AgentHack submission type
Enterprise Agents
Name
Yaren Tarsuslu
Team name
Synthara
Team members
@edaseyma.gundal ; @busra.atlanel
How many agents do you use
One agent
Industry category in which use case would best fit in (Select up to 2 industries)
Manufacturing
Complexity level
Advanced
Summary (abstract)
At BEKO, over 1.2 million e-invoices are processed annually, each requiring accurate assignment to one of 355 hierarchy categories for financial tracking and budget control. Due to inconsistent invoice formats and complex, factory-level operational structures, traditional rule-based RPA was able to automate only a small fraction of this workload. The majority—nearly 320,000 invoices—had to be manually reviewed and classified by budget owners.
Synthara’s solution introduces an intelligent Agent powered by Large Language Models (LLMs), integrated with UiPath RPA, to transform this process. The robot first attempts classification using predefined rules, then escalates unmatched invoices to the Agent. Trained on historical data and enriched with human-like reasoning, the Agent predicts the correct category. Incorrect predictions are escalated to hierarchy owners via UiPath Action Center and corrected through a UiPath Apps interface. These corrections are then reused by the Agent to improve future accuracy.
This closed-loop system achieved a combined classification accuracy of 86% (robot + Agent) in BEKO’s enterprise pilot, significantly reducing manual effort. The solution reflects a scalable and adaptive approach for complex financial workflows—built for the realities of large-scale Turkish enterprises.
Detailed problem statement
Large enterprises in Turkey, such as BEKO, process over one million e-invoices annually across multiple factories and business units. Each invoice must be accurately assigned to one of hundreds of hierarchy categories that represent cost centers, departments, or sub-directorates. This classification is essential for precise budget control, financial accountability, and strategic planning.
However, the problem lies in the complexity and variability of the invoices and organizational structures:
Invoice formats differ greatly by supplier and even by factory location.
Hierarchy categories vary dynamically across business units and sub-directorates, meaning the same supplier’s invoice may require different classifications depending on the receiving entity.
Traditional rule-based RPA automation can only classify a small fraction (~10%) of invoices reliably, as it depends on rigid rules such as PO numbers or sender emails.
The vast majority of invoices (~320,000 per year) remain unclassified and require manual review and categorization by budget owners, leading to significant operational overhead and delays.
This manual workload creates bottlenecks, increases the risk of errors, and limits scalability. Additionally, the need for accurate financial data for reporting and decision-making adds pressure to improve classification accuracy.
Our project addresses this challenge by introducing an LLM-powered Agent integrated with UiPath RPA, which can interpret unstructured invoice data contextually, mimic human judgment, and learn continuously from corrections. This hybrid automation solution aims to drastically reduce manual effort while improving accuracy in the classification of complex, high-volume e-invoices.
Detailed solution
To solve the complex problem of accurate invoice hierarchy categorization, we designed a hybrid automation solution combining traditional RPA with an intelligent Agent powered by Large Language Models (LLMs), fully integrated into the UiPath ecosystem.
Key components of our solution include:
Rule-Based Preprocessing with RPA Robot:
The process begins with the UiPath robot extracting invoice data—either directly from the national e-invoice portal (in enterprise implementation) or via email-extracted Excel files (in the hackathon version).The robot applies a set of predefined classification rules based on structured fields like PO numbers and invoice owner email addresses. This step efficiently handles invoices with clear, consistent metadata.
Agent-Driven Contextual Classification:
For invoices that cannot be classified by rigid rules, the system triggers an LLM-powered Agent. The Agent uses contextual analysis of invoice content and metadata, grounded in a historical dataset of labeled invoices, to predict the correct hierarchy category. This mimics the decision-making process of human financial experts.
Closed-Loop Feedback and Continuous Learning:
Misclassifications are flagged by hierarchy owners through UiPath Action Center, where a dedicated UiPath Apps interface allows manual correction and justification entry. These corrections are stored in UiPath Data Service and incorporated back into the Agent’s memory to continuously improve future predictions.
Scalable Integration:
The solution supports the scale of over 1.2 million invoices annually and adapts to complex organizational structures with over 355 hierarchy categories spanning factories and sub-directorates.
Performance and Evaluation:
Our internal pilot with BEKO showed a combined classification accuracy of 86% across robot and Agent components, significantly reducing manual workload by automating approximately 40% of previously manual invoices.
