Chainet LogoChainet
Back to Home
From Wall Street to AI Agents: Building Vertical AI for Financial Services

From Wall Street to AI Agents: Building Vertical AI for Financial Services

A deep dive with Ankur Patel, founder of Multimodal.dev, exploring how AI agents are transforming financial services through vertical AI solutions, document processing, and systematic workflows that achieve 99.5% accuracy

August 14, 2025
15 min read
By Rachit Magon

Listen on

There's a fundamental difference between building AI tools and building AI that financial institutions will actually trust with their most critical processes. While most companies are wrapping chatbots in fancy interfaces, something far more sophisticated is emerging in the world of finance.

The challenge? Getting AI to work with the same precision that Wall Street demands, where 95% accuracy isn't good enough and mistakes can cost millions. Today's guest has lived in both worlds: trading emerging market credit at Bridgewater Associates and now building AI agents that process 500 page SEC filings in under six minutes.

Meet Ankur Patel, founder of Multimodal.dev and Agent Flow, where he's solving one of the hardest problems in enterprise AI: getting financial institutions to trust AI with their most sensitive workflows, from loan underwriting to investment decisions.

Key Takeaways: Vertical AI and Financial Services Transformation

The Precision Challenge:

Wall Street Meets Machine Learning:

Enterprise Sales Reality:

Q: Can you give us your background? What was it like working at JP Morgan and Bridgewater?

Ankur: It was definitely a huge privilege to start my career there at these places. JP Morgan was coming out of the financial crisis, so I was there 2008, 2009, 2010. In that era there was just absorption of Bear Stearns and trying to deal with the aftermath. You saw firsthand some of the complexity. Very large, very complex operations. I think a lot of the pain that I saw there firsthand still lives with me.

Bridgewater was a very different area because it was under a thousand people. It almost felt like a startup even though it had been about 35 years old at the time, but it was a very focused team, the largest hedge fund at the time coming out of 2008 with great success. So learned a lot firsthand from some of the very best firms in the industry.

🔥 ChaiNet's Hot Take: There's no substitute for seeing institutional complexity firsthand. The pain points Ankur experienced in traditional finance became the foundation for building AI solutions that actually work in production.

Q: You were trading emerging market credit at Bridgewater. How did that experience shape how you think about AI and automation today?

Ankur: Bridgewater thinks so systematically. Before machine learning and data science and AI, these things weren't really mainstays when I joined Bridgewater in 2010. But they had a very systematic process where they took their investment logic and built out essentially rules, very complex set of rules on how to absorb market data and then trade markets with that.

I loved that systematic process of how you had to take logic that was in your head, write it down into rules so you could systematize how you process data and make decisions. That stuck with me because then what I did next was I started applying data science, a more formal technique to do very much the same. And then of course fast forward and you have proper machine learning and AI today.

🔥 ChaiNet's Hot Take: The best AI builders understand systems thinking first, technology second. Bridgewater's rule based approach was AI before AI existed, proving that systematic thinking transcends specific technologies.

Q: You started R Squared Macro in 2012 when people were skeptical of data science for investment decisions. What's changed between 2012 and now?

Ankur: Things have changed slowly but they've certainly started to change. In 2012, if you told the typical hedge fund that more and more of their decisions would be made by algorithms, they would have said no way, this is too complex, we need human expertise, human brain power to make these decisions.

It took some time, but basically every single shop, whether qualitative and fundamental or quantitative, all of them are using algorithms to make decisions now. It's very much a mainstay in investment management.

We're still having some similar conversations today. As good as AI is, people don't feel that it can ever match human expertise for some of the most critical work around investment management or investment banking or underwriting. I think it's going to take some time to change.

But what you realize is that when fed enough data and enough context and also when you have human experts that are providing feedback and coaching, AI can reach human level performance. Just like we started trusting algorithms to make investment decisions 10 years ago with hedge funds, you're going to see a similar arc here where more investment firms and investment banks and insurance carriers are going to trust AI to do some of the most critical work they've ever done.

🔥 ChaiNet's Hot Take: Adoption curves in finance move slowly, but they move completely. What seems impossible today becomes standard practice within a decade. The skeptics of 2012 became the adopters of 2022.

Q: There must be at least 50 companies saying they're doing AI for finance. What's the difference between a fintech chatbot wrapper and something like Agent Flow?

Ankur: We're in the part of the cycle where there are many startups, more startups now, and I'm sure they'll keep rising and there will be some consolidation. But I think going back to Multimodal, we start everything from first principles which is we know the pain that people in finance experience. We've lived in their shoes. We know these workflows.

