Think faxes are a thing of the past? Not in healthcare. Every month this U.S.-based platform that helps patients, doctors, and imaging centers manage imaging procedures online was buried under thousands of referrals, notes, and reports. The result was hours of manual entry that slowed growth and made scaling impossible without hiring more people.
We used Amazon Bedrock Data Automation (BDA) to cut the manual effort and bring fax data into the system automatically. The same approach can help any organization that deals with scanned forms, PDFs, or other messy inputs.
Why simple automation was not enough
As the platform grew, so did the fax pile. To keep up, about 50 agents were dedicated to processing documents day in and day out. Adding more people would have been expensive and inefficient, so automation became the obvious next step.
The first attempt at this was an OCR tool from a third-party vendor, ScriptSender. It helped with some of the work, but OCR only sees letters and numbers. It cannot tell the difference between a referral that kicks off an MRI, a note describing observations, or a final report for a doctor. In healthcare, those differences matter a lot.
On top of that, faxes don’t always arrive in perfect condition. Many are low resolution, some have smudges or scanning artifacts, and others include handwritten notes. The OCR tool struggled with these cases, and staff had to step in often. So the bottleneck wasn’t just reading text on a page. The real challenge was making sense of what kind of document it was, which details mattered, and where that information belonged in the system.
That’s why moving beyond basic automation was critical. To keep up with scale, improve accuracy, and cut down on manual corrections, the platform needed something smarter: an AI-driven approach.
Choosing Amazon Bedrock Data Automation
In search of the most suitable solution, together with the customer, we considered several options. Since their entire technology stack was already running on AWS, we involved AWS engineers in the discussion. After joint consultations, they recommended trying one of their newest services, Amazon Bedrock Data Automation (BDA). At that time it was still very new, but it offered exactly the AI-driven capabilities needed to move from simple character recognition to contextual understanding of documents like fax images.
Behind the scenes, BDA doesn’t ask developers to pick a specific model. Instead, it handles the routing itself, combining foundation models, task-specific models, and prepackaged options under the hood. All we do is set up blueprints and outputs, while Bedrock takes care of the rest. Among the model families behind it are Amazon Nova, built for text understanding, and Amazon Titan, which also covers multimodal tasks.
To make this work in production, we built a serverless architecture around BDA that ensures flexibility, error handling, and security. The diagram below illustrates the solution.
Here’s how it works in practice
Every fax first goes through Bedrock Data Automation (BDA), where blueprints decide what kind of document it is and which fields to pull out. Once BDA does its job, it fires an event: success, client error, or server error. Amazon EventBridge catches that signal and passes it along. From there, messages land in Amazon SQS, which calls a Lambda function to handle the results. If something goes wrong, the message gets parked in a Dead Letter Queue (DLQ) so we can fix it and try again. Lambda then grabs the structured output from S3, applies business rules, and writes everything into the application database, using AWS Secret Manager to keep credentials safe. The end result: every fax is processed reliably, errors are easy to track, and clean data flows straight into the system.
Improving accuracy through blueprint testing
Designing blueprints was only the first step. The real challenge came when we needed to ensure that Bedrock Data Automation produced consistent and reliable results across thousands of incoming faxes. Accuracy in this context depends not only on the AI itself but also on how the prompts inside each blueprint are defined. For example, if a field description is vague, BDA may interpret it inconsistently. If the naming is too generic, results can drift.
To fine-tune this, we built a safe “sandbox” using n8n, an open-source automation tool. Think of it as a mini copy of the production system, but without the risk of touching real patient data. Inside this setup, we could send the same fax through BDA over and over again, tweak the blueprint prompts a little each time, and then compare the outputs side by side. This gave us a clear view of which blueprint definitions worked best and which ones needed more polish.
The diagram below shows how that local testing workflow looked in action:
In the sandbox we can start a test by hand or from chat, pull in the fax data from S3, and break it into smaller pieces if needed. The results get converted into structured JSON, checked for confidence, and routed into the right category like order, clinical note, or report. In the end, everything comes together in one clear view so it is easy to spot which blueprint setup gives the best results.
By pairing blueprint tweaks with this testing loop, we created a simple cycle: adjust, test, review, refine. Over time, that rhythm made classification and data extraction far more reliable before anything hit production.
Results
Before Bedrock Data Automation, every fax was a time sink. It usually took between 5 and 7 minutes to get one document fully processed. Over 90 percent of faxes had to be typed in completely by hand, and the small portion that went through an OCR tool still needed fixes.
With Bedrock in place, the story looks very different. The average time per fax dropped to under a minute. Roughly 70 percent of documents now go straight through without edits, and the rest only need quick touch-ups in one or two fields.
That shift gave the platform the capacity to process about 20,000 requests per month without hiring more people for manual review. Even better, the system is now ready to handle much larger volumes as demand grows, without costs spiraling out of control.
Modernize unstructured data workflows with AI
What worked here is not limited to healthcare. Any organization buried in scanned forms, PDFs, or image files can use the same playbook: blueprints that guide the AI, safe testing loops to tune accuracy, and a serverless setup that makes the whole system resilient. Together, these practices cut down on manual work while keeping human review in place where it is truly needed.
If your team is exploring automation for document-heavy workflows, now is a great time to see what the new generation of AWS tools can do. At ABCloudz, we bring together experience in cloud migrations, system modernization, workflow automation, AI, and beyond. The result is scalable, reliable, and cost-effective transformation that fits the needs of your business. Reach out and let’s talk about how we can help you remove bottlenecks and prepare your platform for the future.