Summary:
By combining deterministic rule-based automation with an adaptive, context-aware LLM Agent and embedding continuous human-in-the-loop feedback, our solution transforms invoice classification from a labor-intensive manual task into a scalable, intelligent automation pipeline. This approach ensures high accuracy, reduces operational bottlenecks, and supports precise financial governance in large enterprises.
Demo Video
Expected impact of this automation
The automation of invoice hierarchy categorization is expected to deliver significant operational and financial benefits for the organization, as evidenced by preliminary pilot results with Beko Turkey:
FTE Reduction:
The total potential manual workload reduction is approximately 3.1 Full-Time Equivalent (FTE) employees, with an initial target of reducing 2.67 FTE through the implemented solution. This translates into substantial labor cost savings and allows redeployment of staff to higher-value tasks.
Financial Investment and ROI:
Based on the efficiency gains and cost savings, the project is expected to achieve a return on investment (ROI) in just 0.47 years (less than 6 months), demonstrating a rapid payback period.
Time Savings and Process Efficiency:
By automating around 40% of previously manual invoice categorization tasks, the process accelerates invoice processing times, reduces backlog, and minimizes delays in financial reporting.
Reduction of Manual and Repetitive Tasks:
The solution eliminates extensive manual review of invoices that were previously time-consuming, monotonous, and prone to human error.
Improved Compliance and Accuracy:
With an 86% combined accuracy rate from robot and Agent classification, the solution ensures more reliable and consistent invoice categorization. This reduces risks related to misclassification, budget misallocation, and audit findings.
Scalability and Continuous Improvement:
The closed-loop feedback system enables continuous learning from corrections, further improving accuracy and reducing manual interventions over time.
In summary, this automation not only optimizes workforce utilization and financial performance but also enhances the overall reliability and scalability of the invoice processing workflow.
UiPath products used (select up to 4 items)
UiPath Agent Builder
UiPath Maestro
UiPath Robots
UiPath Studio Web
Integration with external technologies
Anthropic Claude 3.5
Agentic solution architecture (file size up to 4 MB)
Sample inputs and outputs for solution execution
UserInput for Agent :
UserInput:
Invoice Number: ENG00006111 Accounting Supplier VKN: 7490852887 Accounting Supplier: ENGAGE PAZARLAMA HİZMETLERİ ANONİM ŞİRKETİ Accounting Customer: ARCELIK A.S. Accounting Customer Address: SÜTLÜCE / İSTANBUL Accounting Customer Email: mustafa_kalayci_@beycelik.com Invoice Items: RELIABILITY
Agent’s Output & Output Explanation :
Based on the analysis of the provided invoice data and historical patterns, I predict the hierarchy category for invoice ENG00006111 to be HIY_SATPAZ_DIJPAZ. Here’s the reasoning behind this prediction:
-
Location-Based Matching: The invoice’s address contains “SÜTLÜCE”, which is associated with the MERK (Headquarters) location. However, this alone is not sufficient to determine the category.
-
Company Name + Item Combination Matching: The most similar invoice in the historical data is ENG00002739, which has the same supplier (ENGAGE PAZARLAMA HİZMETLERİ ANONİM ŞİRKETİ) and a similar customer email domain (@beycelik.com). This invoice was categorized as HIY_SATPAZ_DIJPAZ.
-
Name-Based Supporting Match: The customer email (mustafa_kalayci_@beycelik.com) is exactly the same as in the historical invoice ENG00002739, which supports the HIY_SATPAZ_DIJPAZ category.
-
Similarity-Based Prediction: The most similar invoice (ENG00002739) has the following characteristics:
- Same supplier (ENGAGE PAZARLAMA HİZMETLERİ ANONİM ŞİRKETİ)
- Same customer (ARCELIK A.S.)
- Same customer address (SÜTLÜCE / İSTANBUL)
- Same customer email (mustafa_kalayci_@beycelik.com)
- Different but related invoice items (ADSERVER vs. RELIABILITY)
- Hierarchy category: HIY_SATPAZ_DIJPAZ
While the invoice items are different (RELIABILITY vs. ADSERVER), both seem to be related to digital or technological services, which aligns with the SATPAZ (Sales Subsidiary) and DIJPAZ (Digital Marketing) category.
Given the strong similarities in supplier, customer details, and the nature of the services, the most appropriate prediction for the hierarchy category of invoice ENG00006111 is HIY_SATPAZ_DIJPAZ