It is estimated that between 80% and 90% of the world’s data is unstructured1, with text files and documents making up a significant portion. Every day, countless text-based documents, like contracts and insurance claims, are stored for safekeeping. Despite containing a wealth of insights, this vast trove of information often remains untapped, as the process of extracting relevant data from these documents is challenging, tedious and time-consuming. Additionally, the inherent variability in document formats, sources and content adds complexity.
This variability requires tailored extraction approaches for each document type, significantly extending processing times. Legacy systems to address this problem are often inadequate, requiring extensive development and deep expertise in machine learning (ML). Streamlining these processes with advances in technologies like AI could drastically improve how organizations use their document data for better decision-making. Imagine the transformative potential of a system that could automatically and accurately extract crucial information from any document with pinpoint accuracy and very little effort from your business teams.
To help organizations solve this document-processing challenge, Snowflake has created Document AI, generally available soon on AWS and Azure. This exciting new feature allows teams to set up Intelligent Document Processing (IDP) workflows entirely within Snowflake. Using Document AI, key information can be extracted from documents, like invoices and contracts, and directly applied to operational workflows without worrying about scale or variability of your documents. Document AI is powered via a proprietary, built-in, multimodal large language model (LLM), Snowflake Arctic-TILT (Text Image Layout Transformer), which delivers state-of-the-art performance with exceptionally efficient and cost-effective resource usage.
With the power of Document AI, business teams can automate processes, gain valuable insights from their data and improve decision-making. The user experience of Document AI is divided into two main stages: model preparation (using a handful of documents) and inference (across thousands of documents). Neither stage requires any ML- or application-development experience.
Model Preparation Stage: In this stage, the document owner uses the Document AI user interface to create and manage a model “build.” Each build includes the documents, the questions for which you want answers (the eventual extracted data points), and the model itself — all packaged together for a specific document type or use case. The user asks the model questions in natural language and fine-tunes it through corrections as needed. With the easy-to-use and intuitive natural language interface, business users don’t need an ML or AI background to leverage the underlying model and extract information from documents. With a single click of a button, the same users can fine-tune the model by training it to their particular needs. Once the model is successfully evaluated against a handful of documents, the user publishes the model and hands over the next steps of extracting information at scale (e.g., thousands of documents a day) to a data engineer.
Inference Stage: Once the model is ready for production, the data engineer sets up an automated Intelligent Document Processing (IDP) pipeline. Starting with sourcing and loading documents, the data engineer has the option to maintain files on external stages or bring them directly into Snowflake internal stages. They then point the model at the stage of documents for the given use case. The model is called using the PREDICT function, and the results can then be further processed and shared back to the document owner or other stakeholders via dashboards, Streamlit or other applications.
Since Document AI is built on Snowflake Cortex AI, all operations run on managed GPUs, and the model is hosted in Snowflake directly. Simply evaluate the model, fine-tune if necessary and then perform at scale without the need to test, secure, deploy or upgrade GPUs. Snowflake handles all the infrastructure for you, with the option to operationalize into pipelines, exactly where your data is.
Snowflake’s Arctic-TILT, the model powering Document AI, is a Snowflake built LLM that leverages a proprietary and unique transformer architecture, tailored to understand and extract data from documents. By combining multiple data modalities, Arctic-TILT offers unparalleled versatility and performance in document-understanding tasks. Average Normalized Levenshtein Similarity (ANLS) score is a metric used to provide a comprehensive assessment of a model’s performance in handling various textual inputs. Snowflake Arctic-TILT processes documents with a 90.2 ANLS score in its latest DocVQA benchmark, beating GPT-4’s ANLS score of 88.4. That means Arctic-TILT can process a variety of documents accurately even if it has never looked at them or similar documents before — no annotation, handcrafted templates or rules required. The model can still be fine-tuned to your specific business needs by simply annotating a limited amount of documents through the Document AI UI.
The potential applications for this technology are vast — from small financial firms to manufacturing conglomerates, from invoice reconciliation to evidence discovery.
Take, for example, Northern Trust, the 134-year-old financial services company headquartered in Chicago. Using Document AI, the firm expects to greatly decrease the work involved in extracting information from financial documents for futures reconciliation. “Document AI has the potential to streamline how we extract data from financial documents, enhancing our efficiency and accuracy, ” says Robert Ismailov, SVP and Head of Information Delivery at Northern Trust. “It would allow our team to focus more on data analysis and less on manual entry.”
Today, Snowflake customers across all industries are using Document AI for a wide variety of use cases, including:
Snowflake’s own finance team has been leveraging the technology for streamlining the identification of nonstandard deal terms within order forms. “Document AI has the potential to save us hundreds of hours a month, allowing our team to shift focus from repetitive tasks to more impactful work,” says Anh Doan, Director of Revenue and Billing for Snowflake. “The possibilities for transformation are immense.”
To see firsthand the power of Snowflake Arctic-TILT and the value Document AI can bring to your organization, check out this quickstart.
Document AI is only one of the many ways Snowflake is infusing the power of AI into every part of your business. We are accelerating the pace of how organizations can put AI to work to deliver better experiences and drive efficiencies with more advanced automation, thereby achieving tangible value. Learn more here about Snowflake Cortex AI and Snowflake Copilot.
1. Source: https://mitsloan.mit.edu/ideas-made-to-matter/tapping-power-unstructured-data
The post Streamline Operations and Empower Business Teams to Unlock Unstructured Data with Document AI appeared first on Snowflake.
Cybersecurity leader SonicWall has just released their 2025 outlook, including the threats, challenges and trends that will shape the sector […]
AI is proving that it’s here to stay. While 2023 brought panic and wonder, and 2024 saw widespread experimentation, 2025 […]
Earlier this year, Snowflake signed the Cybersecurity and Infrastructure Security Agency (CISA) Secure by Design pledge. As part of that […]