The challenge
The company runs millions of tax documents through OCR, classification, and a large manual quality-control review every season. Sales grew 140% year over year, and because the review layer scales linearly with volume, the one-hour turnaround target was running at a 7.5-hour average with overnight backlogs.
The pain points
A quality-control review layer of 400 to 450 seasonal staff that grows with every point of sales growth
Business impact: Hiring, training, and physical-space costs climb in lockstep with the business, capping how fast the company can grow.
Character-level OCR confidence that means nothing at the field level, so every job gets fully reviewed
Business impact: The system cannot tell which documents are safe to pass through, so nothing is automated and the backlog compounds.
A one-hour turnaround promise running at a 7.5-hour average during peak season
Business impact: Accounting-firm customers wait far longer than promised at the busiest point of the tax year.
How we solved it
- Built an eight-week proof of concept on Azure Document Intelligence against real historical documents
- Introduced per-field confidence scoring with green, yellow, and red routing so low-risk documents can be trusted
- Delivered a side-by-side review and voting tool plus a Power BI control tower for accuracy and quality
- Custom-trained models on the highest-volume form types, starting with brokerage statements
The impact
A working confidence-scoring pipeline that shows which documents need a human and which do not
A Power BI control tower giving leadership one view of accuracy, confidence, and throughput
A validated, managed path to scale the approach into production before the next tax season
Service line
Custom AI & Workflow Automation