Straight-Through Processing

STP rate is expressed as a percentage of total volume: if 9,500 out of 10,000 invoices in a month are processed end-to-end without a human touching them, the STP rate is 95%. The remaining 5% — exceptions — are routed to a human review queue. This is the metric that translates AI performance into operational impact: fewer staff hours per transaction, higher throughput without headcount growth, and a predictable cost per transaction.

A 65% STP rate on a process that previously required full manual handling still transforms the operation: two-thirds of the volume is handled automatically, and the human team focuses entirely on the cases that require judgment.

STP rate is determined by three factors: the quality of the AI's extraction and reasoning, the precision of the confidence thresholds used to route exceptions, and the quality of the underlying data. Poor-quality input documents, inconsistent supplier formats, or missing purchase order data will increase exception rates regardless of the AI's capability.

Improving STP rate over time requires both model-side improvements (better extraction accuracy, better disambiguation) and process-side improvements (standardising input formats where possible, feeding exception resolutions back into the system). This is why continuous improvement infrastructure — monitoring, evaluation, retraining triggers — is part of a production AI deployment, not an optional add-on.

Production AI deployments on accounts payable, order processing, and document classification workflows consistently achieve STP rates between 65% and 95% within the first months of operation. NOS processes over 20,000 supplier invoices per month with 65% straight-through. Sogrape routes 95% of distributor orders directly into SAP without human intervention.

These numbers represent the difference between a team of operators manually handling tens of thousands of transactions and a team of operators reviewing a defined exception queue — a fundamentally different operation, and the business case for the deployment.