Kaaj founder and CEO Utsav on why a $50,000 file costs as much to underwrite as a $5 million one, how to tell a weekend demo from a system that survives real volume, and the part of the job underwriters should never hand over.
An underwriter spends about as long on a $50,000 equipment finance deal as on a $5 million one. Revenue does not scale that way, which is why small- and mid-ticket paper has always been hard to write profitably. Closing that gap is why Utsav founded Kaaj. He spent more than a decade in AI and machine learning, much of it at Uber and then at Cruise, General Motors’ autonomous vehicle unit; his co-founder, Shivi, came out of credit and fraud risk at banks including American Express. Two years in, Kaaj works with more than 45 equipment finance lenders amongst other small business lenders.
In this conversation with Monitor Editor-in-Chief Rita Garwood, Utsav separates the science of a deal, the objective checks a machine can run, from the art of it: pricing, structure, cycle judgment and the relationship that brings a borrower back. He also explains where the hours actually go between an application landing and a credit decision, and how equipment finance leaders can tell a genuinely scalable AI system apart from an impressive demo. They also discuss what data lenders should insist on owning, the smallest credible AI pilot a company can run, and what a high-performing credit operation will look like three to five years from now.
Listen to the full podcast on Spotify or view it below.
This conversation has been edited for length and clarity.
Rita Garwood: Hi everyone, I’m Rita Garwood, editor-in-chief of Monitor. Welcome to Monitor’s podcast. Joining me today is Utsav from Kaaj. Can you tell us a little about yourself and what brings you to the podcast today?
Utsav: I’m the founder and CEO of Kaaj, and my background is in AI and machine learning for over a decade. A good part of that was at Uber. It was a very small company when I joined, then it blew up, and it’s a $150B public company now. At some point everyone has used Uber, so I was fortunate to be there for it. From there I went to Cruise, which was part of General Motors, and I was on the AI team that put the first self-driving cars on the road here in San Francisco.
Then my co-founder, Shivi, and I started Kaaj. Shivi spent years in credit and fraud risk at multiple banks, including American Express, so she is very close to the small business lending ecosystem. We paired that with my AI and machine learning experience. The premise was straightforward: lenders cannot profitably underwrite small-ticket or mid-ticket deals, because it takes the same amount of time to underwrite a $50,000 deal as a $5 million deal. The core economics were broken. Two years on, we serve more than 45 customers in equipment finance, and are hyperfocused on this industry.
Systems, Not Models
Rita Garwood: You spent nearly a decade building AI-powered decision systems at Uber and at Cruise. What did those environments teach you about deploying technology in cases where speed really matters, the inputs are messy and mistakes have real consequences?
Utsav: The core thing you learn in an organization like Uber, dealing with billions of trips and millions of data points every day, is that you are building a system, not just a model. As a customer, what you really care about is whether the car arrived on time, whether you were picked up and dropped off in the right places, whether the pricing is correct. Behind that sits a dispatch system, a routing system, a pricing system and a lot of other AI models. The customer does not care which model you are using.
What we learned very early, especially with the advent of AI, is that the model layer itself is becoming a commodity. You have frontier models from OpenAI, Anthropic, Google, xAI- a lot of options. So the value isn’t at the model level; it is above that. From models you get to agents — specialized orchestrators with context and memory that shape how the underlying model behaves. On top of that is the orchestration layer, an entire system of those agents, each handling something different. That is what Cruise and Uber taught me: how to build orchestration at scale, not how to stand up a chatbot.
That is how we built our infrastructure at Kaaj. Equipment finance is a very complex web of things you have to do. It is not consumer finance and it is not auto finance. It is a niche area where a single deal carries an enormous amount of complexity, and that is exactly why the previous generation of technology has not been able to solve it.
Where the Hours Go
Rita Garwood: It’s definitely something equipment finance needs — everyone in the industry knows we’ve been saying we need to move faster for a long time. Where does time actually disappear between receiving an application and making a credit decision, from where you sit?
Utsav: There are so many hoops between those two points. It’s like saying you want to drive from point A to point B, but there are intersections and objects in the way, and a lot can happen in between.
