Ivo Duran on Automating the Document Work Behind Lending

Ivo Duran is a Product Manager at Heron Data, an AI-native company that automates the document-heavy workflows behind financial services, turning messy bank statements, emails, and attachments into data that lenders, banks, and insurers can act on.

Heron is built around a clear starting point: financial services still runs on documents, and the manual work of reading them is slow, expensive, and error-prone, even though the decisions that depend on it are high-stakes. The company uses the latest advances in AI and large language models to handle that work end-to-end, from ingesting documents to extracting and analysing the data to taking actions like updating systems or flagging risk. Its largest product automates underwriting for small-business lenders, and today more than 150 banks, fintechs, and specialty-finance firms rely on it to sit inside processes that stop without it.

What makes Ivo’s perspective interesting is where he sits, between law, operating, and building. Before Heron he was an M&A junior associate at Linklaters, ran strategy and operations at REVER AI (YC S22), and founded and led his own company, Makito Barcelona. He studied a Master in Entrepreneurship and Law at IE in Madrid after a dual degree in Business and Law, with exchanges in Seoul and Paris along the way. He has seen the document-heavy, high-stakes world of M&A from the inside, the operating side of early-stage startups, and now the attempt to automate exactly the kind of manual document work he once did by hand.

We sat down to talk about why financial services is finally ready to hand its document work to AI, what changes when a former M&A lawyer ends up building the software that automates manual review, where the technology has to earn trust in high-stakes decisions, and where human judgment still matters when the work gets faster.

Q&A

1. Heron describes itself as an AI-native company automating document-heavy workflows in financial services. What does that mean in practice?

At Heron, we sit between funders and brokers in the small business lending market, and our role is to automate the deal process from end to end. Lending has always been an intensely document-heavy business. Every deal arrives as a pile of bank statements, application forms, and email correspondence, and historically someone has had to read through all of it manually before any decision can be made. Being AI-native means we have built our entire system around removing that manual burden rather than layering automation on top of legacy processes.

Concretely, when a deal comes in, we extract and structure all the data the merchant submits, whether that sits in bank statements, supporting documents, or email threads, and we deliver it to the funder in clean, usable form. We then perform the same function on the funder side, processing everything they receive so that they can move directly to a decision. The work we replace is significant. Funders and brokers have traditionally relied on offshore teams whose sole responsibility was opening statements, transcribing the figures, and entering them into their systems. That repetitive, high-volume data entry was our wedge into the industry.

What distinguishes an AI-native approach from a conventional one becomes clear in everything that follows the initial extraction. The genuine difficulty lies in handling inconsistent document formats, multi-step workflows, and the full range of edge cases that emerge once a system is operating in production at scale. This is the layer where most tools fall short, and it is where we have concentrated our efforts. Having established ourselves through document processing, we are now expanding into fraud scoring, identity verification, and broader underwriting, with the long-term ambition of serving as a trusted infrastructure layer for small and medium business lending across the US.

2. You trained as an M&A lawyer at Linklaters before moving into operating roles at AI startups. What felt broken or too slow in that world, and what pulled you toward building?

There were two distinct drivers, and they were largely independent of one another. The first came directly out of my experience practising law, and the second was more personal, rooted in how I see myself and the kind of work I want to be doing.

On the professional side, M&A is an extraordinarily document-heavy, high-stakes discipline. We would read through thousands of pages, run due diligence, draft and review contracts, and work through every file in granular detail. In many ways it is the same shape as underwriting a business or a loan: you go into every document, you read everything, and yet you do all of that simply to extract the small fraction of information that informs a judgment or a decision. Everything that comes before that judgment is manual, repetitive, high-volume work, and it was already clear to me that this was precisely the kind of task that AI could absorb. As a lawyer, I would far rather spend my time on the judgment calls and the genuinely high-value work than on the grunt work surrounding it. At a broader level, it was also obvious that much of this entry-level legal work was destined to become obsolete, and we are seeing exactly that today. My former colleagues at Linklaters now rely on tools like Harvey and Legora, and the work we used to do by hand as juniors is simply handled by AI. 

The second driver was more personal. Around the end of 2022, when the first widely capable version of ChatGPT appeared, I connected immediately with what the technology made possible. The very first AI tool the firm rolled out to us was fairly rudimentary, but I became its single most active user worldwide, which tells you something about how compelling I found it. I could see clearly that this was going to be transformative, and I knew I wanted to be building it rather than simply using it. I have always thought of myself as a builder, and I wanted to be among the people driving this technology forward rather than waiting to adopt whatever others produced.