That deep domain expertise matters because not only do you need the business knowhow but you need to know how AI gets incorporated in these business processes. So you need to incorporate trust and governance and information security and you also need to know what the interaction typically is between business operators and AI.

There are a lot of different experiences, application components that you need to bring together to serve very sensitive business use cases with all of the compliance and regulatory needs that these banks and carriers have and then encourage adoption.

But the number one thing that differentiates us from everyone else is that we realize these workflows are very sensitive. You cannot screw them up. The precision needs to be insanely high and so we've incorporated really great proprietary ways of taking feedback from business operators and incorporating that into the AI. That is the difference maker because a lot of that expertise today lives in the heads of people, investment professionals have expertise in their heads, it's not written down so AI can't easily access it.

We're trying to find clever proprietary ways that we've developed to take feedback from people, incorporate that into AI so AI can think and operate in the same ways, the same levels of precision that people do today.

🔥 ChaiNet's Hot Take: Domain expertise isn't optional in vertical AI, it's the entire moat. Anyone can wrap GPT-4, but understanding financial workflows deeply enough to achieve 99.5% accuracy requires years in the trenches.

Q: Walk us through a real life example. When a bank processes a loan application, what's actually happening and where can an AI agent come in?

Ankur: There are many steps involved in the process. When a loan application comes in it usually comes in through web portal or email. The initial job for a loan officer is to do triage. Read in the application, see if all the supporting documentation is there: bank statements, driver's licenses, tax returns, pay stubs, verification of employment, all of these things.

If items are missing or if the application just isn't a good credit opportunity, the loan officer will need to either turn it down or ask for more information. That's a really good job for an AI to take on because it's administrative in nature. You're just reading in information. So we usually start with that, triaging the volume of applications that come in.

The next step is that if all the data is there, the application's properly filled out, it's a good credit opportunity, all the supporting documentation's there, we can help do some of the diligence that the loan officer and underwriters need to do. Usually in lending, especially if it's something like mortgages, there is a checklist of things you need to see. There are a set of ratios that you need to pull out. This is a very good job for AI to do as well.

So it can support the diligence piece and then go to the underwriter and say we checked for the 20 things that you need to check for that your organization needs to check for and here is what we think the recommendation is and we can pinpoint exactly where that information came from.

Now the nature of the job for the underwriter has changed because they're doing more final review and that allows them to more quickly get the quote out the door. I will tell you one secret here which is really critical: the first lender that gets a quote out has a disproportionate chance of winning. So speed is the name of the game and we're enabling our customers to not only reduce expense but be the first quote out the door which gives them better chance of winning on the opportunity.

🔥 ChaiNet's Hot Take: Speed is the silent killer advantage in lending. While everyone focuses on cost savings, being first to quote can be worth millions in competitive wins. AI's real value isn't replacing humans, it's making them faster.

Q: How do you think the job of an underwriter would change with all the AI coming in? Will there be major job losses like in the tech industry?

Ankur: The nature of the role certainly changes. Instead of doing data entry and data extraction, even applying rules and doing aggregations, performing some math, preparing loan documentation, preparing a quote, corresponding with the end user, I think that's something that you could hand off to an AI.

Where loan officers and underwriters will spend more time is doing deeper diligence, reviewing the initial read from the AI and they'll spend more time shepherding the individuals that are looking at the lending opportunity, explaining the quote, explaining the different details, spending more time doing the high EQ parts of the job that require just human touch and experience.

So I think the nature of the role changes. You'll see more market demand. It's hard to say is that going to result in net job loss or will it actually result in an increase of people working.

Great example: ATMs were thought to make bank tellers extinct but now we have more bank tellers than before ATMs came about and it's because the demand for banking as a whole has increased dramatically because we made it easier for people to bank.

🔥 ChaiNet's Hot Take: The ATM paradox applies to AI. Automation doesn't necessarily destroy jobs, it often expands markets. When processes become easier and cheaper, demand increases, creating new work that didn't exist before.

Q: You've built a document AI that claims 98% accuracy and can process a 500 page SEC filing in under six minutes. How did you achieve that?

Ankur: I've been working on document processing for such a long time. It dates back two startups, three startups ago. My last startup did invoice processing and spend intelligence. So I have a long history of working in this space and when we started Multimodal two years ago we started building all of this again from first principles.

