If you have used ChatGPT or Claude, you know they can type an answer back to you. But what if an AI could do things? What if it could go into your company’s vast archives, find things, and fix them?
This is no longer science fiction. It’s how major financial institutions are tackling massive data challenges today.
Huntington Bank, a major US bank, faced a daunting task: sifting through over 400 million documents accumulated since 2015. The goal? To find and remove sensitive customer information, like account numbers and Social Security numbers, to meet strict compliance rules. Doing this manually would take years. Imagine a detective trying to read every single page of every book in a library the size of a city block, by hand.
But by building a smart system using Amazon’s cloud services, they cut that time from years down to just a few months. And here’s the kicker: it cost them only about 5% of what they initially budgeted.
The Challenge of Scale
For nearly a decade, Huntington’s documents lived on their own servers, on-premises. As of 2025, a push for compliance meant they had to process this mountain of data. The problem? These documents weren’t all the same. Some were scanned papers, others digital files. The solution needed to be flexible enough to handle many types of files and fast enough to chew through millions of documents daily.
Most people think of AI as a chatbot, good at talking or writing. The real power, however, is in AI’s ability to act.
What Most People Think AI Does vs. What It Can Do
A common misconception is that AI is just for generating text or answering questions. This is true for tools like ChatGPT, which excel at understanding and creating human-like language. You ask it something, it gives you words back.
But a newer wave of AI is about doing. Think of it less like a helpful assistant who writes reports and more like a highly skilled technician who can perform complex tasks automatically. This is where AI moves from conversation to action.
Huntington’s project needed AI to perform a specific, critical action: redaction. This means finding sensitive data and either deleting it or covering it up so it can’t be read.
How AI Can Find Sensitive Data
The core of Huntington’s solution involved a service called Amazon Textract. Think of Textract like a super-fast, super-accurate OCR (Optical Character Recognition) system, but with a brain. OCR turns images of text into actual text. Textract does that, but it also understands the structure of documents – it knows what’s a form, what’s a table, and importantly for Huntington, it can identify specific types of information.
For example, if you show Textract an invoice, it can pull out the vendor name, the amount due, the due date, and even the invoice number. For Huntington, it could identify Social Security numbers, bank account numbers, and other personal identifiers. It’s like having a team of paralegals who can instantly read and tag every piece of information in a million documents.
The accuracy requirement was high: over 95%. This means that out of 100 pieces of sensitive data, the system had to correctly identify at least 95 of them.
Making It Happen at Speed
Just finding the data isn’t enough when you have hundreds of millions of documents. You need a system that can process them all quickly. This is where AWS Step Functions comes in.
Imagine you’re baking a complex cake. Step Functions is like the ultimate recipe planner that not only lists the steps (mix dry ingredients, add wet ingredients, bake) but also manages them perfectly. If one step takes too long, it knows to wait. If a step fails, it can retry or flag it for you to look at.
For Huntington, Step Functions orchestrated the entire process:
- Move Documents: Get the documents from their old on-premises storage to a secure cloud storage location called Amazon S3. This involved securely transferring over 400 million files. They used services like AWS DataSync and AWS Direct Connect for this.
- Detect Sensitive Data: Send each document to Amazon Textract to find and tag sensitive information.
- Redact Data: Once identified, use tools to cover up or remove the sensitive text from the document.
- Store Redacted Documents: Save the cleaned-up versions.
- Sync Back: Move the redacted documents back to their on-premises storage.
The key to speed was processing many documents at once. AWS Lambda, a service that lets you run small pieces of code without managing servers, was used to handle individual document processing tasks. Step Functions, using a feature called a “map state,” could run thousands of these Lambda tasks concurrently. It’s like having thousands of chefs working in parallel in the kitchen, each making one part of the cake, all coordinated by the master planner.
Why this matters: Sequential processing, where you handle one document after another, would have taken years. Parallel processing, handling many at once, shrinks that timeline dramatically.
They monitored this massive operation using Amazon CloudWatch, a service for tracking performance and errors. This allowed them to see if any Textract jobs were being slowed down (throttled) and adjust the number of concurrent jobs to stay within service limits while keeping throughput high.
Redacting the Data
After Amazon Textract identified sensitive fields and their locations, the next step was to actually remove them. This is where tools like Python libraries (e.g., PyMuPDF) or image editing software come into play, but automated through the workflow. The system would then take the identified sensitive areas and black them out or remove them.
The catch: While AI is good, it’s not perfect. Textract provides “confidence scores” for its detections. If the score is low, it might flag the item for a human to quickly review before redaction. This hybrid approach ensures high accuracy.
The final redacted files were then sent back to Huntington’s on-premises systems, completing the loop.
The Results Speak Volumes
The outcome was remarkable. Huntington Bank achieved:
- Speed: Processing around 10 million documents per day. This reduced the estimated multi-year project to just a few months.
- Cost Savings: The total cost was approximately 5% of the original estimate. Automating a massive manual task yields huge savings.
- Accuracy: Redaction accuracy exceeded 95%, meeting stringent compliance requirements.
This project shows that AI isn’t just about chatbots. It’s about building powerful, automated systems that can perform complex, high-volume tasks. For businesses managing vast amounts of data, this kind of AI-driven automation is becoming essential for compliance, security, and efficiency. Huntington plans to use this same framework for future needs, like handling data during mergers and acquisitions.
*The agent syncs via DataSync to Amazon S3, then Step Functions orchestrates the process before syncing back to on-premises data stores.*Sources: aws.amazon.com