain really feels like, and what outcomes matter. But just as critical is the ability to encode that knowledge into models, interfaces, and systems that scale. That’s the intersection of deep subject matter and engineering excellence. and its not a new thing. even back in the FinTech era Nubank did the same. David knew this matters a lot:

Without domain leadership early, your GTM efforts will stall. Customers won’t trust you. You’ll miss the nuances. You’ll design for ideal workflows that don’t exist in the real world.

Without tech depth and AI expertise, you’ll build slide decks, not software. Your product will be a service company in disguise, duct-taped together, and doomed to stall at scale.

Too many vertical AI startups fall into one of two traps:

1. Domain-first founders who stay stuck in the box…bound by how things have always been done. Trying to boil the ocean with waterfall specs and slow-motion execution. Stuck optimizing yesterday’s problems. Or trying to build everything all at once (the mega solution like a big company).

2. Tech-first teams who operate like tourists…lost in a foreign land without a map. Chasing shiny demos, endless pivots, pilots that never make it to prod. Looks impressive, many a times delivers nothing. Churn!

The strongest vertical AI companies don’t just bolt on domain knowledge later. They bake it into the founding DNA. And that matters, because DNA doesn’t evolve easily. If you don’t get both sides in sync from the start, one side will eventually dominate and the other will never earn equal influence. Once that imbalance sets in, it becomes nearly impossible to attract top-tier talent from the missing side.

We at gAI Ventures are working hard to make this happen.

Great vertical AI startups are rare not because the markets aren’t big enough but because getting the right team with shared language and mutual respect is brutally hard.

But when it clicks, that’s when real magic happens. Hoping to build a few of them!

Its been a while since I wrote my last post. I was on the road for the entire month of Feb. Covering the West Coast (SF/Bay Area) and East Coast (NYC/NJ/multiple). Met many awesome people who directly or indirectly are thinking/doing something in AI and in many cases influencing it greatly. And I decided to write down my learnings and way forward in AI. and finally got some time over this weekend. I probably should keep all of this internal, but there are a couple of objectives that I want to achieve by sharing it widely — Get feedback on this post from my open network and discuss with other B2B AI founders.

At gAI ventures as some you might know, our focus is on AI products for financial services. There are many areas within banks, insurance companies, capital markets and even Fintechs where dozens or even hundreds of people are employed to do manual tasks and tedious processes. Our aim is to bring productivity and efficiency gains to these departments using AI. We partner with domain experts in niche areas to do that. Vertical AI = financial services

Okay. with that little background, let’s jump in.

Much of the discussion of 2023 and 2024 in AI was revolving around horizontal capabilities of foundational models. In 2025 as foundational model layer gets very competitive with Deepseek/Chinese models and other players launching powerful LLM models, it will get commoditised. As Satya Nadella said, OpenAI is now an AI products company that happens to have foundational models as well.

Now in 2025 attention is shifting to application layer. Today vertical AI is the talk of the town and for the right reasons. We’re entering the Vertical AI decade, where the winners won’t just be those with the best models but those with the deepest integration into workflows. Industry specific AI startups that embed deeply into existing workflows, automate high-value processes, and drive measurable business impact. US AI + FSI/fintech mkt is unraveling now…

US in general is way ahead in adoption than Europe or India or middle-east. From Morgan Stanley and similar Fortune 500 firms to small $25M AUM RIA I met, they are all doing something with AI tools.

Goldman Sachs CEO, David Solomon, said that AI is drafting 95% of an S1 IPO prospectus “in minutes” (a job that used to require a 6-person team multiple weeks)

Stripe Engineering: Engineers reported spending 70% less time writing unit tests after implementing AI tools, fundamentally changing their test-driven development approach.

Amazon reported the impact of its AI system called Q: With Q’s code transformation capabilities, Amazon has migrated over 30,000 Java JDK applications in a few months, saving the company $260 million and 4,500 developer years, compared to what it would have otherwise cost 

Morgan Stanley built an assistant app for financial advisors to improve their productivity using AI and ability to answer client questions with huge digital memory of the org. AI to automate tasks such as summarizing meetings, drafting follow-up communications, and finding relevant insights within seconds. 16,000 financial advisors using AI Assistant. 30 minutes saved per client meeting. 10-15 hours saved per advisor per week. (Small to mid-size cos using tools like Jump, Zochs, and FT

Just to be clear 90% of the market is still seeing it as pilots or POCs though. Good for innovation and good for startups! But its still early days in a 10 year long tech replacement cycle. we believe a lot of Software/SaaS that has been written in the last 25 years maybe rewritten - the unbundling has started and there will be rebundling in the future. This happened in Fintech as well from 2013 to 2023.

but AI is also a lot different. First of all there is too much competition in AI. Everyone wants a piece of the pie. and no one seems to be sitting idle. There is Box and Salesforce and other SaaS incumbents in many enterprise AI conversations. There is Accenture, Mckinsey and IBM in many conversations. And there are 1000s of startups in many conversation — big and small. There are indie hackers. And to top it all many enterprises and corporates want to try out things in-house first. So, as an AI startup, how do you win?* (will come back to this)

Another way to look at it is with this question someone asked - How to think about building products when anyone with Cursor, replit, v0, bolt etc can ship software in 24 hours and you're competing with thousands, not dozens. most founders follow a straight line: find problems, build solutions, repeat. but in an AI world, this approach is a fast track to commoditization. when anyone can identify problems and generate solutions in seconds, following the straight line only leads to Ok companies with no moats.

