The problem
At every monthly close, the firm received hundreds of documents in mixed formats: scanned invoices, payroll PDFs, statements and phone photographs of receipts. Entering all of it into the system consumed most of the junior team's time.
Earlier attempts to automate it with extraction alone had failed on the long tail: the system got most cases right and failed precisely on the documents that were hardest to spot as wrong.
What we did
The design accepts what automated extraction does well and what it does not. The system processes the document; human review concentrates on cases where model confidence is low, plus a control sample across the rest.
The Corpshore AI team runs that review layer from Poland, inside the European Economic Area, a non-negotiable requirement for Spanish clients' tax and employment documentation.
Review criteria were documented with the firm, and edge cases that emerge are added to the guideline rather than resolved differently each time.
Team shape
Five reviewers and a quality lead at the Poland hub, with a firm partner as counterpart on criteria.
Timeline
A 30-day pilot on invoices for a small client portfolio, extended to payroll and statements in month three.
Data handling and compliance
Delivery inside the European Economic Area with no international transfers, a condition set by the tax and employment nature of the documentation. Data processing agreement signed, with the firm as processor towards its own clients and Corpshore as a declared sub-processor.
Results
- -61 %
- Processing time per monthly close
- 99.2 %
- Accuracy after human review
- +64 h/month
- Junior team hours recovered
- 23 %
- Documents requiring human review
Measured across the whole client portfolio.
Against 91 % from extraction alone.
Reassigned to client advisory work.
The rest pass with control sampling.
Figures relate to the period stated in each case and depend on each client's starting point.
Full automation failed us because the problem was never the easy ninety percent, it was finding the ten percent that was wrong.