Architecting the Future of Procurement: The AI-First Approach to Data Standardization

Turkcell is Turkey's leading technology provider, offering integrated voice, data, and value-added services like streaming via mobile and fixed networks. On July 11, 2000, it became the first Turkish company listed on the NYSE. As a key regional player, Turkcell is dedicated to delivering high-quality communication with wide coverage in Turkey and expanding its services abroad.

Through AI-driven standardization of supply chain product and service codes, Turkcell Procurement repositioned artificial intelligence from a supporting analytics tool to the core governance mechanism for procurement data. A new operating model was established in which data creation, classification, and governance are executed by AI, resolving data quality at the source and enabling a scalable foundation for business intelligence.

At the heart of this transformation is the Turkcell Supplier Relationship Management System (SRMS) Code Definition Tool (CDT). Led by the Data Quality Management Team-established in H2 2024-the initiative followed a precise roadmap: The team assumed full process responsibility in H1 2025 and, working in close collaboration with key stakeholders including Obase, Turkcell ICT & AI Teams and Finance, successfully completed the project in H2 2025. This system digitized the previously manual code request process end-to-end, embedding Large Language Models (LLMs) and similarity algorithms as an integral part of the live workflow.

The process flows as follows:

Prevention First: In the initial phase, the system semantically analyzes the user's free-text description and performs a similarity match against the active code data pool. If a functionally equivalent item is detected, the AI automatically suggests the existing code. This mechanism prevents duplication at the very start, ensuring the reuse of defined assets.

Intelligent Extraction: If no match is found, the AI initiates the LLM-based attribute extraction phase. Here, the description is parsed into specific attribute sets, which are then standardized, normalized, and mapped to the correct United Nations Standard Products and Services Code (UNSPSC) hierarchy in real-time.

External Enrichment: For descriptions insufficient to extract reliable attributes, the system autonomously researches the item online to make meaningful assignments. This staged architecture transforms code definition into an AI-governed process that guarantees data integrity. This project transformed procurement data from an operational byproduct into a strategic asset.

By embedding AI directly into the workflow, governance and standardization are automated at the point of data creation, enabling users to quickly and reliably select the correct codes and driving high adoption. As a result, data quality is significantly improved, making UNSPSC-based spend analysis, supplier and tail-spend assessments, and price-cost modeling reliable. Reporting has therefore shifted from retrospective visibility to a strong foundation for strategic, data-driven decision-making.

By centralizing authority from ~400 users to a 7-person expert team, we established ahuman-in-the-loop model. AI executes heavy-lifting tasks at scale, while experts handle validation, ensuring data consistency and control.

Measurable Strategic Impact:

• ~40,000 active codes standardized by AI.
• ~80,000 unused codes were closed.
• 300+ attribute sets were generated by AI.
• 2,000+ UNSPSC mappings were completed by AI.
• All new code requests are now executed through the AI-supported SRMS CDT process.

Beyond these metrics, the project's technical value was validated academically. The system-side development was presented by the Turkcell ICT Team and Obase as 'Enhanced Procurement Through AI-Driven Product and Category Management.'

Launched in 2025, this framework establishes a future-ready, AI-compatible data foundation. By enabling comprehensive data analysis, it clearly demonstrates that sustainable AI value is achieved not by layering analytics over flawed data, but by embedding intelligence directly into the system architecture.