Manual document entry eliminated across six plants — 85+ specialists freed for work that actually needs them
How a major metallurgical group automated incoming document handling, quality certificate processing, and reconciliation acts — and stopped 85 people from spending their days typing the same data twice.
Manufacturing · Industrial group · AI Server · Orchestrator · Robot · 10-min readAbout the customer
A major metallurgical group — six production plants across more than 30 regions, supplying over a million tons of pipe products annually. Procurement, logistics, quality control, and finance each generate continuous document flows that need to be processed, validated, and posted into accounting and ERP systems.
The group began automating in 2020, starting with 16 high-ROI processes across document handling, accounting, and contractor interactions. Over the following years, the team built an internal competency centre of 17 specialists — analysts, architects, developers, testers, and support — managing the full automation lifecycle from process analysis through production maintenance.
The group had been automating since 2020 on a different platform. In 2023, they ran a structured platform evaluation and selected Primo — moving their full automation estate across logistics, procurement, HR, and finance to the new platform and expanding from there. By that point the programme covered 75+ processes.
Every document meant a person reading it and typing it — across six plants
Manufacturing at scale generates documents constantly. Incoming goods arrive with supplier documents, quality certificates, and delivery notes. Procurement generates purchase orders and reconciliation acts. Finance handles bank statements, SAP postings, and intercompany reconciliations. Each document type had its own format, its own validation rules, and its own destination system.
The non-negotiable that shaped the entire design: the accounting and ERP systems could not change. SAP configuration at a group of this scale is fixed infrastructure — any change triggers a change management process that takes months. The automation had to work around the existing systems, not through them.
The work was being done manually. A specialist would receive a document, read it, extract the relevant fields, open the appropriate system — SAP, accounting, quality management — and type the data in. Then move to the next document. Across six plants, with different teams doing the same work independently, the aggregate person-hours were substantial.
How we automated it
Phase 1 — Process audit and document taxonomy
Every document flow across the six plants was catalogued in Primo Idea Hub: document type, volume, source system, destination system, validation rules, exception patterns. The output was a prioritised automation order — which document types to tackle first based on volume, error rate, and system complexity.
The audit surfaced an important pattern: most of the document types shared a common structure despite coming from different suppliers in different formats. Quality certificates, for example, had the same required fields regardless of supplier — but the field layout varied widely. This made them good candidates for AI-based extraction rather than rule-based parsing.
Phase 2 — AI Server deployment and extraction training
Primo AI Server was deployed on-premise inside the group's infrastructure. The first capability was document classification — identifying each incoming document as a quality certificate, delivery note, reconciliation act, or other type — followed by field extraction.
Extraction models were trained on historical documents from each plant. The team configured the extraction rules without involving an external data science function; the internal competency centre handled the training and tuning cycle directly. This was a deliberate choice — the group wanted the capability to be owned and operated internally, not dependent on external support.
Phase 3 — RPA integration and plant-by-plant rollout
Primo Robot, orchestrated by Primo Orchestrator, takes the AI-extracted output and executes the posting steps in the downstream systems — SAP, accounting, quality management — through the same interfaces a human operator would use. The accounting systems themselves were untouched.
The rollout proceeded plant by plant, starting with the highest-volume document flows. At each plant, the automation ran in parallel with manual processing for a validation period before full cutover. The rollout structure also meant each plant could be handled as a contained unit — a failure at one plant could not affect others.
The question from management was always: what happens when a supplier changes their document format? The answer is now: the team retrains the model and it is done. We do not call anyone.
The numbers
What changed for the team
The specialists who had been doing manual entry did not leave. The work shifted: from document reading and data typing to exception triage, supplier quality analysis, and process improvement. The competency centre expanded its scope — with the routine document entry automated, the team had capacity to take on more complex process categories.
The accounting systems received data in a more consistent and validated state than they had before. Errors that had been introduced through manual transcription — wrong fields, transposed numbers, missed validations — largely disappeared from the exception queue.
The plant-by-plant rollout structure also produced a reusable playbook. Each new document type or new plant could be onboarded using the same framework: audit, train, validate, cut over. The group now treats document automation as a standard capability rather than a project.
Document processing architecture — six-plant deployment
What's next for this customer
The competency centre is extending AI Server to additional document categories — customs documentation and customer billing reconciliation — using the same on-premise architecture. The team is also building a feedback loop into the extraction pipeline: when a human reviewer corrects an AI extraction, the correction is logged and fed into the next training cycle.
A second initiative is cross-plant analytics. With document data now flowing through a single orchestration layer, the group has visibility into document patterns across all six plants for the first time — supplier quality trends, reconciliation discrepancy rates, and processing volumes by plant and category.