Scaling to Millions of Transactions with +90% Spend Classification Accuracy Using Agentic Systems

Industry Context

Enterprise procurement classification is increasingly automated using AI, particularly for complex, multi-taxonomy environments. According to a 2024 BCG report, 60% of large enterprises plan AI investments exceeding $25M, with procurement a key focus area. HITL classification models are gaining traction as they balance automation efficiency with compliance, accuracy, and enterprise-specific rules.

The Problem: Multi-Taxonomy Procurement Classification

The client’s classification challenge included:

• Inconsistent Naming – Same items appearing under varied names, spellings, and abbreviations across ERP systems.

• Free-Text Descriptions – Missing codes or inconsistent formatting reduced automation accuracy.

• Multiple Taxonomies – UNSPSC, ISIC, and internal schemas required cross-mapping.

• Decentralized Procurement – Different business units using disparate standards.

• Supplier & Product Diversity – Thousands of suppliers and millions of items from raw materials to MRO.

• High Volume & Velocity – Continuous inflow of transactions requiring timely classification.

• Enterprise-Specific Rules – Special handling for sister company transactions and designated suppliers.

• Manual Inefficiency – 12,000 hours spent annually on classification, with 20% misclassification risk.

The Solution: PSA Classification Capability

PSA’s classification capability addressed these issues through an integrated AI workflow

1. Input Data Integration

- Procurement taxonomy with 200+ categories

- Historical expense data from ERP, including supplier names, descriptions, and amounts

2. Enterprise Rule Encoding

- Sister company exclusions

- Category overrides for specific suppliers

- Non-sourceable transaction flags

3. Model Training

- Random Forest and Gradient Boosting trained on multi-year expense data to capture seasonal and category patterns

4. Real-Time Classification

- AI model classifies incoming transactions against taxonomy

- RAG retrieves matching historical classification and rules

- LLM analyses new or ambiguous entries using NLP for best-fit category inference

5. HITL Oversight

- Low-confidence classifications (score < 0.7) routed for human review

- Reviewer corrections fed back for continuous retraining

6. Continuous Improvement

- Six-month retraining cycles improved accuracy from 92% at pilot to 95% at scale

Implementation

• Data Preparation – Three years of ERP transactions cleaned, normalized, and deduplicated

• Rule Configuration – Cross-unit alignment on enterprise rules before encoding

• Pilot Deployment – Logistics and IT spend categories tested first, then scaled

• User Training – Procurement team trained on HITL interface and accuracy scoring

Results and Impact

• Accuracy – Misclassification reduced from 20% to < 5%

• Time Savings – Manual workload reduced from 12,000 hours to 1,800 annually (85% reduction)

• Cost Savings – $30M in annual opportunities identified through improved supplier and contract alignment

• Consistency – AI-predicted spend shares now differ by < 1% from original categories at Level 1 taxonomy

Testimonial

Chief Procurement Officer:

“PSA has freed our team from repetitive work while maintaining classification accuracy at a scale we could not achieve manually. The system adapts to our rules and taxonomies without disrupting existing processes.”

Procurement Analytics Lead:

“The HITL interface is intuitive. We’ve closed the gap between AI predictions and our internal standards while keeping pace with transaction volumes.”

Business Value

• Operational Efficiency – 70% faster classification turnaround

• Data-Driven Decisions – Reliable spend analytics for sourcing, supplier performance, and market monitoring

• Risk Mitigation – Enterprise rules enforced automatically, ensuring compliance

• Strategic Reallocation – Resources shifted from manual classification to supplier strategy and innovation

Conclusion

By combining machine learning, RAG, LLMs, and HITL review, PSA on the Zen Platform delivered enterprise-grade, multi-taxonomy spend classification at scale.

‍The solution has set a repeatable model for AI-driven procurement analytics in global organizations, improving accuracy, efficiency, and decision-making resilience.

Data Integration:

Extracting and cleaning three years of transaction data from the client’s ERP system, addressing inconsistencies in naming and free-text descriptions.

Rule Configuration

Working with the procurement team to define and encode enterprise-specific classification rules, ensuring alignment across decentralized units.

Training and Testing:

Training the model on historical data and validating it on a holdout dataset, achieving an initial accuracy of 92%.

User Training:

Conducting workshops to train the procurement team on the HITL interface and interpreting accuracy scores.

Deployment:

Rolling out the model in phases, starting with a pilot for the logistics and IT spend categories before scaling to all categories, ensuring compatibility with multiple taxonomies.

