$10M+ a year saved and 15x less manual work with AI invoice processing at 10+ companies

Client story

Our clients in logistics, legal, retail and construction were looking for a way to automate invoice processing, reduce manual labour and raise operational efficiency. We built a customized tool that automates document data extraction, processing and analysis, and delivers substantial cost savings.

Adoption and scalability across different business operations were essential. We have since implemented the solution at more than 10 companies. Together they generate hundreds of thousands of invoices a month, and they report higher revenues since adopting the tool.

“Leveraging AI to automate our invoice processing was a game-changer for our operations team and exceeded customer expectations. We worked with Erbis to develop and implement a comprehensive solution that eliminated data entry, reduced processing times, and most importantly, allowed us to scale business operations.”

President, Supply Chain SaaS company

Challenge

Manual invoicing is a burden for companies in every industry. Automating it lets accountants improve efficiency, reduce errors and free time for more strategic work.

The main challenges were:

• High volume of paper invoices. Companies receive thousands of paper and scanned invoices a month from many suppliers.
• Manual data entry. Employees and external accountants digitize these invoices by hand, which is labour-intensive and error-prone.
• Lack of standardization. Invoices vary in format and include handwritten annotations, which makes digitization chaotic and complex.
• High error rate. Human intervention is frequently needed to correct errors, which slows the process further and reduces accuracy.

There were also technical hurdles, such as integrating the new solution with existing software. Overcoming them required close collaboration between the finance departments and our team so that adoption went smoothly.

Solution

We developed an AI solution that uses Amazon Textract for intelligent document extraction, with Amazon SageMaker Ground Truth as the human-in-the-loop (HITL) layer. The combination automates data extraction and delivers a scalable system that handles complex invoices with minimal errors.

The pipeline works in six steps:

• PDF acquisition. Scanned PDF invoices arrive in an email inbox. A serverless AWS Lambda function periodically checks the inbox for new emails containing invoices.
• Preprocessing. PDFs are converted to JPEG images, and quality-enhancement techniques are applied to improve readability.
• Template-based processing. We created a template for each supplier’s invoices, because the AI needed guidance to interpret the data correctly. Templates describe the expected tables and column layouts.
• Amazon Textract extraction. The preprocessed images are fed into Textract, which extracts text data according to the predefined templates. The extracted data is evaluated against a quality threshold.
• SageMaker Ground Truth human-in-the-loop review. If parsing accuracy meets the standard, the data is saved as CSV and sent via FTP. Otherwise SageMaker Ground Truth applies a two-tier review strategy: skip review when confidence is high, send for human review otherwise. Custom HTML/JavaScript review interfaces guide reviewers with pre-populated fields, dynamic dropdowns and real-time validation. Human corrections feed an active-learning loop that refines the extraction rules over time. Lambda orchestrates the flow, with private-workforce notifications and real-time status tracking through the SageMaker APIs.
• Table parsing. Tables were particularly hard to parse because of overlapping columns and complex layouts. Templates combined with the HITL review guided the AI to interpret tables correctly and avoid common errors, which lifted overall accuracy.

Team: Senior ESB Engineer, two ESB Engineers, Senior Java Engineer, Data Scientist, Software Architect.

Architecture and tech stack

The solution runs entirely on AWS serverless services. Lambda handles ingestion and orchestration, Textract performs the extraction, SageMaker Ground Truth manages the review workforce and the active-learning loop, S3 stores documents and outputs, and API Gateway exposes the integration endpoints.

• Amazon Textract
• Amazon SageMaker Ground Truth
• AWS Lambda
• Amazon S3
• Amazon API Gateway

Results

We completed the project in under 6 months and integrated the solution into 10+ of our ongoing projects across several sectors.

• Cost savings. Annual savings across the companies using the solution exceed $10 million.
• Automation efficiency. The manual labour required for invoice processing fell by 15 times, freeing significant time and resources.
• Digitization rate. Almost 95% of invoices are now digitized without human intervention, which means faster turnaround and fewer errors.

The $10M+ figure is the aggregate across the 10+ companies using the solution. One of them, a supply chain SaaS company, puts its own operational saving at roughly 1,700 manhours a year in processing time (see the quote at the end of this page).

What made it work

Erbis’s takeaways from this engagement:

• Put a confidence gate in front of the human. Routing only low-confidence extractions to reviewers is what let the digitization rate climb while the review workload shrank.
• Treat corrections as training data. Every human fix flowed back into the extraction rules, so the system kept improving after launch instead of degrading.
• Per-supplier templates beat a generic model for tables. Overlapping columns and handwritten notes were solved with structure, not with more model capacity.
• Build the integration path first. Email in, CSV over FTP out: adoption across 10+ companies depended on fitting the existing finance workflow rather than replacing it.

“It's difficult to truly quantify the value of the project since customer satisfaction and scalability were headline benefits. From a purely operational perspective, we believe this project saves roughly 1,700 manhours a year in processing time. I'm happy to have Erbis as our long-term technology partner, I believe they care as much about our business as we do.”