Affinda's Invoice Extractor solution aims to accurately and quickly return key data from an invoice with as little human intervention as possible. The use of confidence levels and auto-validation is key to minimising this human intervention.
Artificial intelligence models are probabilistic. This means that the AI model will return what it believes is most likely to be correct, which will be the data it is most confident is the correct answer. Affinda returns the confidence levels to users, either via API or through the validation interface, so that users can more easily direct their attention to those fields that the model has less confidence in and therefore be more likely to be incorrect.
The confidence levels shown take into consideration:
- That the data point selected by the model is correct
- In the case of scanned images, the confidence that the model has that the capture of the text via OCR is correct
To reduce the amount of human intervention, auto-validation rules and thresholds can be set so that users only need to validate a subset of all data fields. Within the Affinda web app, users can set their auto-validation threshold. Any data fields whose confidence level is above this user-set threshold will be auto-validated and not require any human to validate this data point.
By its nature, from time to time an auto-validated data field may be incorrect. As such, we recommend setting a high auto-validation threshold above 90%. Any auto-validated fields that are identified to be incorrect can still be corrected within the validation interface.