The solution enables faster, guideline-compliant protocol creation, boosting researcher productivity and accelerating time-to-market for new therapies.
Meet our client
Client:
Industry:
Market:
Technology:
In a Nutshell
Client’s Challenge
The company’s post-sales service team processes approximately 30,000 customer emails per month. Each email can contain multiple attachments — often scanned or handwritten PDFs — requiring manual review and data entry into internal systems. The result: unsatisfactory service time, inconsistent classification, and a growing backlog that impacted both customer experience and operational efficiency. In addition to addressing operational inefficiencies, the client emphasized the need for security, compliance, and scalability from the outset. Handling sensitive customer data under strict GDPR regulations required a solution that guaranteed privacy, auditability, and explainability. The project also had to confirm technical feasibility and scalability to support long-term AI-driven automation across multiple business lines.
Our Solution
We developed a custom web-based AI platform to automate the classification and data extraction of inbound emails and attachments. The system analyzes each message, identifies one of 12 key service categories, and extracts key fields (policy number, name, date, etc.) from email text and scanned, handwritten documents. Our solution was built with security, compliance, and scalability as integral design principles rather than additional layers. To handle GDPR and security requirements and to accelerate the PoC, we hosted the open-source LLM on an internal, secure infrastructure. The 3-week PoC confirmed both the technical and business viability of the solution, ensuring key stakeholder buy-in and laying a strong foundation for full-scale development.
The solution includes:
- An LLM-driven classification engine
- OCR-based extraction with structured output (.CSV, .JSON)
- A web dashboard for batch processing and audit trails
Client’s Benefits
- ~85% accuracy in automatic email classification
- Processing time reduced from ~5 minutes to ~30 seconds per email
- Scalable throughput: up to 5,000 emails/day without performance loss
- >90% reduction in manual effort, freeing staff for higher-value tasks
- Established a data foundation for predictive analytics, trend analysis, and continued AI expansion
- Validated approach to GDPR-compliant solution hosted on secure infrastructure in Europe, with no third party data processing
A Deep Dive
1. Overview
The project aimed to verify whether AI could effectively automate InterRisk’s post-sales email handling – from message triage to structured data extraction. By combining open-source LLM with OCR pipelines, the PoC demonstrated high efficiency in automating repetitive classification tasks and preparing structured data for downstream systems.
The Proof of Concept confirmed that full-scale AI deployment was both feasible and strategically beneficial. It validated not only the technical viability but also the significant operational upside in terms of time savings, automation consistency, and scalability. As a result, the organization decided to move forward with a complete production rollout to realize these gains across departments.
The second phase will integrate the tool into production systems, expanding automation and introducing real-time analytics.
2. Client
A leading European insurance provider serves hundreds of thousands of customers in Central Europe. The company is focusing on digitization and AI-driven process optimization, with this PoC being a major step toward operational automation across departments.
3. Challenge
The client faced growing inefficiencies as manual processing of thousands of complex customer emails strained operations, hindered scalability, and demanded a secure, explainable AI solution. An additional challenge was posed by strict GDPR and Polish Financial Supervision Authority (UKNF) requirements that regulate how personal and financial data can be used and processed.
Business Challenge
- ~30k customer emails per month managed manually
- Delays in assigning cases to the right departments
- Costly and repetitive work for back-office teams
- No consistent framework for classification or tracking
- Strict GDPR and UKNF regulations
Technology Challenge
- Attachments with low-quality scanned documents or handwriting
- Fragmented infrastructure (shared mailbox, legacy systems)
- Limited automation and no standardized data pipeline
- Need for explainable and auditable AI integration
- Secure data handling and storage
4. Solution
The solution combines an advanced multimodal model into a web-based platform that automatically classifies customer emails, extracts key data from attachments, and delivers standardized outputs ready for seamless system integration. Our solution was built with security, compliance, and scalability as integral design principles rather than additional layers. To handle GDPR and security requirements and to accelerate the PoC, we hosted the open-source LLM on an internal, secure infrastructure.
Approach
The developed Proof of Concept (PoC) automated the intake and categorization of customer emails, including structured data extraction from attachments. A web-based application allows employees to upload, process, and download results in standardized formats. The PoC was designed to ensure easy future implementation and continuous improvement, thanks to its ragbits-based stack, which modular design supports straightforward substitution of components, integration with production systems and scalability for upcoming enhancements.