Information is missing. The intake team is swamped, with a huge pile to get through — 1,500 documents. You have to open all of those PDFs. You have to research the company. You have to make sure the names match across documents, that the information matches across documents, and that it matches what the public data sources say. There are probably 150 different data points an intake or underwriting team works through before anybody can even start making a decision.
So the actual decision-making, which is what underwriters are great at, is literally 20% to 25% of their job. The other 60% to 80% is just preparing the data in a format that makes it easier to decide. That is the time sink behind the turnaround times brokers and vendors feel. From a broker’s perspective, they send you a package at noon, and if they don’t hear back by 12:30 they get very finicky, because 30 minutes in this day and age is a long, long time. Talk to anybody and average turnaround times in this industry are in hours — sometimes days.
“The actual decision-making, which is what underwriters are great at, is literally 20% to 25% of their job. The other 60% to 80% is just preparing the data.”
Rita Garwood: That makes a lot of sense. The speed is really important, but so is the underwriting process — you have to get that right and make sure you’re not making mistakes. How do you see the difference between shortening the turnaround time and rushing credit judgment?
Utsav: Picking up the self-driving car example again: you have to detect an object. There is a cyclist on the right-hand side, and you have to detect that. At any point in an autonomous vehicle there are nearly 40,000 machine learning models giving you the ability to make the final decision about what the car’s next maneuver should be. The passenger does not care about any of those details.
It is the same for a borrower or a broker. They don’t care about the tasks underneath or what models you are using. They care about clarity on a decision in the fastest time possible. They are not looking for an approval in 10 minutes — that’s not the idea. They are looking for clarity on the next steps within the first five or 10 minutes. That is the gap we are really passionate about solving.
So the system abstracts away a lot of the model work that happens under the hood. Kaaj is really the state of the art when it comes to analyzing bank statements, at OCR and recognizing characters, but is that the underwriting system? No. That is one part of it. Kaaj does hundreds and hundreds of other tasks that prepare the packet so the underwriter can begin decisioning.
Rita Garwood: So it’s essentially helping them organize and make sense of the data so they don’t have to spend so much time doing it.
Utsav: Absolutely, and it’s very important that they can trust the preparation, not just have it. If you can’t trust what the model outputs, or what the orchestration layer outputs, you end up doing the same work again, checking and rechecking.
The way we get around that is by following the memory and context these organizations already have. A company that has been in equipment finance for 25 years has 25 years’ worth of knowledge about processing applications a certain way. It’s not as though every deal runs the same 15 steps down the happy path. Most deals have peculiarities — edge cases, rare instances. That is what lenders are really good at, because they have seen so much variety. We can mimic the policies they have written, which is the institutional knowledge they already own.
Rita Garwood: Getting back to the speed of the process: if a lender has a slow underwriting operation, how does that affect more than the customer experience? Do you see it affecting deal quality, the confidence of their brokers, operating margin?
Utsav: Customer experience is first and foremost. The average age of a small business owner in the U.S. is declining — it used to be in the late 40s, even 50s, and now it’s in the late 30s. So people’s expectation of everything they touch is just faster.
That part is visible. What’s not visible is the back end. Brokers, vendors and partners expect the same speed, because they are the ones being pinged all day by customers asking, “Am I getting this funding or not?” The easier you make that interaction, the lower the friction for everyone in the industry.
How that relates to operating margin: if you are doing something in six hours and somebody else is doing it in 16 minutes, or even six, that’s a huge gap. The broker’s next deal goes to whoever has the shorter turnaround time. Which means the quality of deal a slow lender gets is an adverse selection — they only get what didn’t get done in the first hour or two somewhere else.
Strategically, it is also just not preparing for the world we are already in. Losing relevance as a brand is the core underlying risk for companies that are set in their ways. Those processes have worked for 25 years. They may not work for the next three to five, because so much is changing under the hood.
The Science and the Art
Rita Garwood: The world is changing rapidly, and the pace of that change seems to accelerate every day. A big conversation we have been having is what AI or automation should be handling and what should stick with human judgment. How do you see that in equipment finance?