Results and Success Metrics

The deployment of Oraczen’s AI-powered spend classification model yielded transformative results for the client’s procurement operations:

Accuracy Improvement:

The model achieved a 95% classification accuracy rate, reducing misclassifications from 20% to less than 5%. This improved the reliability of spend analytics and sourcing strategies.

Time Savings:

Manual classification time was reduced by 85%, from 12,000 hours to 1,800 hours annually. This freed up procurement professionals to focus on strategic tasks like supplier negotiations and sustainability initiatives.

Cost Savings:

Accurate classifications enabled the identification of $30 million in annual cost-saving opportunities through optimized supplier selection and contract negotiations.

Example:

Improved Classification Accuracy Through HITL Feedback

Following the initial deployment of Oraczen’s AI-powered spend classification model, iterative feedback loops were implemented using human-in-the-loop (HITL) corrections and model retraining. As a result, discrepancies between the AI-classified and original category spend percentages were significantly reduced. For example, as shown in the latest evaluation, the AI’s predicted spend shares now differ by less than 1% from the original across all Level 1 taxonomy categories. This demonstrates the model’s ability to self-correct and improve through continued exposure to enterprise-specific data and expert validation—resulting in highly reliable and consistent spend classification over time.

Client Feedback

The success of Oraczen’s solution is reflected in the feedback from key stakeholders at the Fortune Global 500 client:

Chief Procurement Officer:

“Oraczen’s AI-powered spend classification model has been a game-changer for our procurement operations. The 85% reduction in manual classification time has allowed our team to focus on strategic initiatives, while the 95% accuracy ensures we’re making data-driven decisions with confidence.”

Procurement Analytics Lead:

“The human-in-the-loop interface is intuitive and empowers our team to refine classifications seamlessly. The model’s ability to handle our complex taxonomy and enterprise rules has significantly improved our spend visibility.”

IT Director:

“Integrating Oraczen’s solution into our existing systems was smooth, and the scalability of the model has supported our growing transaction volumes without compromising performance.”

Market Figures and Trends

The procurement industry is undergoing a significant transformation driven by AI adoption. According to a 2024 study by BCG, 60% of enterprise companies plan to invest over $25 million in AI-related projects, with procurement being a key focus area. Key market trends include:

AI Adoption Growth:

Up to 94% of procurement teams now leverage AI tools, particularly for spend analytics and supplier management.

Generative AI (GenAI):

GenAI is increasingly used for tasks like contract analysis and supplier communication summarization, enhancing efficiency.

Sustainability Focus:

AI is being applied to evaluate supplier ESG performance, aligning with corporate sustainability goals

Human-AI Collaboration:

HITL models are gaining traction to balance automation with human expertise, ensuring fairness and compliance. Oraczen’s solution aligns with these trends by combining ML, RAG, LLMs, and HITL capabilities to deliver a robust and future-ready procurement platform.

Impact on Procurement Business

Oraczen’s AI-powered spend classification model has redefined procurement operations for the Fortune Global 500 client and set a benchmark for the industry:

Operational Efficiency:

Automation of repetitive classification tasks reduced processing times by 70%, enabling faster decision-making.

Data-Driven Decisions:

Accurate spend categorization provided actionable insights into spending patterns, supplier performance, and market trends, empowering strategic sourcing.

Risk Mitigation:

The model’s ability to flag low-confidence classifications and incorporate enterprise rules minimized errors and ensured compliance with internal policies.

Strategic Focus:

By reducing manual workload, procurement teams could prioritize high-value activities like supplier relationship management and innovation.

Competitive Advantage:

The client gained a competitive edge by leveraging AI to optimize costs and respond proactively to market changes, aligning with Industry 4.0 principles.

Conclusion

Oraczen’s AI-powered spend classification model has transformed the procurement operations of a top Fortune Global 500 manufacturing client, delivering 95% accuracy, 85% time savings, and $30 million in cost-saving opportunities. By addressing the multifaceted procurement classification problem—including inconsistent naming, free-text descriptions, multiple taxonomies, and decentralized procurement—Oraczen ensured reliable and scalable spend categorization.

The model’s use of a RAG system to first retrieve relevant information, followed by LLM classification for novel cases, optimized the process for efficiency and accuracy. The flowchart illustrates the end-to-end process, highlighting the model’s ability to self-improve, as evidenced by reduced discrepancies in spend percentages across taxonomy categories. The enthusiastic testimonials from the client’s leadership further underscore the solution’s impact on efficiency, decision-making, and strategic focus. As AI adoption continues to accelerate, Oraczen remains at the forefront, empowering procurement teams to navigate the challenges of a dynamic global market.