A financial services company needed decades of legacy records extracted fast. The big AI platforms were too expensive. Other providers couldn't move quickly enough. Here's how a custom multi-tool pipeline delivered when nothing else could.
A financial services company was sitting on decades of legacy records locked inside 260,000 PDF pages. The documents spanned multiple time periods and included dot matrix printouts, handwritten corrections, multi-column layouts, and inconsistent formatting that varied across years and document types.
They needed the data extracted, structured, and delivered as clean Excel output — and they needed it at very short notice. This wasn't a project that could wait weeks for a lengthy scoping process or a slow delivery pipeline.
The three most obvious routes had already been considered and ruled out before the project reached us.
The large AI data extraction platforms — the ones that advertise bulk PDF processing — charge per page or per template. At 260,000 pages across constantly changing formats and layouts, the cost was prohibitive. These platforms also require upfront template configuration for each document type. With inconsistent legacy documents spanning decades, that configuration work alone would have taken longer than the delivery deadline.
Other specialist providers either couldn't turn the work around fast enough or couldn't scale to the required volume on short notice. Reliability and responsiveness were as important as capability — a provider who could theoretically do the work but needed three weeks to start was not a viable option.
Full manual data entry was accurate but completely impractical at this scale. At 260,000 pages, even a large team would take weeks. The budget and timeline ruled it out entirely.
The solution was a purpose-built extraction pipeline assembled specifically for this document set — not an off-the-shelf tool, but a layered workflow combining the right tools at each stage.
Python handled the initial organisation of the document set — sorting, batching, and preparing the files for processing in a sequence that maximised extraction consistency across the varying document types. PDF extraction tools handled the base data pull, converting raw document content into structured output. VBA consolidation brought the extracted batches together into clean, consistent Excel output, handling the formatting normalisation that extraction alone couldn't achieve. Human checks ran throughout — not processing every record manually, but targeting the exceptions and edge cases that the automated stages flagged as uncertain. Delivery happened in batches, meaning the client was receiving usable data throughout the process rather than waiting for a single end-of-project handover.
Each layer handled what it was best at. No single tool was asked to do everything.
Constantly changing formats and layouts are the core challenge in legacy document extraction. Any approach that relies on a single tool or a fixed template structure will fail when the documents don't conform to expectations — which with decades-old records, they frequently don't.
A multi-tool pipeline is flexible by design. When one stage encounters a document type it handles poorly, the human check layer catches the errors before they reach the final output. When the document set has consistent sections that can be processed reliably, the automated stages handle those at speed without human involvement.
The other critical factor was responsiveness. The pipeline was operational within hours of project start, not days. Batch delivery meant the client had working data almost immediately, rather than waiting for the entire job to complete before seeing any output. That's only possible when the person building the pipeline is directly involved and available — not routing requests through account managers or offshore teams.
260,000 pages of messy, multi-decade financial records delivered as clean, structured Excel data in 4 days. Work that would have taken a large manual team many weeks was completed in under a working week, at a fraction of the cost of the per-page AI platforms, and to an accuracy level that fully automated tools could not have matched on this document set.
This approach works well when you have large volumes of PDFs with inconsistent or legacy formatting, when per-page pricing from the big platforms makes bulk extraction uneconomical, when you need the work done fast and can't wait for lengthy onboarding processes, and when accuracy requirements mean you can't accept the error rates that fully automated tools produce on complex documents.
Common scenarios include legacy financial records, insurance documents, historical business archives, scanned invoices and purchase orders across multiple suppliers, and any document set that spans many years or originates from multiple sources with different layouts.
Whether it's hundreds or hundreds of thousands of documents — messy PDFs, legacy records, inconsistent formats — let's find the right approach for your volume, timeline, and budget.
Book a free 30-minute call →The same problem-first approach applied to different business challenges.
Book a free 30-minute call and we'll find the right approach for your volume and accuracy requirements.
Book a free call →