Document Parsing Done Right: Speed, Accuracy, Reliability

I’ve recently been writing about how to introduce Artificial Intelligence in the highly regulated banking industry. The most straightforward application of AI is to automate manual processes, improving efficiency. However, accuracy and reliability are paramount. A clear example is in the complex process of commercial real estate (CRE) underwriting, where data must be extracted from various documents for financial projections and risk analysis. This is commonly known as 'document parsing.' Instead of spending hours manually collecting data from documents, AI can handle this tedious task, allowing you to focus on making informed decisions based on the extracted data."

The Basics: OCR, AI, ML, NLP, NER. What does it mean and what’s in it for me?

While many refer to document parsing as “OCR” (Optical Character Recognition), this is inaccurate. OCR technology (now even integrated in most smartphone cameras) can quickly identify and extract letters, numbers, and symbols in an image and convert them into usable text. However, they cannot preserve the original layout (for example data in a table) or understand the context in which the words reside

Documents, such as a CRE Offer Memorandum (OM) often contain structured and unstructured data, such as tables, figures, and annotations, that need to be interpreted in a very specific way. It's not just about reading the text, it's about understanding the context, relationships between data points, and formatting it correctly for financial analysis. Data extracted from a financial statement must be parsed and arranged into rows and columns, ensuring that revenue, expenses, liabilities, and other figures are accurately mapped to the appropriate categories. This is why OCR is only the first step to document parsing. Once the text is extracted, it needs to be processed by other Artificial Intelligence (AI) models, each responsible for specific tasks:

• ML (Machine Learning) is the broad technology that learns from data. The more data that it processes, the better the model becomes
• NLP (Natural Language Processing) is a subset of AI that focuses on processing human language. For example, it can turn a description of a CRE property in plain text, into a unit mix table.
• NER (Name Entity Recognition) is a specific task within NLP aimed at identifying key entities within text. For example, in the sentence "John works at Google in California," NER would recognize "John" as a person, "Google" as an organization, and "California" as a location.

While AI and automation provide speed and efficiency, they don’t guarantee accuracy. As I often explain, AI is science, not magic, so while some models can extract accurately specific data from specific file types in specific formats, expecting consistent results across all high variability of file types and formats is unrealistic.

To understand the complexity, take a rent roll document for example. It might look standard but might have hundreds of different variations derived from factors such as file types (Microsoft Excel, PDF, image, etc.), is the rent roll part of a larger document (such as an appraisal), number of columns, existence of subtotals and summary table, number of pages, number of rows per unit, rent amounts (monthly or annual), the need to aggregate columns to get the monthly rent (for example when the unit is subsidized), are utilities included in rent or not, etc. You get the point.

To add this, errors are magnified by the compounding effect. If you need to parse 5 data points from a document each using an individual model at a 90% accuracy, the ability to get ALL 5 data points correctly is only 59% overall accuracy (0.9⁵). This reduces to 32% overall accuracy with 80% individual model accuracy, etc.)

The Challenge: AI Without Oversight

Some document parsing services focus solely on technology and speed and avoid human input and validation. This allows them to deliver results to users fast and at a low cost. While this approach may indeed save time, it often comes at the cost of both accuracy and file format restrictions. When data is extracted solely by AI models and immediately sent back to the user, errors are inevitable. Whether it’s misreading a character or misunderstanding the context of a sentence, the consequences of these mistakes can range from minor inconveniences to critical errors in data integrity. This approach is like throwing ingredients into a microwave to prepare a meal. It’s quick, but the quality may leave much to be desired. The meal may not taste good or offer the nourishment you need. Similarly, without human validation, AI in document processing can serve up results riddled with inaccuracies, potentially jeopardizing a business's data-driven decisions.

To try and address the lack of quality and inaccuracy, they either restrict file types, support only certain templates, burden the user with mapping the AI inputs or output themselves or fixing errors, and finally, correcting errors themselves.

Why Blooma's Approach Sets the Gold Standard in Document Parsing for Commercial Real Estate Lenders

Blooma operates in the highly regulated banking industry, and particularly the complex process of CRE underwriting, where efficiency gains are a must but even the smallest mistake is intolerable. Blooma’s Supervised Intelligent Document Processing (Supervised IDP) refers to a specialized approach where a blend of AI technologies and automation tools is combined with the contextual understanding, subject matter expertise, and oversight of a human-in-the-loop, to ensure accuracy and quality, along with efficiency gains in document processing.

