Across manufacturing, distribution, logistics, and healthcare operations, the most persistent manual bottleneck sits at the boundary between external documents and internal databases. Purchase orders, invoices, delivery notes, and itemized bills arrive as mismatched PDFs, scans, and faxes, while staff re-key them line by line into an ERP.
Vendor pitches treat AI document processing as drop-in OCR. Production engineering write-ups show something more nuanced. Below is a teardown of two detailed public deployments published on the AWS for Industries blog in late 2026: Oldcastle APG (B2B purchase order entry) and WellRithms (itemized medical bill preparation). Neither company is an Insonl client; both are third-party public cases worth studying for their concrete architecture choices—and for the operational limits their numbers leave unspoken.
Case 1: Oldcastle APG — automating B2B order entry into Infor M3
In a 13 November 2026 AWS for Industries post co-written with Oldcastle APG architect Avdhesh Paliwal, the North American architectural products and construction materials supplier detailed its order-intake bottleneck. More than 200 customer service representatives (CSRs) were manually entering roughly 13 million order lines per year into Infor M3, spending four or more hours per person per day transcribing multi-page PDFs (up to 100 line items each), faxes, and scanned images in highly variable customer layouts. Manual keying caused peak-season backlogs and introduced quantity, SKU, and pricing errors that flowed into fulfillment and invoicing.
Two developers and one architect built and deployed an event-driven pipeline in eight weeks on Amazon Bedrock using Anthropic’s Claude Sonnet 4.6. The pipeline runs in six stages:
- Ingest: Customer emails with attachments (PDF, image, HTML, or text) are captured via Amazon SES and Microsoft Power Automate and stored in Amazon S3.
- Extract: Amazon Bedrock parses documents into structured fields—customer, PO number, ship-to address, line items, quantities, and pricing—handling English and French purchase orders from Quebec in the same pipeline without extra configuration.
- Match: Extracted values are resolved against Oldcastle APG’s master customer and product records in Amazon Aurora PostgreSQL via fuzzy similarity search.
- Verify: CSRs check matched orders in a verification dashboard displaying extracted lines beside the original PO, with ranked suggestions for uncertain matches.
- Create: Confirmed orders are written into Infor M3 automatically with a full audit trail.
- Learn: CSR corrections are saved and reused so repeat orders from the same customer skip the AI step entirely, running faster and cheaper over time.
In the first two weeks of production, Oldcastle APG processed 1,100 orders and more than 10,000 order lines—redirecting roughly 170 hours of manual effort—with 95% of lines fully automated and 5% flagged for CSR review, at an AI processing cost of less than $0.04 per page.
Case 2: WellRithms — preparing complex medical bills for expert review
In a 5 October 2026 AWS for Industries post, medical bill review company WellRithms documented a harder input profile: itemized hospital and clinic bills with inconsistent layouts, poor scan quality, and provider-specific billing conventions. WellRithms already applied downstream analytics and physician-informed review logic at the line-item level, so raw OCR text could not feed its system directly. Documents had to be interpreted, structured, normalized, and validated first.
Adapting the open-source AWS Generative AI Innovation Center IDP Accelerator, WellRithms built a tiered processing pipeline rather than pushing every page through one model:
- Clean digital bills route to traditional OCR extraction.
- Low-quality scans and layout-heavy itemized statements route to LLMs and vision-language models (VLMs) on Amazon Bedrock.
- The engineering team continuously benchmarks extraction methods against real billing documents on accuracy, consistency, and cost to assign document classes to tiers.
During its validation period across 2,820 pages and 98,377 extracted lines, WellRithms cut manual effort from roughly 8 hours per bill to 15–20 minutes of expert review alongside an average AI runtime of 70.1 seconds per bill (described as roughly 30x faster processing capability). End-to-end client turnaround fell from up to 5 business days to up to 1 business day (5x faster). Measured AI throughput reached 1,167 pages per hour (3.1 seconds per page), an estimated 9.5x capacity buffer—explicitly footnoted as assuming continuous AI processing time and excluding human review and quality assurance.
Side-by-side teardown
| Dimension | Oldcastle APG (third-party, not an Insonl client) | WellRithms (third-party, not an Insonl client) |
|---|---|---|
| Input profile | Variable B2B POs (PDF, fax, scans, HTML, text; up to 100 lines/PO) | Itemized hospital/clinic medical bills (clean PDFs to degraded scans) |
| Extraction architecture | LLM extraction + Aurora PostgreSQL fuzzy match + repeat-order cache | Tiered routing: traditional OCR for clean bills, LLMs/VLMs for complex scans |
| Human role | Side-by-side CSR verification UI; 5% of lines flagged for review | 15–20 min expert review retained per bill for clinical and billing judgment |
| Published results | 95% line automation; 10,000+ lines in first 2 weeks; < $0.04/page AI cost |
Turnaround down from up to 5 business days to 1; ~8 hrs prep to 15–20 min review |
| Unpublished details | Full-year accuracy across 13M lines; silent-error rate; total engineering spend | Field-level precision/recall; poor-scan error rates; total build and cloud spend |
Four engineering patterns synthesized from these two architectures
Oldcastle APG and WellRithms built two distinct architectures for different document domains. Synthesizing both public write-ups highlights four engineering patterns worth considering when designing a document pipeline:
1. Strict data contracts and master-data matching
Pulling text out of a PDF is only half the problem. Production accuracy depends on enforcing a strict output schema and resolving messy vendor wording against canonical records inside your ERP or analytics warehouse. Oldcastle APG did not ask an LLM to invent ERP SKU codes; it extracted structured line items and ran fuzzy similarity search against Aurora PostgreSQL. Without clean master data and deterministic lookup logic, AI extraction just pushes bad records into accounting faster.
