PO/Invoice IDP → ERP Automation (ezIDP + RPA)
Problem
Procurement and finance teams manually retyped data from purchase orders, invoices, and delivery orders into Sage and Bassnet, then manually chased down whoever needed to know a payment had gone out — slow, error-prone, and entirely dependent on whoever was available that day.
What I built
Two related but separate automations, both on Cyclone’s RPA/IDP platform:
PO/invoice/DO intake — configured and prompt-tuned Cyclone ezIDP (Cyclone’s OCR + LLM document-extraction product) to pull structured purchase-order data out of scanned and PDF documents, then built the Cyclone RPA pipeline that takes the extracted data and submits it directly into Sage, removing the manual re-keying step entirely.
Payment voucher automation — a separate, partially-automated batch process: a user drops files into a designated folder, and an RPA robot picks them up, pre-processes and validates the data, performs the data entry into Sage, generates the payment voucher, syncs the transaction status to Bassnet, and emails the output file to the relevant PIC automatically.
Key engineering decisions
- Configure the vendor tool, don’t reinvent it. ezIDP already solves generic OCR+LLM extraction; the real engineering work was tuning its prompts and templates for this client’s specific document formats and wiring its output into the actual downstream systems — that integration and RPA orchestration is where the value was, not rebuilding an extraction engine that already existed.
- Folder-drop as the trigger, not a fully unattended pipeline. The payment voucher flow is deliberately “partial auto” — a human still decides what goes into the folder — rather than fully autonomous, so a bad batch can’t silently generate and submit incorrect payment vouchers unattended.
- Multi-system sync as a first-class step, not an afterthought. Keeping Sage and Bassnet’s transaction status in sync, and notifying the PIC automatically, closes the loop manual processes usually dropped — the automation isn’t done when the data entry happens, it’s done when the right person has been told.
Result
Reached 95% extraction accuracy on production documents and eliminated manual purchase-order re-entry and payment-voucher generation for the client’s procurement and finance teams.