Utsav: It’s very interesting how decisions are made in equipment finance. There is a science of the deal and there is an art of the deal.
The science is the math behind the equation. Do the numbers make sense? Is this a fraud or not? Do the names line up? Do we have the right documents, the right OFAC checks, the right KYCs? That is a lot of objective analysis of the deal.
Then there is the art. An underwriter who has spent 20 years in the industry knows exactly what the boom and bust cycles are for transportation, for trucking. They know that a sole proprietor who started six months ago, but has a history of running businesses successfully before, is much more valuable than somebody who has been in business for two years. And there is the art of structuring. Okay, I’m going to approve this — but at a 9.3% or a 10.2%? How can I structure this deal? What should the payment terms be? It’s an agricultural business, so do I give them a three-month off window?
The science, though, can be automated away almost entirely, and that is what we’re seeing with these agents and architectures. That frees humans for the decision-making, the art of the deal and the relationship — with partners and with direct customers — because this is not a one-deal-and-done industry. A restaurant that buys an oven this year may come back for another one, or for a different piece of equipment. You are playing a long-term game of repeat business, which means the experience matters a lot more. Relationship building, decision-making and deal structuring are what people should focus on. The objective science is what should be automated.
Telling a Demo from a System
Rita Garwood: That makes a lot of sense — the repeatable science goes to AI, the work that needs nuance stays with people. There is a lot of AI in the market, and I feel like I hear about a new provider several times a day. How can a CEO distinguish an impressive demonstration from a system that will actually change the economics of their underwriting?
Utsav: We are in San Francisco, so we see AI companies every hour, not even every day. There are three things that are absolutely critical.
First, is this a company focused on the success of this particular industry? You don’t want a generic solution — a CRM with some AI functionality retrofitted into equipment finance. This industry requires very deep domain expertise, because it is so nuanced. The reason we don’t move as fast as other industries is that every deal is so complex; every deal in itself is a whole new system you could build. A vendor without that domain may still be valuable for one thing here or there. But from an overall systems perspective, you need somebody with deep domain experience.
Second, with the advent of AI, tools like Codex, Cursor and Claude Code, creating a demo is now a weekend’s worth of work, if not less. That doesn’t mean you can scale it, and this is why I bring up Uber and Cruise: scale does something very different to infrastructure. Going from 10 applications a month to 20,000 — or, for us, hundreds of thousands a month — changes what the infrastructure has to be, and you can’t vibe-code your way there. A lot of demos are impressive at first and have no substance, because they break the moment you see volume, the moment you see more intricacies in the product.
Third, and probably most crucial: what happens after you launch? Who do you call when things don’t work? If I call in the middle of the night, will they respond in a couple of hours? Will they customize the product the way I want it? Otherwise you’re just buying generic software — we’ve even heard of employees at equipment finance companies building their own tool as a capstone project and calling it a solution. So: domain expertise, the volume they have already handled, and the support after launch.
“Creating a demo is now a weekend’s worth of work, if not less. That doesn’t mean you can scale it.”
Rita Garwood: All definitely important things to measure. Say a company wants to get started with AI. What is the smallest credible pilot they can run, and what should leadership measure to know whether it worked?
Utsav: Monitor does a great job publishing the Monitor 100 and the Monitor 30, and we are fortunate to work with a lot of those names. The ones that do this well sit down and ask: what are my top three priorities this quarter, this year? Usually the first is time — time to decision, time to get back to a broker, entering applications without a human touching them. Then they break that down to the task level and see where exactly they are losing time and losing deals.
That is the easiest ROI to prove. You took a process that used to take an hour and 23 minutes, and now it takes two. That is an hour and 20 minutes saved on every application. Do 100 applications a month, or 500, and it becomes a real dollar figure over a year. We like starting there because it gives a very quantified view of the ROI and makes an easy proposal for a CFO to approve: bring in this technology and it’s an instant 7x on the dollars we are committing.