We treated what people think of as the abnormal document types and we took them as a given. Handwritten notes, poor quality scans, screenshots of documents, just messy documents in addition to different tables and charts that are pretty much the more dominant part of SEC filings and analyst reports and such. We had to build all those components from the ground up.

When you look at a sell side analyst report, it's going to have appendices, footnotes, sidebars. It's going to have different charts, some with line charts, vertical bars, horizontal bars, and have various tables. We had to build an agentic workflow that used both text processing and image processing to go and find all those elements and process them independently. Some using code, so when you take a look at a table, you want code to then read that in as a CSV or a markdown file and then work off of it.

We had to build this and we had to build it in a way where you get to 98% out of the box. To be honest though, we use feedback cycles to then get to 99.5%.

When you compare our product even out of the box, but certainly after feedback, it is dramatically more robust and resilient versus just throwing this at ChatGPT. And for the workflows we target, that's everything because in finance you cannot make mistakes. If it's 95% it's unusable. You need to get to 99.5%. And we enable that through the feedback cycles we do.

🔥 ChaiNet's Hot Take: The last 4.5% from 95% to 99.5% is where vertical AI companies win. Generic models get you to "good enough" for casual use, but mission critical workflows require specialized architectures and feedback loops that only domain experts can build.

Q: Does AI also help you in integrating to these tools or is that more manual?

Ankur: That's some of the biggest insights that we've had over the last two years. It's become easier to use AI and AI agents, some of the newer capabilities around agentic AI, to help with the integration piece itself.

Now all of a sudden it takes less effort and you could do integrations more quickly and we build that as part of our business model. Our teams that do forward deployed engineering work on integrations are using agentic AI to speed up the integration piece which means that you could do more complex, more custom, more longtail integrations that would have been impossible historically because it would have been too expensive and too time intensive to do.

🔥 ChaiNet's Hot Take: AI building AI infrastructure is the meta layer most people miss. When your integration team uses agents to build integrations faster, you've achieved compound leverage that competitors can't match.

Q: What does a vertical AI solution handle that a horizontal AI solution just cannot?

Ankur: I think there's deep value in horizontal software platforms because they can do a lot of use cases out of the box and they could serve different departments in different verticals. It's a different game.

The reason why vertical AI is really important, and usually a company will use both horizontal and vertical players, is that a vertical AI player like us, when we go into insurance we know what a particular type of insurance carrier given the geography that they operate in and given the products that they cover, what they're trying to do, what are the fields that they're trying to extract, what are the decisions they're trying to make, how should we corroborate how the AI made those decisions so they get explainability built in.

Those things are pre-built automatically and you do have the ability to configure certain things that are specific to your organization, but we have a lot of that pre-built. We also have designed experiences where we strive for that 99.5% or higher precision. So you have feedback cycles, you have citations, you have confidence scores, you have these trust capabilities that we've built in that a horizontal platform player just doesn't need to think about or consider because their users aren't going to need that 99.5% precision. They're okay with 97% precision.

Eking out the last few percentage points, having the governance and trust features built in, we've done that. So vertical specificity, a lot of the governance and trust features, these are things that we've built into our platform and it makes it a much better fit from day one with a large carrier versus if they had to try to make a horizontal platform work.

🔥 ChaiNet's Hot Take: Horizontal platforms optimize for breadth, vertical AI optimizes for depth. In regulated industries where mistakes have legal consequences, that depth isn't a nice to have, it's the price of entry.

Q: As a founder, especially dealing with financial institutions that are risk averse with long sales cycles, how do you actually convince them to trust AI with mission critical work?

Ankur: The first step is not even to trust the AI, they need to trust you. The company, you as a founder, you as a product team. When you talk about enterprise sales cycles they are insanely long, two years sometimes two years plus, and it comes down to trust and relationship building.

They need to feel and know that you can be a trusted partner, that you know their pain, you know their domain. You understand what they're trying to do and that you have the capabilities as a company, as a team, to serve them well.

Every single touch point is an opportunity to either engender trust or erode trust. So POCs and pilots and everything we do in between really important in building trust and snowballing that over time. So I view it as a people problem first and then the technology part of course is important, which is the technology itself needs to be technology that they could understand so there's explainability, it's not so black boxy.

They could see how they can convince the rest of their organization to adopt it and we help and shepherd it. But a really good relationship is one where the company that we're selling into, the enterprise, we have champions there that then become part of our united front. We're almost on the same team trying to shepherd the adoption of AI within the larger organization.

