Even in an era where ChatGPT can even handle coding, the real world of finance and accounting still struggles with Excel. Ultimately, it's Excel that's the answer.
Jung Min-gyu, CEO of ClovaNote, which oversaw the AI engine used by millions of people at Naver, observed what changed and what remained unchanged when AI was introduced into the workplace. The problem he identified was simple: technology advances rapidly, but work methods remain largely unchanged. Even with the introduction of AI, it merely adds tools to existing processes, leaving structural inefficiencies behind.
He left a stable leadership position at a major corporation to found KernelSpace in 2024. His goal was to redesign the very structure of work to fit the AI era.
CEO Jeong stated, "My career has always been a series of choices to get closer to customers. Only by directly listening to customers can I develop truly necessary products." He continued, "While I experienced significant impact at Naver, the nature of a large company meant there were clear limitations in directly hearing customers' voices and interacting with them quickly."
His conclusion to solving the problem is clear: rather than simply adding ways to better utilize AI, we need to redesign our work methods to fit the AI era.
"That kind of change is impossible without innovative work from within. I'm experimenting with new business grammars I believe in and trying to prove the results through products."
KernelSpace's product, "Greedy," is an AI-based spreadsheet automation platform. It's not just a chatbot; it thinks table-centrically and directly assists with spreadsheet tasks. It overcomes the limitations of general-purpose LLMs, which struggle to properly understand the context between rows and columns of structured data.
"LLM is fundamentally a text-generating model. It has inherent limitations when working with spreadsheets, which require strict adherence to row and column constraints. 'Greedy' uses natural language for explanations, but direct data manipulation is done through code. Instead of using words, it uses computable code."
When a user request comes in, the need is first translated into code. The request, expressed in natural language, is converted into executable logic, the code is executed, and the results are reflected in a spreadsheet.
"All code is generated with the spreadsheet's structure in mind. It's designed to reflect context—cell location, formula relationships, column and row meanings—so values aren't randomly generated, but calculated within the structure."
The execution results are then verified against the spreadsheet format and business context. Greedy's AI isn't a model that immediately provides answers; rather, it operates in a cyclical structure that translates the user's intent into code, executes it, and then re-verifies the results before finalizing the solution.
Excel files aren't isolated. Behind each sheet lie other linked files, repetitive workflows, and the essential context of the business. However, many AI tools focus on solving immediate problems and fail to neatly document the criteria and work history behind their actions. This can easily lead to repeating the same tasks from scratch or repeating similar mistakes.
"We need a structure that understands not only the state of the table, but also the state of the business. When the context and criteria of the process are recorded along with the results of the work, and when they are preserved as reusable workflows and accumulated as assets, the essence of the work is revealed and efficiency is achieved."
In integrating these critical issues and quality standards into the product, co-founder and CBO Park Sang-jung played a crucial role. A domain expert with experience at leading accounting firms, including the Big 4, and Channel Corporation, Park provided early market insights. From the product design stage, she closely monitored how Excel users work, the units within which accounting practices are structured, and the context in which business data is interpreted.
Greedy's core differentiator is workflow automation. When a user says, "Merge sales data," AI designs a workflow and saves it as reusable code. This shifts data power, previously reserved for a select few who know how to use macros or VBA, to the hands of everyday workers.
"Workflows, as we see it, represent the democratization of technology and a key means of capitalizing on practical work. User-defined workflows, defined in natural language, are organized into code-based execution logic, and in the process, the work context and judgment criteria become repeatable assets."
Workflows are much easier to create than VBA, and they're also more powerful. They can be created and modified using natural language, and unlike traditional VBA, where a single cell error can crash the entire system, they operate much more flexibly. He emphasized, "Users are experiencing everything from simple financial data organization to complex auditing tasks, all with the click of a button."
The value of repeatable execution lies not only in speed but also in consistency. For accounting tasks, accurate and consistent results are crucial. Manual or LLM-based workflows cannot guarantee this consistency. However, code-based workflows always execute the same code, so if properly configured, they guarantee consistent results.
Even major models like MS Copilot and Claude are enhancing their Excel support. However, CEO Jeong emphasizes that the focus isn't on the number of features, but on how the results are preserved and structurally accumulated.
Integration follows the same principle. Instead of relying on traditional methods, where users understand the API and configure detailed settings through a UI, agents should understand natural language requests and perform the necessary integration. Users don't need to know the technical details of the integration; they simply need to state their request: "Fetch the necessary data and settle the account."
More than half of the early testers moved beyond simple experiments and immediately transitioned to real-world automation. The biggest appeal of Greedy to practitioners is that they can start working the same way they're doing now. While many tools require installation, setup, and migration during the initial implementation process, Greedy allows users to simply upload their existing files and continue working on them.
Team plans, in particular, are expected to generate synergy effects because they share work outcomes and workflows. Because tasks performed once by someone are preserved as execution logic and can be reused, the more often they are repeated, the more valuable they become to the entire team. Individual experiences naturally translate into team standards, allowing organizations to leverage AI based on common principles without relying on the capabilities of specific individuals.
The future envisioned by CEO Jeong Min-gyu is clear.
"Simple data processing tasks will virtually disappear. The biggest difference between humans and AI is ultimately will. AI can calculate and organize, but it can't decide what to choose or in what direction to move. Human workers without willpower are increasingly likely to be excluded from their jobs."
Soon, we may find ourselves staring at the Excel UI less and less. The central screen might be a single table encapsulating all insights, or a single-line summary in natural language. Humans will have to make decisions based on data organized by AI.
In the short term, it's crucial to establish a product that's reliably used in the field. Currently, it's integrated with Gmail, Google Drive, Analytics, payment systems like Stripe, and the corporate expense management platform Spendit. Going forward, we plan to expand integrations with various SaaS (Software as a Service) platforms, including commerce, accounting, and internal databases, ensuring business data operates seamlessly within a single execution context.
In the long term, KernelSpace aims to move beyond spreadsheets and connect databases and dashboards, creating an environment where data is accumulated, insights are shared, and execution logic remains an asset. Ultimately, KernelSpace's mid- to long-term goal is to evolve beyond tools providing individual functions into a business execution infrastructure that reliably enables AI to perform corporate tasks.
"Ultimately, future competitiveness doesn't depend on the ability to handle more work, but on the power to set a clear direction with determination and then push through to execution. 'Greedy' aims to be a tool that helps users with such determination move one step ahead into the future work environment."
Leaving behind the success of Naver Clova Note, the developer's tenacity as he jumped back into the rough startup world is drawing attention to how it will change the way all practitioners who "wrestle" with Excel work.
You must be logged in to post a comment.