Blooma's Human-in-the-Loop: human intelligence combined with artificial intelligence for speed AND precision

At Blooma, we understand that accuracy in OCR and document parsing is paramount, especially for businesses handling sensitive financial documents, contracts, or other data-heavy materials. This is why we have adopted a human-in-the-loop approach, where AI is not the final arbiter but a highly efficient assistant to our expert human analysts. This method combines the strengths of AI with human oversight to ensure that the results delivered to our users are not just fast but also highly accurate and reliable 100% of the time

Rather than relying solely on AI to handle the entire parsing process, we equip our human experts with a complete “toolbox” consisting of AI, OCR tools, data mapping widgets, and scripts. AI is used to automate and speed up the more repetitive tasks, while other tools are used for complex formats and document types where either the OCR or the AI accuracy is insufficient. This collaborative approach ensures that the right AI model is used for each specific task and that the AI’s errors are caught and corrected before any data is returned to the user, providing both speed and accuracy without compromise.

While human input and validation introduce a slight delay compared to completely automated systems, our AI-enhanced workflow significantly reduces the time our analysts spend on mundane tasks. By using AI as a tool to streamline their work, we empower them to work faster without sacrificing precision.

In fact, Blooma’s system is designed to get smarter over time. As our data analysts interact with the AI, correcting and validating its work, the machine learning models learn from these corrections. This continuous feedback loop means that the system becomes more accurate with each use, further reducing the chances of errors and speeding up the validation process over time.

Serving the Banking Sector: Why Accuracy is Non-Negotiable

The stakes are even higher when it comes to automating loan origination and portfolio monitoring for financial institutions. Blooma understands that banks require absolute accuracy in every document they process. Whether it's extracting crucial financial data from loan applications or monitoring complex portfolios, even a small error can have a significant impact on decision-making, compliance, and risk management.

Blooma tailors its models and tools based on the type of document being processed. This level of customization ensures that each document is parsed with the highest level of accuracy possible, considering its unique characteristics. For example, some PDF documents may not be correctly read by standard AI models. In such cases, Blooma employs a dedicated app or specialized model to first convert the document into a format that the AI can effectively process.

Multiple Approaches for Enhanced OCR and Parsing Accuracy

One size doesn’t fit all when it comes to document parsing, especially in the financial industry. Blooma employs multiple OCR approaches, including text extract and machine vision, depending on the document type and its complexity. Text extracting models are excellent for extracting structured data, while machine vision works better for documents that require image-based analysis. The same goes for using different AI models for each particular task: ML, NLP, NER, etc. By using the right tool for the job, we ensure that even the most complex documents are processed with precision.

Additionally, our analysts are equipped with specialized tools and models that adapt in real-time to handle document complexities. This does not only help streamline workflows but also enhances the accuracy of the data being extracted, especially when dealing with documents that have unusual layouts or require contextual understanding beyond what AI can interpret on its own.

A Tailored Approach for Complex Documents

Not all documents are created equal. Some are straightforward, while others are complex and filled with nuanced data that requires contextual understanding. This is where Blooma’s human-in-the-loop approach truly shines. While AI can handle standard forms and documents with relative ease, it can struggle with complicated language or non-standard layouts.

In cases where AI might stumble, our data analysts step in to ensure that every piece of information is parsed correctly. This is particularly important in financial documents, and rent roll, where a single mistake can lead to serious consequences. Blooma’s system is built to handle these complexities, ensuring that users can trust the data they receive.

Blooma’s Approach is a Path to the Future

While AI keeps improving and, in the future, will get good enough to perform tasks unsupervised, financial institutions today still can’t rely on it for lack of consistent results and accuracy. For businesses that require precision in their document processing workflows, Blooma is the clear choice. We’re not just automating; we’re enhancing human capabilities to ensure that every document processed is accurate, reliable, and ready for action.

In the end, it’s not just about speed. It’s about delivering results you can trust. And with Blooma, that’s exactly what you get—a perfectly crafted solution that ensures both quality and efficiency every time.

Consider Blooma as the bridge between the AI of today and the AI of the future. Let us be your guide on a safe journey into the future of AI.