I will say that we work a lot on enterprise, we're also doing government work, but we also serve the mid market players as well. Regional banks, credit unions, community banks. They have the same set of problems as the larger enterprises, but they don't have quite as many people to solve this problem. They don't have quite as many options frankly because they're not debating do they build this in-house or not. They're looking at partners to work with.

Those tend to be quicker. So we're basically moving those more quickly through the sales funnel, getting them to value, also getting good reps ourselves in terms of product and customer success and such. That also ends up being a very good referenceable customer list for the enterprises that want to know that you've done this well now many times over with other institutions in some capacity.

🔥 ChaiNet's Hot Take: Enterprise sales is a trust game disguised as a technology sale. Your first customer isn't buying your product, they're betting on you as a founder. Mid market customers give you the battle scars and proof points you need to win enterprise deals.

Q: Going from zero to one is the hardest part. How did you get your first client when you just had an MVP or POC?

Ankur: It helps that this is my third startup. I've been doing this a long time. I have a good network. I can get into more conversations more quickly.

But the advice that I have is if you're trying to do zero to one, you want to make sure that you really spend time with prospects and would be clients or customers and you try to understand what's the pain they're trying to solve for. What are they experiencing? What have they tried? Where's the skepticism?

You start with then building product with that pain in mind. If you do a good job in identifying the first set of prospects to pursue and you build the right product and you're like, "Hey look, I heard you through the conversations we've had. You've articulated this pain. Here's what we've built. We'd love for you to try."

Once they see that and you've really addressed the pain that they've articulated, it starts to become a more organic conversation where they want to try and then pay for it. That's how we got our earliest customers where we really understood and heard what pain that they had. We built product in light of that and then we started getting revenue.

Now over time the product changes but our mission at large, our vision has been very clear which is we're trying to automate some of the most complex work that shouldn't require as much people effort just given where AI is. We're on that mission with a very clear north star and I think the means to get there have changed a bit over time as they always do. Agentic wasn't quite as built out two years ago but I think we delivered good value from the very beginning.

🔥 ChaiNet's Hot Take: Product market fit starts with pain recognition, not feature lists. The best first customers are the ones who articulated the problem to you before you built the solution. They're pre-sold because you listened.

Q: What's one thing you stopped doing when you became the CEO of your startup?

Ankur: One thing I stopped doing was saying yes to things that maybe didn't move the needle quite as much. Time is the most precious asset you have as a founder. For me it was where do I spend the precious time that I have to make the company successful?

You have to be insanely ruthless about saying yes and no to things. I got much better at saying no far more frequently than at my previous startups where I was more open to exploring and saying yes to different things and different paths and different meetings and conversations people wanted to have.

Now we just have to be very ruthless in terms of what we take on and what we don't. That's the number one thing that's changed.

🔥 ChaiNet's Hot Take: Your calendar is your strategy. Every yes to something unimportant is a no to something critical. Third time founders know that ruthless prioritization isn't optional, it's survival.

Final Thoughts: Building Trust in the Age of AI

The financial services industry is at an inflection point. What seemed impossible a decade ago, trusting algorithms with investment decisions, is now standard practice. What seems impossible today, trusting AI agents with loan underwriting and complex document processing, will be standard practice tomorrow.

But the path from skepticism to adoption isn't straight. It requires deep domain expertise, systematic thinking, precision that goes beyond "good enough," and most importantly, trust built through countless small interactions over years.

For developers and founders building in this space, the lesson is clear: horizontal AI platforms will capture breadth, but vertical AI solutions will capture the highest value workflows. The companies that win will be those that understand both the technology and the domain deeply enough to achieve the precision that regulated industries demand.

Don't try to build everything. Build something specific, build it with domain experts, and build feedback loops that turn 95% accuracy into 99.5%. That last few percent is where the billion dollar companies get built.

Q: How can people connect with you and learn more?

Ankur: The best way to reach me is through LinkedIn. You can see what I'm doing professionally and what I'm currently pursuing. If you have questions, I'm open to providing guidance and can chat anytime.

If you're interested in what we're building at Multimodal.dev and Agent Flow, reach out. We're always happy to discuss how vertical AI can transform financial services workflows.

Final words: The authentication revolution in finance isn't coming, it's here. AI agents are moving from experimental to essential, from nice to have to competitive necessity. The question isn't whether financial institutions will trust AI with critical workflows, but how quickly you'll build the solutions that earn that trust.

Related Shorts

Explore short-form videos that expand on the ideas covered in this blog.