Challenges of Generative AI and Datumo's Role in Enhancing Trust and Safety

Many organizations face challenges in fully preparing to implement generative AI in a safe and responsible manner, according to a recent report by McKinsey. Among the most prominent of these challenges is interpretability, which is concerned with understanding how and why AI systems make their decisions. While 40% of McKinsey survey participants view this as a significant risk, only 17% of them are actively working to address it.

The Emergence and Development of Datumo

The company Datumo was founded in Seoul as an entity specializing in AI data classification, and has expanded to help companies build safer AI models. Datumo provides tools and data that allow models to be tested, monitored, and improved without the need for advanced technical expertise. The company recently succeeded in raising $15.5 million, bringing its total funding to approximately $28 million, with prominent investors such as Salesforce Ventures, KB Investment, and SBI Investment participating.

David Kim, Datumo's CEO and a former researcher at Korea's Defense Development Agency, was frustrated by the long time it took for data classification. From this, an innovative idea was born: a reward-based application that allows anyone to classify data in their spare time for money. This idea proved successful in a startup competition at the Korea Advanced Institute of Science and Technology (KAIST). In 2018, Kim co-founded Datumo, formerly known as SelectStar, with five KAIST graduates.

Datumo's Success in the Korean Market and Service Expansion

Even before the full launch of the application, Datumo managed to achieve tens of thousands of dollars in pre-contract sales during the customer discovery phase of the competition, with most of these clients being companies and startups led by KAIST graduates. In its first year, the company's revenue exceeded $1 million and it secured several major contracts. Today, its client list in Korea includes major companies such as Samsung, Samsung SDS, LG Electronics, LG CNS, Hyundai, Naver, and telecom giant SK Telecom. As the market grew, clients began demanding that the company provide services beyond just data classification. The seven-year-old startup now has over 300 clients in South Korea and achieved revenues of approximately $6 million in 2024.

Michael Hwang, Datumo's co-founder, told TechCrunch: "Our customers wanted us to evaluate the outputs of their AI models and compare them to other models. At that moment, we realized we were already evaluating AI models without realizing it." Hwang added that Datumo has strengthened its efforts in this area and released Korea's first benchmark dataset focusing on AI trust and safety.

Competition and Excellence in the AI Classification and Evaluation Market

Kim told TechCrunch: "We started with data classification, then expanded to include pre-training and evaluation datasets as the large language model (LLM) ecosystem matured."

Meta's recent acquisition of data classification company Scale AI for $14.3 billion demonstrates the importance of this market. Shortly after this deal, OpenAI, a competitor to Meta, stopped using Scale AI's services. The Meta deal also indicates that competition for AI training data is intensifying.

Datumo is similar to companies like Scale AI in providing pre-training datasets, and to companies like Galileo and Arize AI in AI evaluation and monitoring. However, it distinguishes itself with its proprietary licensed datasets, particularly those extracted from published books, which provide rich and structured, yet difficult to clean, human inference, according to CEO Kim.

Unlike its competitors, Datumo offers an integrated evaluation platform called Datumo Eval. This platform works to create test data and automated evaluations to check for unsafe, biased, or inaccurate responses without the need for manual programming, as Kim explained. The flagship product is a no-code evaluation tool, designed for non-developers such as policy, trust, safety, and compliance teams.

Attracting Global Investments

When asked how they attracted investors the size of Salesforce Ventures, Kim explained that the company had previously hosted a discussion session with Andrew Ng, founder of DeepLearning.AI, at an event in South Korea. After the event, Kim shared the session on LinkedIn, which caught the attention of Salesforce Ventures. After several meetings and Zoom calls, the investors made an initial commitment. The entire funding process took approximately eight months, according to Hwang.