Artificial intelligence, from its name, suggests two characteristics: automation and intelligence. From these two perspectives, the level of intelligence is insufficient, leading to an inability to truly achieve automation in practice. For example, a private enterprise has bank reconciliation statements, each bank’s statement has different formats and a large quantity, resulting in a significant workload! Therefore, many small artificial intelligence companies focus on using AI methods (OCR) to automatically input bank reconciliation statements, reducing human workload.
The technical implementation of bank reconciliation statements carries a key premise: bank reconciliation involves financial reconciliation, and it must be ‘error-free’ and 100% correct! Further breaking this down: if all content is 100% correct, that’s best. If it cannot be 100% correct, at least it should strictly confirm which content is 100% correct and which content is 100% incorrect. In this way, only the content that is 100% incorrect needs to be manually verified, thereby reducing the workload. Otherwise, if each recognition result is merely a probability value (even if it’s 0.9 or even 0.99), and cannot ensure the reliability of the result, then manual verification is still required, and true automation cannot be achieved. Some readers might be curious why the model’s threshold has been set to 0.99, which is already very high, yet the results are still untrustworthy. This is easy to understand: Model A, due to good training, might be quite accurate at 0.9. Model B, lacking good training, might be inaccurate even at 0.99.
Moreover, in real-world scenarios, factors such as damage to the documents and shooting angles can indeed easily lead to inaccurate recognition results. Therefore, even in the era of large models, for such an old and classic scenario of OCR recognition, manufacturers still only repeatedly claim that the accuracy rate has improved by more than 10 percentage points (relative value), and no company can claim to have achieved 100% accuracy (absolute value) in engineering practice, thus realizing true automation.
Automation is an important standard for measuring the practicality of artificial intelligence! This is also the core issue that artificial intelligence truly needs to focus on and solve!