Gemini 2.0 and Gemini 2.5 Flash Models for Financial Document Intelligence

At our core, we believe the best solutions are born from deep curiosity, research, and a relentless commitment to our clients’ goals. The research-driven approach encourages our employees to explore the latest technologies, ensuring we deliver the most innovative and tailored solutions for our clients.

This mindset guided us through our partnership with our client, one of Australia’s most innovative technology-powered lenders. When the client approached us with the complex challenge of streamlining their financial document processing, our team immediately set out to investigate the state of the art in AI document intelligence. We didn’t rely on generic, off-the-shelf tools. Instead, our AI engineers benchmarked, experimented, and iterated, ensuring that any solution we proposed would be robust, secure, and precisely tuned to Plenti’s needs.

In the current stage of the project, we have compared leading AI models, including Gemini 2.0 and Gemini 2.5 Flash, to identify the technology that would deliver the highest accuracy and efficiency for the client’s growth document volumes.

A Custom AI-Driven Workflow

The first step was to build a robust system that could accurately identify the type of document being processed. For this, we trained a custom deep learning model on a vast dataset of financial documents. This bespoke solution was designed to be highly accurate and secure, ensuring that sensitive information was handled appropriately.

Our research included comparing several models, with a specific focus on the performance of Gemini 2.0 and the more advanced Gemini 2.5 Flash models. We also benchmarked their performance with VGG16, a well-established convolutional neural network architecture. The result of this testing, as seen in the accompanying chart, provides a clear picture of the models’ capabilities.

Key takeaways:

• Gemini 2.5 outperformed Gemini 2.0 in most categories, showing notable gains in challenging document types such as Payslip, Roadworthiness Certificate, and Small Scale Technology Certificate.

• Gemini 2.0 still achieved near-perfect accuracy in many high-volume document types, such as Invoices and Notice of Assessment.

• VGG16 lagged behind the Gemini models in consistency, though it performed competitively in some document types, like Green Equipment Invoice.

A Performance Deep Dive: Comparing Gemini Models

We evaluated the classification accuracy of three models: Gemini 2.0 (LLM), Gemini 2.5 (LLM), and VGG16 (Baseline CNN) on both Document Classification and Document Extraction.

1. Document Classification

As demonstrated in the chart, the Gemini models showed a remarkable ability to classify various financial documents.

• Overall Performance

On most document types, Gemini 2.5 Flash achieved the highest accuracy, often reaching and exceeding 99%. This includes critical documents like License (99.14%), Individual Tax Return (99.25%), and Notice of Assessment (99.12%).

• Edge Cases and Specific Strengths

While Gemini 2.5 showed near-perfect performance on many documents, the testing also revealed specific strengths. For example, VGG16 slightly edged out the Gemini models on Vehicle Registration Document and Payroll Letter, showcasing the importance of evaluating multiple models. Conversely, on Payslip documents, Gemini 2.5 Flash’s 99.68% accuracy was significantly higher than the other models.

• Insights into Model Behavior

The testing also highlighted areas for further optimization. For instance, the Warehouse Comprehensive Insurance Policy document posed a significant challenge, with Gemini 2.5 reporting 0% accuracy. This suggests that while Gemini 2.5 is exceptionally good at generalizing across a wide range of document types, it may require further fine-tuning or a more specific training approach for highly specialized or unique document formats. Gemini 2.0, in this specific case, performed better with 76.63% accuracy.

2. Information Extraction

While classification is the essential step, the true power of our solution lies in its ability to accurately extract information from these documents. Our testing confirmed that Gemini 2.5 Flash could achieve 100% accuracy on critical extraction tasks, allowing for the automation of a wide range of verification steps.

• Bank Statements

• Transaction coverage (90 days): 100% accuracy

• Statement recency (within 30 days): 96% accuracy

• Customer name verification: 100% accuracy

• BSB and account number detection: high reliability, though some edge cases remain

This capability allows for real-time feedback to be provided to the customer if an issue is detected, such as a statement being out of date or not showing a full 90-day transaction history.

• Payslips

• Customer name checks: 100% accuracy

• ABN validation: 100% accuracy

• Year-to-date earnings recognition: 100% accuracy

• Document recency and issue date checks: reliable, with minor limitations where dates were missing or unclear

• Notice of Assessment (ATO)

• Document classification: 100% accuracy

• Customer name, financial year, and income verification: 100% accuracy across tested samples

• Individual Tax Returns

• Document classification: 100% accuracy

• Customer name and financial year checks: 100% accuracy 

• Income verification: 100% accuracy

These results prove that Gemini Flash 2.5’s ability is not just to recognize document types, but to extract granular, compliance-critical details with high precision. This ensures that the client’s employees can trust the data provided by AI, significantly reducing the need for manual checks while enabling faster loan approvals.

Conclusion and Future Outlook

The adoption of the Gemini models has been central to the success of this project. Gemini 2.0 delivered exceptional performance out of the gate, accurately classifying most financial documents and proving its value in high-volume categories. Gemini 2.5 Flash took that foundation and refined it further, consistently delivering higher accuracy across more challenging documents and offering greater reliability in real-world conditions.

Together, these models have transformed our client’s document workflows from a slow, manual process into a fast, intelligent, and highly accurate pipeline. By utilizing Gemini’s strengths, our client now processes documents at scale with unmatched precision – setting a new standard for efficiency in loan approvals and reinforcing our commitment to pairing the right technology with the right problem.