TOOL FOR EXTRACTING INFORMATION FROM SCANNED INVOICES

Tired of manual invoice processing? Our AI-powered invoice processing solution streamlines your operations, saving you time and money while ensuring accuracy.

Key Features:

• Intelligent Invoice Recognition: Our solution utilizes advanced OCR and Google Vertex AI to accurately identify and process invoices, even those with multiple pages and complex formats.
• Data Extraction and Validation: Automatically extracts relevant data fields from invoices, such as customer ID, total amount, and line items, with minimal error rates.
• Customizable Templates: Adapt the solution to your specific invoice formats and data requirements, ensuring seamless integration with your existing systems.
• Seamless Integration: Easily integrate with your current accounting software or ERP systems for efficient data transfer and automation.

Complexity Aspects:

1. OCR
2. Invoice or non-invoice
3. Lines in invoice
4. Fields in invoice lines
5. Common fields for the invoice
6. Multi-page invoices
7. Common fields for invoice pages
8. Invoice formats
   
   

• We analyze the task and the provided data. 
• We could not find any ready-made solutions on the market that solve this task with sufficient quality.
• Tests showed that existing tools only address a single aspect of the complexity. 
• A tool that meets all aspects of complexity requires research and engineering work.
• Google models demonstrated the best performance and cost efficiency. We expect the final solution to operate using the Google Vertex AI platform.
  
  

We conduct a series of tests on a provided set of scanned documents, including invoices and other documentation. The latest OpenAI model and a specialized Google model will be used.

Our tests indicate that documents can be analyzed and fielded with high accuracy. Ready-made tools cannot fully complete the task. Most parts of the task are performed by existing market tools. Due to the multiplicity of complexities listed in this document, several problems will require research and engineering solutions.

- Models can identify invoice/non-invoice with high accuracy (>90%) without additional instructions. Simple measures are expected to increase accuracy to 95-99%.
- Models handle single-page invoices in one format well.
- Models handle a series of invoices in multiple formats without field structuring effectively.
- Text recognition is error-free and less affected by typos in source documents.
- The quality of executing complex instructions (field distribution) decreases with an increase in the number of pages processed at once.
- Multi-page invoice batches can be hundreds of pages long, necessitating either page-by-page processing followed by text and metadata processing or developing a solution for multi-page invoice identification and processing.
- Models accurately determine the number of items per package if mentioned in the description and extract this into a separate field if instructed.
- Given the potential hundreds of invoice formats, a solution independent of pre-configured mappings is preferred. System behavior needs to be studied. It is hypothesized that a Vector Database might be required. This is manageable if working with Vector Embeddings from text. Developing a custom transformer for vectorizing scanned invoices may be necessary.
- Fields relating to multiple lines or pages of invoices could also pose an engineering challenge to be addressed.

Disclaimer:

We take personal data and commercial information very seriously. As such data may be present in the documents received, we take measures to ensure their safety. Data leaks on our part are highly unlikely. We use third-party models with solutions and settings that imply that these data cannot be used or published.

Stages:

1. Demo

   - Select a solution that potentially outperforms the current one.

   - Demonstration on a limited number of variants.

   - Only mandatory line fields (name, quantity, unit price, total, date).

   - Only the most common invoice general fields (VATID, total, net, gross, customer ID, phone, IBAN, supplier name, invoice date).

   - Assess real bottlenecks.

   - Hypothesize the architecture.

   - The solution might work on a single agent or may need to be split across several; VDB might be added, invoice format/field clustering might appear, a custom transformer/vectorizer for invoice images might be needed.

Duration: 5-10 working days

Cost: €2,400

2. MVP

   - Develop a solution that meets metrics and works with an expanded range of formats. Resolved bottlenecks.

   - Presumably ready solution without integration into the company's business processes.

Duration: 10-15 working days

Cost: €6,000
The duration and cost are preliminary. They will not change if the task and terms remain unchanged.

3. Deployed Solution

   - Integrate into the company’s business processes.

   - Modify the solution based on the results of operation within the business processes.

   - Deploy the solution on company servers.

   - Handle business process errors.

   - Debug and launch.

Duration: 10-15 working days

Cost: €6,000
The duration and cost are preliminary. 

FINAL TIMELINE AND COST

Duration: 25-40 working days

Cost: €14,400

Alternative approach

If you want a more detailed, phased approach for individual functions, here is our proposal for the invoice detector.

In this case, we offer work in hourly packages starting from 40 hours at a rate of €60 per hour.

We provide regular reports with justifications, where you can see and appeal the hours spent.

Hours can be used for enhancements to the existing solution and development of other functions.

Unused hours do not expire for 12 months.

Development of a solution for invoice/non-invoice detection

Testing and tuning on larger datasets

Integration into your existing platform

Post-implementation enhancements considering real processes

Duration: 10-25 working days

Cost: 50-80 hours