Image Recognition Versus Manual Processing: Achieving The Most Value
By Greg Council, Parascript, LLC
When measuring accuracy in manual data processing, humans are inaccurate from 2% to 6% of the time. This accounts for factors such as boredom, repetition and fatigue that affect an individual's overall performance.
Alternatively, with recognition, there are two important variables affecting success rates — errors and rejects. These are specific to each implementation and associated business rules.
Errors are issues involved with false data recognition. In this instance, incorrect data is passed on to another system. Rejects occur when the confidence rate of a particular recognition process does not meet a specific threshold and requires manual intervention. Both incur a cost, but both costs can be quite different.
How so? Errors involve sending incorrect data to another system. Costs here could be argued as trivial — for instance if the error is associated with a name used for marketing correspondence. The result is that the person receives correspondence but with an inaccurate name. Or it could be large: the error could be associated with a social security number, which is associated with a financial transaction, having much larger ramifications.
For rejects, the cost is typically associated with the effort required to review the data and correct it manually. If the business is outsourcing this effort to a low-cost area, then it might be trivial. The key here is to balance the cost of manual interaction with the cost of bad data going into the system.