From an investment perspective, you will feel like 2021 all over again. There is a lot of investment happening in AI and some of it at crazy multiples. Reason is that large AUM investors look at things differently. They get exits from such investments in 7 to 15 years. So they are looking at a future value or terminal value and discounting their way backwards, so they feel that some of these companies could one day become so valuable. It will probably be worth it.

But for Pre-Seed & Seed Investors, Trace Cohen wrote a nice post on Why Vertical AI is the Best AI Investment and another good one by Euclid Ventures

As it happens with every new revolutionary tech that after the initial super bullish exuberance dies down … when shiny startups from silicon valley can no longer take the AI hammer and say everything looks like a nail. Well, we have seen this movie before.

The “AI can do everything” phase starts to cool down (not today), and we enter the “sober execution” phase in the future sometime where:

Hype-driven startups may fail – Those that didn’t find real product-market fit or relied too much on AI as a gimmick will struggle.

AI integrates into the workflows – Instead of being standalone AI tools, AI gets embedded into existing software and enterprise systems.

Regulation always catches up – Governments start shaping policies around AI safety, copyright, and employment impact.

ROI expectations will rise – Companies demand real cost savings and productivity gains, not just “cool demos.”

We’ve seen this before with cloud computing, blockchain, VR/AR, and even the internet itself. The survivors are usually the companies that:

1. Solve a real problem (not just “AI for the sake of AI”).

2. Have strong distribution & sales execution (AI startups need more than just good tech).

3. Evolve with AI’s limitations (understanding where AI is actually useful vs. where it’s just hype).

Now let’s deep dive into the main part of this story.

*So dear founders, how do you win?

1. Industry first and problem first approach - In every technology hype cycle we have seen that it comes back to identifying problems first and then thinking about the solution. As an example from one of our own startup - theAudit.ai

First thing is to have discussions and brainstorm about the problem statement for months before you decide - Lets go and solve it. Audit is a lengthy, costly and hard to deal with process. Building Audit AI, Anzar (has 18 yrs+ experience in compliance and audits), Kushal (CTO) and I asked lot of questions to ourselves:
Why compliance and audit space? Why is this bundle of tasks bundled this way? Why did they design Audit like this in the first place? Why only sample transactions, we could look at (controls testing) all the transactions. And so on.

—> then comes the solution part that maybe we can use AI to build the best product possible. this is the right order I guess.

To further illustrate the point read this note:

2. Choose the right business model and how will you position your AI offering:

   1. you can start with Service as software - This is the most interesting. people are calling it by various names. Dharmesh shah of Hubspot called it Waas and Raas . Owning the flow end to end = new AI-native way = Service as software.

      essentially you promise outcomes and you deliver outcomes and you get paid for it. Infact, today when many enteprises/corporates are reluctant and unsure about data security and other issues it is even better that you don't ask the customer to adopt AI. In fact, in most cases, the customer does not have to learn about AI and how to adopt it. They just want to get things cheaper, better, and faster. Gushworks is a good example https://www.gushwork.ai/

   2. another way is to follow the vertical saas playbook. As Sangeet Paul said - Vertical software repeatedly followed a playbook of

      (1) gaining adoption with a focused use case like Harvey AI (offering tools primarily for drafting, reviewing, and researching legal documents)

      (2) rapidly rebundling adjacent capabilities around that initial use case.

      Long-term competitive moats in vertical AI are created by rebundling around the initial use case, using the unique rebundling benefits that AI provides.

3. Market selection, strong market need and GTM:

   Pick a great market - first hand experience - the US is way ahead in the adoption of AI than EU/India or many other regions. 3 large fortune 500 cos I met there recently have so much going on in AI and so does small companies in the same space (already spending on AI tools). Across the length and breadth of corporate America you can take a niche and you will get customers for your B2B AI. And your product will actually make money and have margins. Compare that to India or ME where it can be challenging.

   It's difficult to sell in the US as well, especially for first timers. You need previous relationships and connects. You need face time. You need to embed in the culture. You need local teams. You need capital for events and wining and dining folks. And most importantly you need to excel on the product side - in the US the best product wins. Unlike some other markets where the best distribution wins

   Strong market need - Dharmesh Shah of hubspot recently said - The best way to win is not to start by developing a strong product but to discover a strong market need. If there's strong market need, but you have a weak product, you can iterate and make the product better over time. You can control your product iteration. You can't control market pull. I couldn’t have said it better.

   Also what will the GTM team look like (probably needs another full blog).

   and what is your sales philosophy? I read an AI startup founder wrote it well - Sales isn't just about closing a deal. it's about creating relationships where everyone wins. You help your customer win first and make money in the process and not the other way round. Try to be an amazing partner first. I can’t emphasize how the long term thinking helps your sales function. think about your sales philosophy

4. Pricing is another area which needs a lot of thinking. depending on your AI product proposition and whether you are selling seats or outcomes or some other way.

   Old SaaS Model → Per-seat pricing (e.g., Salesforce).

   AI Model Shift → Usage-based, value-based, or AI-Agent-driven (e.g., Harvey AI, Gushworks).

   Look at some data below:

   

For further reading on pricing models there is an interesting read - X post by Aaron Levie of Box - This provides a completely new growth vector for software companies in AI, and has major implications to software business models. For a wide number of use-cases, there may be little to no connection between the number of users on a platform and the total amount of usage of AI Agents that use the software -- and what the value is that those Agents can deliver. And for many use-cases, the idea of a user seat being tied to the agentic workflow altogether is unnecessary when agents are just running around completing work in the background for an enterprise.

If you’re an AI founder in financial services, the time is now. Let’s connect!

what’s working for you, and what challenges are you seeing?