The next phase is much more interesting. If turnaround time drops, what does that do to the business? Does it let me process 1,000 applications instead of 500 without growing my team, without throwing bodies at the problem? That is the revenue side, the growth curve. The strong leaders we see keep an eye on the short-term cost advantage, but the best return on investment is what they are planning from a future perspective — how much do they want to grow, and what does that growth look like from a technology, application, and business standpoint. The second is far more powerful in the long run. Start with the first, because it is measurable and provable.
What a Lender Should Refuse to Give Up
Rita Garwood: I want to talk about data for a moment. Everybody is talking about the value of their data and how to use it to do exciting things with AI. When lenders use third-party models and platforms, what data and decision history should they make sure they continue to own?
Utsav: This is exactly why we advise lenders against going straight to a model provider and building on top of it. A lot of those companies are out there to get your data. Look at the enterprise agreements of the biggest frontier model companies — you are giving away years and years of history, knowledge and context to a company that may come back and build a similar product, or use the data in a way that isn’t conducive to your business.
The way we handle this at Kaaj is that we have enterprise agreements and zero data retention with every model provider we use — Google, Anthropic, OpenAI, everybody. They don’t get to see the data under the hood and do not get to leverage it or train in any way.
So those are the questions to ask. How are they hosting the data, and where? Is it only going to U.S. servers, or somewhere outside? Do they have zero data retention policies with the model providers they call? And how are you protecting your own employees from uploading documents to random model companies? We see employees do exactly that — here’s the application, here’s the credit report — straight into a model provider, which is extremely risky.
This also comes back to judging a vendor. Don’t stop at their SOC 2 policy. At a minimum you have to be SOC 2 Type II compliant to work in finance. On top of that: if you’re using any model providers, can you ensure that none of our data is hosted anywhere else, or used in training models, or used by any of the other AI models?
Three to Five Years Out
Rita Garwood: That’s so important. So three to five years from now, what will a high-performing equipment finance credit operation look like, and what work will underwriters no longer be doing?
Utsav: In our initial process we shadowed hundreds of underwriters and hundreds of intake teams to learn exactly what they do and what the jobs to be done are. The most interesting thing was that they actually don’t enjoy a lot of their work. Nobody wakes up saying, “I want to open up a bank statement. I want to key amounts into a system.” That’s not the fun part.
The fun part is the creative aspect. There is a sort of high you get from approving a deal, declining a deal, making a decision. People will continue owning that. They will move to higher-leverage, higher-order thinking and start working from a systems perspective, rather than owning a lot of minute little tasks.
Everything that is a task within a workflow will get automated away — probably not even in three years, probably in a year or so. Some of the top-performing companies among our customers have already reduced 90% of the human tasks they were doing, and most of their underwriting time now goes to the creative side of the job.
“Nobody wakes up saying, ‘I want to open up a bank statement. I want to key amounts into a system.’ That’s not the fun part.”
Rita Garwood: That sounds like it would be good for underwriters in the long run. One final question: say an equipment finance leadership team is ready to get started with AI. What is one concrete action they should take within the next 90 days?
Utsav: I want to be a bit biased and say sign up for Kaaj. Outside of that, they really need to sit down and build a strategy for the entire operation. A lot of the leadership teams we talk to have a good idea of what their teams are doing, but they don’t have a concrete, written-down policy of what exactly needs to happen.
So the first thing is to look inward before looking outward. Look into your own operations. Talk to your teams as much as possible. Figure out the jobs and tasks that may not be relevant in the future. Then start talking to external companies like us. Because if you are very clear about what you want to solve, you will find the vendor that solves it.
The problem happens when you go in blind — “I want to use AI in what I do right now,” without being clear where exactly. That is the recipe for the failures we see: “We’ve spent nine months introducing AI and it hasn’t moved the needle.” It’s because the discovery process wasn’t done correctly. So spend the next 30 to 90 days on internal discovery, then go out in the market with a very clear vision of what you want and pick the best vendor. Hopefully that’s Kaaj.
Rita Garwood: Excellent advice. Thank you so much for being on the podcast today. I enjoyed our conversation, and I hope we get to talk again soon.
Utsav: Absolutely. Thank you so much, Rita.
