How a multi-model AI pipeline replaced manual NetSuite data entry for a ready-mix concrete producer, with no added headcount
The Challenge
Every day, staff at a regional ready-mix concrete and building materials producer reviewed as many as 1,500 unstructured PDF supplier tickets by hand, then keyed the data into NetSuite line by line. Two full-time employees did nothing else. During peak season, the company brought on temporary and part-time staff to keep up.
The math did not work going forward. Leadership planned to extend the same process to a sister operating company, and every increase in ticket volume meant another increase in headcount.
The data itself was the hard part. More than 30 distinct ticket formats were in circulation across the two businesses, with inconsistent vendor name formatting, plant address variations, and differing ticket-number prefixes. Extracting text from a scanned ticket was only half the job. Each value had to be translated into the exact vendor, plant, and product records NetSuite expected, or the downstream entry would be wrong.
The client did not have the internal engineering capacity to build this. They were introduced to AIM through Bridgepoint Consulting, an Addison Group company already leading their NetSuite implementation, and selected AIM for the practicality of the proposed approach and the technical solution behind it.
What Had to Be True
Three risks defined the engagement from the start:
- Extraction accuracy high enough to trust. If the output required checking, the manual work would simply move rather than disappear.
- Correct entity resolution. Misidentifying a vendor, plant, or product would compromise every downstream field on the record.
- Economics that scale with volume, not headcount. The solution had to absorb growth, including new operating companies, without adding people.
The Approach
AIM started with a rigorous proof of concept rather than an architecture decision. The team cataloged more than 30 ticket formats, evaluated 8 or more large language models alongside specialized document intelligence platforms, and tested the candidates against 300+ real tickets.
The results reset the plan. Neither a purpose-built document intelligence platform nor a single general-purpose model held up against the variety of ticket layouts, and any individual model would occasionally hallucinate a value with full confidence. That finding drove two design decisions that shaped the final system:
A multi-LLM consensus approach. Rather than trusting one model, the pipeline runs extraction across multiple vision-capable models and reconciles their outputs, so a single model’s error does not become a database record.
Client-maintained cross-reference tables. Vendor, plant, and product mappings live in reference data the client can update themselves, with hot reload and no redeployment. Accuracy improves as the business onboards new suppliers, without an engineering ticket.
AIM also built a realistic operating cost model during the POC so the client could make an informed go or no-go decision on the full build before committing.
The Solution
An event-driven automation pipeline that ingests supplier tickets, extracts and validates the data, writes it into NetSuite, and surfaces only anomalies for human review.
Architecture
- Azure Container Apps with KEDA auto-scaling workers
- Azure Service Bus queues and Blob Storage
- SQL application journal for full run and ticket traceability
- Multi-modal LLM OCR across several frontier vision models, routed through an OpenRouter API gateway
- NetSuite integration via NetSuite REST API and SFTP
Operations Portal
A custom React application gives non-technical staff direct control of the system: dashboard, ticket browsing, run history, ticket and run diagnostics, issue reporting, supplier and reference-data management, and role-based access control. The client operates and governs the pipeline themselves rather than depending on a black box.
The Results
In its first 30 days in production, the pipeline processed 24,058 tickets, with more than 80% completing straight through on a scheduled run with no manual intervention. Staff moved from full-time data entry to exception handling only.
Actual operating cost came in at $0.0585 per ticket, or roughly $16,832 per year at current run-rate volumes. Infrastructure cost now scales with ticket volume instead of headcount, saving an estimated 3,600 FTE hours annually and requiring zero additional FTEs as the business grows.
The engagement ran across four phases from October 2025 to June 2026 and landed within 2.5% of the original budget.
The proof of the design came before the project even closed. AIM extended the same pipeline to the client’s sister operating company with no significant redesign, and both are live in production today.
“We engaged AIM Consulting to create a custom solution for us to handle the processing of thousands of receiving tickets each week, autonomously. The manual process of scraping the data off the tickets took about 80 hours each week involving 2 full-time employees.
We wanted a solution that would allow us the ability to scale for future growth but not increase head count. Going in we knew a solution using AI could be built, but being on the forefront of the technology, we weren’t sure of the issues that would be encountered or what the final solution would look like.
AIM never had any doubts. They overcame our challenges positively and finalized a product within the allotted budget. We have not been disappointed with the final product. It has allowed the other two employees to pursue more engaging, higher priority duties. We were so pleased that we had AIM implement the same framework for the solution in one of our other companies.”
Chief Financial Officer, Ready-Mix Concrete and Building Materials Producer
What This Enables
The client now has a repeatable automation pattern rather than a point solution. Near-term opportunities include extending the pipeline to other manually processed supplier documents such as invoices and purchase orders, automatic NetSuite synchronization, and adaptive fine-tuning as new vendors and formats are onboarded.
Facing the same problem?
If your teams are keying unstructured documents into an ERP by hand, the constraint is rarely the technology. It is knowing which approach will actually hold up against your real documents. AIM starts with evidence, not assumptions.