Key Components & Technologies
| Layer | Description |
| Model | Google-Gemma-3-12b-it LLM for classification |
| Hardware | 2× RTX 4090 24 GB GPUs |
| Language / Stack | Python 3.12, Gradio, LiteLLM, uv, ragbits |
| Data Processing | OCR for text & image PDFs; JSON/CSV export |
| Interface | Web dashboard for batch processing and file registry |
Functionality
- Upload and analyze single or multiple emails (up to 100 at once)
- Extract structured data and download results
- Display classification categories and keywords
- Filter and browse processed items
- Export results for downstream systems
5. Process
The project followed a structured, end-to-end approach that combined data preparation, model development, and user-centered design to ensure technical accuracy and practical adoption across teams.
Steps Taken
- Defined key 12 classification categories for post-sales operations
- Collected and labeled ~750 email samples for training & validation
- Developed OCR + NLP extraction models for key fields
- Designed a user-friendly web dashboard (Gradio interface)
- Evaluated model performance and benchmarked against the manual process
Expertise Involved
- Data Scientists (LLM fine-tuning, OCR pipelines)
- ML Engineers (architecture, automation)
- UX/Product (workflow design, dashboard)
- DevOps (infrastructure setup, deployment guidance)
6. Outcome
The PoC delivered measurable improvements in speed, accuracy, and scalability, proving that AI-driven automation can significantly enhance both operational efficiency and service quality. The project established a clear roadmap for expanding from a proof of concept to a production-grade platform.
Core components—classification logic, extraction pipelines, and dashboard architecture—were designed with modularity and extensibility in mind, ensuring the solution could evolve alongside the client’s digital transformation initiatives.
The 3-week PoC confirmed both the technical and business viability of the solution, ensuring key stakeholder buy-in and laying a strong foundation for full-scale development.
Following successful validation, the next phase focuses on:
- Integrating with live mailbox systems and back-office APIs
- Deploying within secure, scalable cloud infrastructure
- Adding monitoring dashboards and human-in-the-loop validation
- Expanding automation across departments and communication channels
These steps will transform the prototype into a full-scale AI automation solution embedded within the organization’s service ecosystem.
Quantitative Results
- 85% classification accuracy (based on validation set of 750 emails)
- Extraction accuracy up to 71% for names and 62% for document dates
- 30 seconds average processing time per email
- Up to 5,000 emails/day scalability confirmed in PoC environment
Qualitative Results
- Standardized and explainable classification across departments
- Faster registration of incoming cases (minutes instead of hours)
- Improved transparency and consistency in email processing
Lessons Learned
High-volume email handling can be effectively automated using a hybrid LLM + OCR approach. However, moving from PoC to production requires robust infrastructure, stronger models (closed-source), and full system integration.
7. The Road Ahead
In the next phase, the solution will move from proof of concept to full production deployment over an eight-week period. The plan includes creating secure connectors for live mailbox integration, linking outputs directly to core business systems, and deploying the platform within a scalable cloud environment. Dedicated dashboards will provide real-time monitoring of model performance and business KPIs, while a human-in-the-loop interface will ensure continuous quality control.
Following deployment, a six-week enhancement phase will expand the system’s scope by introducing new categories and departments, enabling sentiment and trend analysis, and leveraging insights from document processing to define standardized, automation-ready templates.
8. Expected Impact After Deployment
The production rollout is expected to deliver significant efficiency gains. By replacing manual email processing with AI automation, the organization can handle the same communication volume in a fraction of the time-improving consistency, accuracy, and scalability while reducing operational costs by up to 70%.
| Metric | Pre-AI | Post-AI (Projected) |
| Avg. processing time per email | 5–10 min | <30 sec |
| Manual review load | 100% | < 20% (exception handling only) |
| Classification accuracy | N/A (manual) | ~85% + continuous learning |
Strategic Value:
- Reliable, scalable automation framework across communication channels
- Immediate productivity gain for back-office teams
- Data foundation for advanced analytics and customer insight
Summary
The AI-powered post-sales automation platform will redefine how the client handles customer communication—turning a manual, time-intensive process into a fast, intelligent, and fully traceable workflow.
The Proof of Concept demonstrated that AI can process thousands of customer emails daily with high accuracy, reducing handling time from minutes to seconds.
With the foundation now in place, the organization is positioned to:
- Accelerate response times and deliver a superior customer experience
- Unlock real-time operational insights through structured, data-driven decision-making
- Scale automation enterprise-wide, setting a new standard for efficiency and innovation in customer operations.
