In a recent benchmark, it was found that handwritten forms continue to pose a significant challenge for artificial intelligence, even in 2025, with no model exceeding 95% business-level accuracy.
The benchmark, which evaluated seven leading models, revealed that OpenAI‘s GPT-5 Mini leads in terms of accuracy for high-stakes workflows. Meanwhile, Nvidia‘s Gemini 2.5 Flash Lite offers near-top accuracy at a lower cost and increased speed. Other models, including those from Azure and AWS, as well as Claude Sonnet, demonstrated moderate performance. In contrast, models from Google and Grok 4 fell short in terms of accuracy. The choice of model ultimately depends on factors such as risk tolerance, volume, and cost, with benchmarking on individual data being essential.
The results of the benchmark highlight the need for hybrid AI pipelines that incorporate validation and human review, given that even the best models rarely exceed 95% business-level accuracy. As a result, it is likely that businesses and organizations will need to continue investing in the development of more accurate AI models, as well as implementing robust validation and review processes to ensure the accuracy of handwritten document processing.
Leave a Reply