2. Tiered routing and bypass paths
Sending every page to a frontier vision model wastes money and slows queues down. WellRithms keeps conventional OCR on clean digital documents and reserves LLMs and VLMs for irregular scans. Oldcastle APG goes further with its “Learn” stage: verified mappings for recurring customers bypass AI extraction on repeat orders. Pragmatic generative AI engineering calls a model only where layout variability makes rules brittle.
3. Exception handling and side-by-side human review
Neither team waited for 100% straight-through automation. Oldcastle built a verification dashboard showing extracted values beside the original PO so CSRs can clear the flagged 5% quickly. WellRithms retained 15–20 minutes of domain-expert review per bill because payment decisions must withstand scrutiny.
Insonl took a related pre-publication screening approach in a different text-heavy domain on Review Nguoi Yeu. As outlined in our moderation write-up, submitted Vietnamese posts run through pre-publication AI screening for harmful content, PII, and safety risks, with a design allowing review of flagged content rather than treating automated detection as infallible. In document pipelines like Oldcastle APG’s, the verification UI serves a similar role as the control loop for uncertain cases.
4. Continuous evaluation on real documents
WellRithms continuously benchmarks extraction tiers against real bills on accuracy, consistency, and cost. Oldcastle advises scoping a proof of concept with 50 to 100 real orders from your highest-volume document type. Before any pipeline writes into a production ERP, assemble a labeled evaluation set from messy historical documents so every prompt, schema, or model change is measured before release.
What the published numbers leave out—and where document AI fails
Operations leaders budgeting a document pipeline should keep four caveats in mind:
- Early slices are not steady-state scale. Oldcastle APG’s 95% rate covers 10,000+ lines in the first two weeks of production—a fraction of 13 million annual lines. WellRithms’ 2,820-page / 98,377-line metrics come from its validation period. A rollout risk to plan for is that long-tail vendor templates can raise exception rates as coverage widens beyond the initial sample.
- Compute speedups are not end-to-end SLA speedups. WellRithms’ 30x processing headline compares ~8 hours of manual prep to 15–20 minutes of review plus 70.1 seconds of AI runtime, whereas client turnaround improved 5x (5 business days to 1). That 5x improvement is meaningful—just do not promise a 30x SLA cut when human QA and batch schedules stay in the loop.
- Per-page inference cost is not total cost. Oldcastle’s
< $0.04per page covers AI processing, not developer time, ERP integration, verification UI upkeep, or production monitoring. - Silent errors are not disclosed. Neither write-up publishes the error rate hidden inside auto-approved lines. Production teams need periodic random audits of auto-approved outputs, not just exception queues.
Document AI also does not pay off when structured intake is achievable. If major suppliers can submit EDI, XML, or web-portal orders, deterministic ingestion beats probabilistic extraction. If monthly document volume is low, or if your underlying enterprise software lacks clean SKU tables and reliable APIs, fix the core system first.
Start with 50 real documents, not a vendor contract
Before buying a document-processing platform, pull 50 to 100 real purchase orders or invoices, map every ERP lookup and write step, and check where deterministic code ends and model extraction begins.
Insonl offers a free AI workflow audit for operations and engineering teams. We inspect your document flow and systems of record, then hand back a written workflow map, honest recommendations, and a fixed-scope estimate. Follow-on Map, Build, or Hold work is optional.
Sources
- AWS for Industries (13 Nov 2026), How Oldcastle APG Automates 13 Million Order Lines with Amazon Bedrock: https://aws.amazon.com/blogs/industries/how-oldcastle-apg-automates-13-million-order-lines-with-amazon-bedrock/
- AWS for Industries (5 Oct 2026), How WellRithms achieved 30 times faster bill processing with AWS: https://aws.amazon.com/blogs/industries/how-wellrithms-achieved-30-times-faster-bill-processing-with-aws/
- AWS Solutions Library Samples (GitHub), Accelerated Intelligent Document Processing on AWS: https://github.com/aws-solutions-library-samples/accelerated-intelligent-document-processing-on-aws
- Insonl, Free AI workflow audit
- Insonl, Enterprise Solutions (ERP, CRM, BI)
- Insonl, AI Moderation for Vietnamese UGC on Review Nguoi Yeu
