August 10, 2026

For all that large language models (LLMs) are really good at – and they are great at a few things – understanding large reports generated by IBM i servers and System Z mainframes is not one of them. A Southern California company called Astera has identified this gap and developed its own AI-based tool for extracting information from this valuable resource.

Foundation AI models and LLMs are probabilistic systems, and they work best in the gray zones: Inputting words and images, and outputting other words and images that are similar to the originals that it was trained on but not exact copies. Just as humans can get a little creative (or fuzzy) in trying to call up previously established neural connections, so too do the LLMs as they reflect stuff back to us based on what “learned.”

But modern computer systems and their applications are deterministic systems, which means they (hopefully) work the same way every time. Our tolerance for downtime may be measured with five or seven or nine 9s. But the actual mathematical and logical accuracy of our applications is very close to 100 percent. Bits flip all the time, but thanks to error correction, it’s detected and fixed before we notice it. Today’s 64-bit systems are totally accurate out to about 15 digits, at which point we start to get rounding errors. Most applications don’t need that level of accuracy.

As a standard computer output, the standard green-bar report also reflect this penchant for total accuracy. Whether printed out on a line printer and sent via US Postal Service or saved as a PDF and emailed to customers, legacy reports are standard fare for modern business.

Seeing the wealth of information contained in these reports, some people have turned to AI for help. We use AI to analyze all sorts of information stored on our PCs and across the Internet, so why not point ChatGPT at the old green bar report and let her have a go?

That would be a bad idea, according to the folks at Astera.

“The enthusiasm for large language models in document processing is understandable, but applying them universally ignores the unique requirements of mainframe-generated reports,” said a spokesperson from Astera. “These systems process billions of dollars in transactions daily and support regulatory compliance across industries. A probabilistic approach to data extraction for financial services simply cannot meet the accuracy thresholds these operations demand.”

There are several major reasons why LLMs are lousy at understanding legacy reports, Astera said. For starters, the data in reports is delivered in a highly structured format.

“The difficulty is not that mainframe reports are unstructured. They are actually highly structured, but the structure is positional rather than semantic,” writes Astera’s Ayesha Amjad in a paper titled How ReportMiner Processes Mainframe Reports at Enterprise Scale, which you can see here.

The fixed-column in legacy reports deliver information with ruthless efficiency. It’s just not amenable to LLMs understanding it.

“There are no delimiters between fields, no column headers that repeat on every page, and no markup to indicate where one field ends and another begins,” Amjad writes. “If the extraction logic is off by even one character position, the entire row of data shifts.”

What’s more, the legacy reports, such as a daily FedWire report, also frequently change column layouts within the same report, which humans can adapt to easily but digital eyes cannot.

“The transaction summary has seven columns. The settlement position detail has six. The reserve summary has no columns at all, just label-value pairs,” Amjad writes. “An extraction system that applies a single template to the entire document will either miss sections or corrupt data by applying the wrong column definitions.”

Maintaining accuracy in data extraction requires being nimble enough to preserve the hierarchical relationships between the different segments of a report. LLMs don’t natively do that, but ReportMiner has been programmed to do exactly that.

Another big problem with legacy reports is the sheer size of the reports. At a large company, such as Walmart, it’s not uncommon to have end-of-day reports that run into the thousands of pages. Trying to run a report like this through an LLM would be expensive, slow, and glitchy.

Astera says it gets around these challenges with ReportMiner by processing documents as continuous text streams rather than individual pages.

“A template defined on page one applies consistently across all 1K or 10K or 10M pages because the extraction logic operates on character positions and pattern rules that repeat throughout the document,” Amjad writes. “Processing time scales linearly with document size, typically completing a 1,000-page report in seconds.”

ReportMiner specializes in ingesting legacy reports and outputting them into a standard database or data lake. After ingesting the reports, it analyzes the report structure, including the column layouts and header and footer patterns. It creates an extraction template that can be used to process the report.

Before outputting the report data in its final format, it previews the transformation and verifies that it works. After the data types are validated – numbers stay numbers, dates stay as dates – it exports it to a final destination, such as SQL Server or Oracle.

The company has been developing ReportMiner for over a decade, and counts more than 500 enterprise users. The recent emergence of LLMs and the poor experiences that some customers are having with them – particularly customers trying to migrate databases – motivated the company to highlight its particular approach.

Astera recently added an AI Agent for generating extraction templates, which it says reduces the time from 10 minutes to merely seconds. It’s also added new optical character recognition (OCR) engines that deliver better speed and performance on poor quality scans. For more information, check out the company’s website at www.astera.com.