Introducing Toss Place's Data Bot 'PANDA': How Every Team Member Can Work Like a Data Expert

PANDA stands for Place Analytics & Data, a data analysis assistant that helps Toss Place team members directly query and analyze the data they need within their security clearance.

The motivation behind its introduction was clear. 70% of all data requests were not complex analyses but simple extraction tasks to check exact figures, and analysts had to check dozens of dashboards or write SQL themselves. PANDA reduced these repetitive requests and allowed data analysts to focus on higher-density work.

To improve accuracy, Toss Place's Data Analysis Team and Data Platform Team first organized a standard data mart that serves as a single source of truth (SSOT).

• Organized core data such as store information into a unified table
• Organized a DW standardization convention so the same concept isn't referred to by different names
• Densely filled in table and column names and Descriptions to help AI better understand meaning

Since data structure alone wasn't enough, business language was also connected. To answer questions like "What is the criterion for an installed store?" or "What is the basis for industry classification?", domain terminology and metric definitions were bundled together with the standard mart to align on a single standard.

PANDA also applied a Scoring & Ranking system so that it doesn't pick tables arbitrarily. It looks at similarity and reliability together, and only brings tables managed with dbt tags into the Manifest, narrowing the search scope to ensure accuracy and consistency.

Its operation resembles an Agentic Loop. PANDA repeats table exploration, query execution, result verification, re-questioning the user if necessary, and retries and corrections to produce the final answer. Rather than simply being a chatbot that gives an answer in one shot, it's a structure that solves problems by choosing and reviewing tools on its own.

Answers were also designed to not just throw out numbers. To help users immediately understand and use the results, PANDA aims to provide the data analyst's thought process alongside the answer.

The response after launch came quickly.

• 1 in 3 team members used it on the first day of opening
• Half of all team members experienced it within a week
• Over 4,000 messages were generated during the same period
• Currently, 70% are using it

Interestingly, it was actively used by unexpected job roles such as developers, and data professionals also directly asked PANDA questions to check its reliability. Toss Place plans to continue advancing PANDA toward its goals of 90% or higher data coverage and 97% or higher answer accuracy.

The key was not AI itself, but the way it solved actual pain points. Rather than complex technology, PANDA's goal is to reduce repetitive data requests within the organization and enable anyone to check, understand, and use data on their own.