AWS Summit Zurich 2025 - Implementing AI Document Processing at Scale: Syngenta

Intelligent Document Processing Overview [01:43]

• 80% of organizational data is unstructured, posing challenges for downstream applications and processes.
• Companies struggle to efficiently handle documents, audio, and video at scale, with less than 20% having established processes.
• IDP enables the extraction of data from various unstructured formats to feed digital business processes, applicable across industries like warehouse management, health and life sciences, fintech, and human resources.
• The IDP pattern involves classifying documents, extracting data, and performing post-processing for enrichment and normalization, with continuous learning and monitoring to improve accuracy.

"And this is where less than 20% of companies have processes in place because most most of the time is their core competency of their core business to handle such document at scale."

Senta and AI Integration [05:05]

• Senta is a science-based company operating in 90 countries, with 30,000 employees, producing products for crop protection, seeds, and vegetables.
• The company aims to feed a growing global population (projected 10 billion by 2050) with 50% more food sustainably, without using 50% more land.
• AI is utilized across Senta, including in IT for digital products for farmers (disease diagnosis, chatbots) and in back-office processes.
• Senta has a history with AI in demand planning, forecasting, and has an in-house LLM used for various internal processes and embedded into SAP applications.
• The company has long used AI, particularly machine learning, in document processing, automating AP processing which previously used older OCR and basic ML for three-way invoice matching in SAP.

"And ideally 50% more food without chewing up 50% more land, right? So being able to be able to feed the world as a as a population increases and being able to do that sustainably is what is what Senta is all about."

The Wingman Project: AI for Invoice Processing [09:50]

• Wingman is the project name for Senta's AI initiative focused on intelligent document processing, previously named Co-pilot.
• Senta processes approximately 1 million invoices annually, with an additional 2.5 million related documents like credit notes.
• Current document processing capabilities are described as "dumb," sufficient for three-way matching but lacking deeper insight extraction.
• The core message is that leveraging the content of invoices beyond what's typically moved into SAP can solve numerous business problems, with LLMs enabling this at scale.
• The vision is to create a pipeline of scanned documents in an "invoice data lake" to derive greater insights and solve business challenges.

"And the one message I want to leave with you, right? if you if you don't remember anything else is if you can leverage the the content of an invoice all the things that you wouldn't normally find in move into SAP you can solve a lot of business problems with that right and and LLMs will enable us to do that and do that at scale."

Wingman Use Cases [11:41]

• Indirect Spend Reduction: LLMs can categorize spend by analyzing invoice content, enabling cost reduction opportunities and better negotiation deals by clustering products by type.
• Vendor Master Data Accuracy: LLMs can intelligently understand vendor details on invoices (including bank details) and compare them with SAP MDG data, intercepting and correcting discrepancies before processing to avoid payment interruptions.
• T Compliance (Expenses): LLMs can identify spend types on expense claims and compare them against expense policies, flagging compliance issues. This can be done in real-time within systems like Concur or in batches.
• Tax Optimization and Compliance: Wingman analyzes invoices against tax law and company tax policies. In Argentina, with complex state-specific and product-specific tax laws, Wingman verifies tax classifications, suggesting claims or additional payments where incorrect, saving significant manual effort.

"All right, so use case number one for us is really procurement spend reduction."

Wingman Architecture and Functionality [17:45]

• A Proof of Concept (PoC) for batch invoice processing and tax compliance in Argentina has been completed and works.
• Dashboards in Power BI are used to visualize matched and mismatched tax calculations between SAP data and Wingman's suggestions.
• Wingman provides insights into tax classification discrepancies, explaining why a particular tax code might be incorrect and suggesting alternatives.
• It can also identify currency mismatches, flagging when an invoice currency differs from the expected local currency (e.g., USD vs. ARS in Argentina).
• Future phases include integrating a chatbot for AP staff to check tax calculations before processing and developing Wingman to continuously monitor and propose updates to tax policies based on changes in tax law.

"And on the right hand side, the LLM is intelligent enough also to kind of describe why it's incorrect. All right, it provides a summary."

Amazon Bedrock Data Automation (BDA) [22:06]

• The implemented architecture uses AWS services with the ERP in EC2, data extraction to Redshift, and document archiving on S3.
• Wingman's dashboard displays SAP data, LLM suggestions from Wingman, and is stored on Amazon OpenSearch Serverless.
• Data flows from a data lake via Glue ETL jobs to OpenSearch Serverless, while invoices in PDF format are fed into Amazon Bedrock Data Automation (BDA).
• BDA, a combination of OCR and LLM techniques, is a fully managed service that linearizes any document modality (audio, video, document, picture) into a JSON output.
• BDA is integrated with Amazon Bedrock for model selection (e.g., Anthropic, Cohere) and context window for LLM processing, providing structured insights and suggestions.

"BDA is basically a combination of OCR and LLM technique. It's a fully managed service."

BDA Capabilities and Demo [26:48]

• BDA can handle various modalities (audio, video, document, pictures) and extract data into a linearized or structured schema.
• It provides accuracy with confidence scores and integrates with the AWS service stack.
• Development is simplified as it's fully managed, removing the need for in-house ML expertise.
• BDA offers tailor-made output, simple and intuitive interfaces, and is built on responsible AI principles with features like toxic detection.
• It integrates with Amazon Bedrock Knowledge Bases, allowing its parsing capabilities to enrich RAG systems with insights from text, pictures, graphs, audio, and video.
• The service supports both standard output (linearization) and custom output, where users define blueprints for specific document types.
• A demo shows BDA processing a bank statement, generating a summary, page-level linearization, and element-level descriptions, and then demonstrates custom output using a predefined bank statement blueprint, extracting 7 out of 10 fields.

"So with a standard output you have already a lot of information a lot of information that you can use for your for from your document and that you can feed your processes."

Future of Wingman and Lessons Learned [39:13]

• Next Steps: Ramp up the tax use case rollout to more countries, expand procurement use cases, and explore new use cases as awareness grows (e.g., certificates of analysis for chemical products).
• Foundation Building: Scale up the pipeline of scanned and LLM-processed documents to create an "invoice data lake" of raw material for analysis.
• Lesson Learned/Tip: The technology is powerful, but foundational elements are crucial. Organizations need to have the "something else" to compare against – e.g., expense policies, tax policies – readily available as the LLM can help create them, but they need to exist as a grounding truth.

"Wingman's all about taking information in a document and comparing it to something else. You need to have that something else, right?"