3. Underwriting and lending decisions are high-stakes and heavily regulated. How do you make document processing fast without making it shallow or hard to trust?

The first thing to be clear about is that speed and risk are separate questions. Underwriting and risk evaluation should always be done to the same standard regardless of how quickly the data moves, so we never treat speed as something that comes at the expense of rigour. What speed does change is competitive advantage. For a funder or a broker, being first matters enormously. If you can underwrite and return a decision before anyone else, or as a broker submit to a funder and pass on the strongest deals first, you win the business. That is why speed is the primary lever in this market, and it is the one we focus on, precisely because it is the one that does not alter the underlying risk equation.

That said, speed is worthless without accuracy, because every underwriting and funding decision rests directly on the output of our process. If we are not accurate, the consequences are not theoretical or delayed. In this industry an error is flagged the same day it happens, not a year later, which is actually a healthy dynamic. It means we operate inside an extremely tight feedback loop. If a customer identifies an issue, we own the mistake, fix it, and redeploy immediately, so it does not recur.

That combination of a tight feedback loop and a genuinely white-glove deployment model is one of our core value propositions. It is what allows us to deliver underwriting that is not only fast but also accurate and scalable at every level. In a high-stakes, heavily regulated environment, trust in the accuracy of the data is paramount, and our model is built specifically to earn and maintain that trust rather than trading it away for speed.

4. A lot of companies now claim to do AI document processing. How do you separate what works in production from an impressive demo?

It comes down to distinguishing two very different things. One is handling document extraction and data entry, taking a clean application and moving it through to an offer. That part is moreeasy to demonstrate, and it is what most vendors show you. The other is handling the messy PDF, the fifteen-step workflow, every edge case, and all the maintenance required once a system is live in production. That last stretch is the hard part, and it is where most providers fall short. We take care of it end to end.

The clearest signal in our industry is word of mouth, and our customers consistently refer us as the vendor that can actually deliver this. A common pattern is that prospects come to us from failed vendors. They had tools that could pass a straightforward PDF into the system without trouble, but the moment things got even slightly messier, those tools broke down. We are the ones who can reliably handle that complexity.

Everyone in the competitive landscape can claim to underwrite and go from application to offer in a minute. The difference is that this is high-stakes work. You are funding a company on the basis of that output, so you have to be able to trust that the data is accurate. In a demo, accuracy is easy to assert. In production, it is the only thing that matters, and that is the bar we hold ourselves to.

5. Where do you think AI matters most across financial services today: underwriting, fraud detection, onboarding, or ongoing portfolio monitoring?

The honest answer is that AI matters wherever the work is high-volume, document-driven, and unstandardised, and by that definition every one of those areas is a strong candidate for automation. Underwriting, fraud detection, onboarding, and portfolio monitoring all share the same underlying characteristics, so none of them is off the table.

In our own experience, though, the right approach is not to attack all of them at once. We deliberately started with the most urgent and most painful problem for underwriters, which was application scraping and field entry, and that remains the core of what we do today. The key insight is that this initial wedge functions as a Trojan horse. Once you have automated that first step and proven it works reliably, you earn the right to widen out into adjacent parts of the funnel. That is exactly the path we are on now, moving into fraud scoring and identity verification, with the goal of becoming the trusted layer for small and medium business lending across the US.

6. If we sit down again in a year, what do you think will look most different about how lenders and banks handle document-heavy work?

For the past few years, funders and brokers have been focused on automating the extraction and scoring of individual deals. The model has essentially been to get the data into the system, generate a score, and then have a person decide what to do with it. That has been a meaningful step forward, but it still keeps a human in the loop at the final stage of every decision.

What is changing now is the shift toward genuinely end-to-end systems that can escalate and handle each case autonomously, without that final human judgment on every deal. For the roughly 95 percent of cases that are straightforward, an autonomous agent can scrape everything, reach an informed decision, and fund, offer, or decline within minutes. The remaining 5 percent, the genuinely complex edge cases, will always need a human, and that will not change. But the proportion of work that runs without any manual intervention is about to grow dramatically.

More broadly, I think this reflects a transition from one era to the next. Up to now, the work has been about automating specific parts of a workflow and applying AI to discrete, repetitive tasks. We are now moving into the agentic era, where you rely on an AI system that handles an entire workflow without you needing to intervene at any single step. My expectation is that within a year, or perhaps even less, this will be the standard across the market rather than the exception.