Work and processes
What Is STP: The Direct Processing Rate That Shows Real Savings, Explained
Korea Deep LearningPublished
STP (Straight-Through Processing, direct processing rate) is the share of all received cases that were completed from start to finish without a person ever touching them. It is the most direct measure of whether automation actually reduced the workload.
Why the Rate Matters More Than Accuracy
A higher accuracy figure does not mean the workload has gone down.
The number usually discussed when weighing adoption is recognition accuracy. But there is a large gap between a 98% accuracy figure and the time staff spend on their work.
If you cannot tell where the remaining 2% is, staff have to check 100%. In that state, even if accuracy rises to 99%, the checking work stays the same. Conversely, if the cases likely to be wrong can be picked out accurately, staff only need to look at that 5%, even at 95% accuracy. In one actual case, replacing a full recheck with a review of exception cases greatly reduced the checking burden. The number that explains labor savings is the direct processing rate, not accuracy.
What STP Is Exactly: How It Differs from Recognition Accuracy
Three Axes That Separate Recognition Accuracy and STP
First, the unit being counted differs. Recognition accuracy counts characters or items. STP counts work cases. If a case has twenty items and one of them is wrong, item accuracy is 95%, but that case has failed direct processing.
Second, what is reflected differs. Recognition accuracy reflects only model performance. STP also reflects validation rules, baseline settings, and exception handling design. Even with the same model, STP can vary greatly depending on operational design.
Third, where they are used differs. Recognition accuracy is used to compare products. STP is used to explain the effect of an investment. When you need to explain savings to executives, the second kind of number is the one required.
The two metrics are not substitutes for each other. If recognition accuracy is low, STP cannot rise either. However, high recognition accuracy does not automatically bring STP up with it.
Four Criteria for Measuring STP Accurately
Four Criteria Used in Measurement
First, fix the denominator. Whether it covers all received cases or only the forms chosen for automation changes the value a lot. Leaving out documents outside the scope makes the result higher than reality.
Second, decide how far human intervention extends. You need to decide whether only editing a value counts as intervention, or whether opening a screen just to check it also counts. Checking alone still takes time.
Third, decide how to count reprocessed cases. Whether a case that dropped into an exception, was corrected, and rejoined the flow counts as a success or a failure changes the value.
Fourth, fix the measurement period. Measure monthly, and if certain documents concentrate at particular times, record their effect as well.
How to Raise STP in Practice
Group Failures by Cause
When STP is low, the causes usually fall into a few groups. Failures repeat for a specific form, confidence comes out low for certain items, or validation rules are too strict.
Grouping failed cases by type shows where to make changes. For a form problem, adjust the item definitions. For an item problem, adjust the extraction criteria. For a rule problem, adjust the baseline.
Adjust Validation Rules and the Baseline Together
Tight validation reduces errors, but it increases exceptions, which lowers STP. Loosening it raises STP, but incorrect values then pass through.
The two values must be adjusted together while looking at both. The safe approach is to relax rules that the review history shows caused no real problems, and to add rules at the points where errors occurred.
STP in the Korean Environment
Korean finance and public-sector work carries statutory deadlines and audit requirements, so there are sections where STP cannot easily be pushed up. For work where an error has a large impact, such as payment review or accident intake, a conservative baseline is reasonable.
At the same time, manage the processing speed of exception cases as well. Even with the same STP, total lead time varies depending on how quickly exception cases are handled. In one actual case, payment lead time was greatly reduced.
Read It Separately from the Task Completion Rate
There are cases where the no-intervention rate is high but staff workload does not fall. This happens because a case was decided without human hands, but the result was never registered in the work system.
So look at the two metrics separately. The no-intervention rate shows whether a person was involved up to the decision stage. The task completion rate shows whether a case was finished through registration in the system. If the gap between the two values is large, the link is broken, so the integration section rather than recognition needs fixing.
The direction of improvement follows from this split too. If the no-intervention rate is low, many cases are going to exceptions, so look at recognition and validation criteria. If the no-intervention rate is high but the task completion rate is low, look at system integration and approval procedures.
Frequently asked questions
What target level should be set?
It depends on the work. For work where one error has a large impact, it is safer to start low and raise the target based on review history.
Recognition accuracy is high, but STP is low.
Validation rules are likely too strict, or the baseline is set too high. Divide the failed cases by type and review them.
Can it be predicted before adoption?
A rough estimate is possible. While validating with your own documents, looking at the distribution of confidence by item shows the range of the exception rate.
How is the savings converted?
Multiply the directly processed cases by the previous processing time per case to get the saved time. It is common to convert that value into full-time staff equivalents for the explanation.
How do the no-intervention rate and recognition accuracy differ?
Recognition accuracy shows how often the model was right. The no-intervention rate shows the share of cases that finished without anyone touching them.
Why is the task completion rate tracked separately?
Even if judgment is completed automatically, if the flow breaks at system registration, staff must enter the data again. The gap between the two metrics reveals problems in the integration section.
Tell us your current review method and monthly volume, and we will review the expected direct processing rate with you.
Contact us



