years of manual process automated in one quarter
order intake, quality paperwork, and supplier matching — moved from inboxes and spreadsheets to agents with human sign-off.
- client
- an industrial components manufacturer, 4 plants
- sector
- manufacturing ↗
- duration
- 13 weeks
- team
- 5 engineers, 1 ml engineer, 1 process lead
every customer order arrived as a pdf, an email, or an edi message — and each was retyped into the erp by hand. quality certificates and supplier confirmations followed the same path, adding days to every order.
previous automation attempts relied on templates that broke whenever a customer changed a form.
13 weeks, 4 phases, one team.
shadowed the order desk, measured volume, variance, and error rates.
extraction and matching agents evaluated on six months of history.
erp write-back with approval queues and full audit trail.
evals, monitoring, and retraining owned by the plant it team.
- every order retyped by hand
- templates that broke weekly
- two-day order acknowledgement
- errors found at invoicing
- 71% touchless processing
- layout-agnostic extraction
- same-day acknowledgement
- validation at intake
the team stopped being typists and started being the people customers call first.
incoming orders, certificates, and confirmations are read, validated against the erp, and queued for one-click approval. anything uncertain is routed to a person with the evidence highlighted.
the order desk moved from data entry to exception handling and customer contact.
- python
- azure openai
- langgraph
- sap
- postgres
- mlflow