Why existing approaches fall short

I’ve long had a personal connection to this problem, experiencing it first hand as an early engineer at Brex in 2018 just as we were launching the initial corporate credit card product. As a startup taking on Amex, we knew our survival would come down to having the best product on the market. One problem that kept coming up: no matter how good our product was, employees were still missing expense reports due to the dreadful process of manually collecting and uploading receipts.

Our goal was to build a magical experience for employees, by parsing receipts and automatically matching them to the correct expense. Sounds straightforward, right?

Turns out, since users were uploading these receipts in real time, there were an infinite number of edge cases to consider. Millions of vendor formats aside, the receipts could be upside down, blurry, crumpled up, covered in coffee stains, have handwritten tip amounts, and so much more.

We tested nearly every option on the market, from traditional OCR vendors to specialized ML solutions. Some of them seemed promising at first when we tested them on a handful of examples. However, our team had enough experience to know that while demos are easy to build and quick to inspire, the real problems awaited us in production at scale. Despite our best attempts and multiple members of our team tediously annotating hundreds of complex examples and edge cases we expected in production, we never achieved a level of accuracy we could put in front of customers.

Plus, as an ambitious company built on modern software from the ground up (we even built our entire card processing infra in-house!), we came to the realization that we needed full control and flexibility over the end user experience. Many of our requirements, such as low latency, multi-language support, and custom data fields, simply couldn’t be met by the “one-size-fits-all” approach that comes with off-the-shelf solutions.

So we decided to build a large portion of it in-house, quickly becoming one of our most complex engineering projects spanning months of implementation, iteration, and maintenance. We used every trick in the book, combining ML models with custom code and even an in-house regex rule and heuristic scoring system to build the foundational data infrastructure necessary. Only then, with the proper regression testing and performance monitoring in place, were we able to begin building the user experience on top. We had a very talented engineering team, and in the end, we shipped that magical feature we envisioned. For many years, it consistently showed up as one of the main reasons customers would give Brex a 10/10 NPS score.

But six years later, the amount of code we had to write and maintain still haunts me. We added new models to keep up with our growing data stream, tuned heuristics every time edge cases popped up, and deprioritized new feature requests when engineers simply didn’t have the bandwidth.

Not every company can afford to do that, and even if they can, they shouldn’t. Ambitious teams want to focus on innovating, not data plumbing and fire-fighting edge cases. And businesses shouldn’t have to choose between speed to market vs. building something in-house that becomes a competitive advantage.

How Extend works

Legacy OCR, IDP, and point solutions have existed in the market for a long time. If you can get away with using one of them, you should — we’re probably not the right choice for you (in fact, we make that recommendation all the time on customer calls).

But often times, as we learned at Brex, those legacy solutions fall short because:

1. They lack the capability to handle your complex data and requirements

2. They cannot be customized to power the unique experiences that teams envision

Our customers are ambitious companies trying to innovate, leverage the latest advancements in AI, and don't want to be tied down by off-the-shelf options. So they decide to build it in-house — and while impressive demos are easy to build with LLMs, there’s a world of difference between that demo and confidently deploying an enterprise-ready, production use case.

Extend closes that gap by providing a platform that accelerates in-house builds to production-level confidence in days, not months. Companies use our toolkit to train, deploy, and monitor an in-house AI workforce for processing their unstructured documents. To do this successfully at scale, AI agents need to be onboarded onto your team, taught the intricacies of your data and workflows, know when to ask for help, and get feedback to consistently improve — much like any other employee.

Having now deployed robust AI workflows into production on mission-critical use cases at startups and Fortune 500 companies alike, we’ve seen first-hand both the opportunity at stake, as well as lessons learned from landmines waiting around the corner:

• Documents are complex, varied, and full of ambiguity. Giving teams the right set of tools and empowering non-technical domain experts to work together with engineers is how you can drive impact quickly, instead of spending months iterating and fire-fighting with edge cases.

• AI systems are indeterministic and can fail in unexpected ways. Proper guardrails, interpretability, and human oversight are necessary to confidently deploy into production.

• Data complexity, driven by customer demands and growing data streams, only increases over time. Self-improving systems that constantly learn and adapt are the only way to keep up.

Throughout all of this, one lesson has become very clear: OCR is dead. The problem is no longer “can we extract text from a PDF?” That’s table-stakes.

Rather, the problem becomes: "how can we effectively teach AI models with PhD-level intelligence the intricacies of our documents, business, and workflows in a way where they can drive business impact?"

That’s what Extend solves, and we’re the only ones who do it.

The future is here

Extend is already enabling several ambitious companies to drive business impact from their documents.

• Real estate enterprises are helping families move into homes faster with fleets of agents automating real estate transactions across all 50 states

• Fintechs are enabling their customers to pay and get paid faster with embedded agents parsing financial documents in real-time

• HR and payroll platforms are empowering workers to onboard and get approved for jobs faster with agents verifying education and employment documents

• Procurement platforms are racing ahead of the competition with agents ingesting sales documents to surface data insights

• Healthcare companies are driving better patient outcomes with agents trained by professional nurses to surface medical insights

And this is just the start. As large language models continue to get better, they’ll begin to understand the most complex unstructured data in ways that even humans can’t.

Organizations will have hundreds (or thousands) of specialist agents looking at every piece of unstructured data, connecting the dots, and surfacing insights we never even knew to look for. And we’re excited to help the world get there.