Introduction Artificial intelligence (AI) agents are beginning to move beyond answering questions and toward acting on users' behalf. In online commerce, they can research products, identify a suitable payment method, and complete a purchase. Their potential applications could eventually extend to deposits, loans, investments, insurance, and other financial transactions. But the closer an AI agent comes to moving money or selecting financial products for a consumer, the more difficult the questions become: What authority can a user safely delegate? Which activities require a license? And who is responsible when an autonomous system makes an error? In the Republic of Korea (ROK), fully autonomous AI systems that directly execute payments and financial transactions appear unlikely to be deployed in the near term under the current regulatory framework. Advisory, comparison, and recommendation services, however, are likely to emerge much sooner. This creates a policy challenge. Existing rules help contain risks, but they may also limit how quickly regulated financial institutions can experiment as technology companies expand into financial services. Policymakers need to examine how regulation can keep pace with technological advances without weakening consumer protection. Analysis From assistance to autonomous payment Discussion of AI agent payments has accelerated alongside the development of agentic commerce. Major AI companies, including Google, OpenAI, and Anthropic, as well as cryptocurrency firms, have announced protocols such as the Universal Commerce Protocol, Agent-to-Payments, and x402 and have begun pilot programs for agentic commerce services. Visa and Mastercard have also formed consortia with banks and fintech firms to test AI agent payments and explore approaches to payment tokenization. AI agent services typically combine three elements. Model Context Protocol connects an AI system to external tools. A large language model performs reasoning, such as deciding which product to purchase or which payment method to use. Orchestration logic coordinates the interaction between the model and the tools. In an AI agent payment, this architecture connects the system to payment functions so it can help determine whether and how a transaction should proceed. Payments raise higher stakes than many other AI-enabled tasks because they depend on payment tokenization and the delegation of functions such as processing, authorization, authentication, and payment authority. Tokenized credentials require governance over what the agent may do, including spending limits and restrictions based on category or time. They also require clear authentication requirements, approval rules, and audit mechanisms. The risks do not end with security. Errors by an autonomous payment system could cause significant financial losses and disputes over liability. Fintech firms, AI service providers, and financial institutions have therefore generally approached real-world deployment cautiously while they assess its potential. Why advise may arrive before autonomous transactions AI agents could eventually extend beyond online purchases into the wider retail financial sector. Many products, including deposits, loans, investments, and insurance, can be compared using objective criteria such as interest rates and insurance premiums. This makes comparison and recommendation a natural early application for AI in finance. Such services are already emerging. One AI-driven service operating under a regulatory sandbox uses MyData services to exercise a borrower's right to request an interest rate reduction automatically when the borrower's credit profile improves.[1] Services that automate investment decisions have also appeared. Yet comparison, recommendation, and advice are not outside financial regulation. Under Korea's Financial Consumer Protection Act, the use of AI to compare or recommend financial products, including investments and insurance, may be regarded as solicitation and therefore be subject to existing licensing requirements. AI-powered financial advice may likewise fall within licensed financial advisory or quasi-investment advisory services. These services also raise substantive consumer-protection concerns. An AI recommendation may involve conflicts of interest or fail to satisfy the suitability principle. Because generative AI outputs have limited transparency, questions also arise over whether an AI-based service can provide the reliability, ethical standards, and professional expertise expected from a human financial advisor. Direct transactions face still greater constraints. The ROK's financial system is based on account holders acting on their own behalf, supported by real-name verification. Under the Electronic Financial Transactions Act, users are prohibited from delegating payment authority or access credentials to third parties.[2] The Act on Real Name Financial Transactions and Confidentiality also requires financial transactions to be conducted in the account holder's name, with strict procedures governing delegation to an agent.[3] Under this framework, delegating financial transactions to AI remains practically infeasible. The likely sequence is therefore clear: AI agents may first gain traction in advisory and recommendation services, rather than in direct payments or financial product transactions. But earlier adoption does not mean lower regulatory importance. A finfluencer's sale of a self-developed automated stock-trading program without registration as a discretionary investment manager illustrates the risk of AI-enabled services operating without the required license. A widening gap between financial institutions and technology firms As AI-based services gain ground, they could reshape competition in financial services. Regulated financial institutions may face a relative disadvantage under the current regulatory environment. Network segregation requirements have encouraged them to adopt the latest AI technologies cautiously in credit information systems. Concerns about liability have made them especially reluctant to deploy AI agents in customer-facing or payment-related applications. Technology firms, meanwhile, are actively experimenting with new financial services. The growing scalability of large language model-based services has lowered entry barriers to finance for technology companies and other non-financial firms. Anthropic, for example, has introduced AI agents designed to automate financial tasks that include research, modeling, fund management, and advisory services. Model Context Protocol-based services have also emerged, including Alpaca MCP Server and TradingView MCP Agents. Powered by models such as Claude, these services are designed to connect AI agents directly to brokerage accounts on platforms such as Robinhood. This produces a difficult regulatory balance. Existing laws restrict the delegation of financial authority and help contain the risks associated with AI agents. But if licensed institutions remain unable to test new services while technology firms advance, the ROK may fall behind developments in the global market. The challenge is therefore not simply whether to permit AI agents in finance. It is how to create room for controlled experimentation while maintaining the safeguards required when systems advise consumers or act on their behalf. Implications A gradual approach can allow regulators and financial institutions to learn from controlled applications before AI agents are used more broadly. In the short term, financial regulatory sandboxes should be used to test AI agent payments. Sandboxes can give financial institutions space to experiment with new technologies within a supervised environment while policymakers observe how existing rules apply. Over the medium term, eligibility for this experimentation should be expanded to cover a broader range of financial products, accompanied by refinement of the existing regulatory framework. This would allow regulation to respond to the different issues raised as AI moves from comparison and recommendation toward more consequential financial decisions and transactions. Over the longer term, policymakers should strengthen institutional safeguards for consumer protection, including consumer education. Unlicensed AI-powered financial services may cause consumer harm, and consumers need better information about the risks of AI deployment and the liability issues associated with its use. AI agents could change both how financial services are delivered and who is able to provide them. The ROK's current framework means that advice and recommendations will likely advance before autonomous payments. That window should be used to test technologies carefully, clarify regulatory responsibilities, and strengthen consumer protection. Regulation need not abandon existing safeguards to support innovation, but it must begin adapting before AI agents move from advising consumers to acting for them. [1] Under Article 30-2 of the Banking Act, borrowers that have entered into credit agreements with banks may request a lower interest rate if their credit profile improves, including through increases in assets, credit ratings, or personal credit scores. [2] Article 9(3) of the Electronic Financial Transactions Act and Article 8 of its Enforcement Decree. [3] Article 3 of the Act on Real Name Financial Transactions and Confidentiality. Resource Korea Institute of Finance website. Ask the Experts Yeonju Baik Research Fellow, Banking and Financing Holding Companies Research Division, Korea Institute of Finance Yeonju Baik is a research fellow in the Banking and Financial Holding Companies Research Division at the Korea Institute of Finance. Her research focuses on banking, digital and AI transformation in finance, and bank competition. She holds a PhD and an MA in economics from the University of Wisconsin-Madison and an MA in economics from Seoul National University. Korea Institute of Finance (KIF) The Korea Institute of Finance provides expert analysis for the development of the Republic of Korea's financial sector and financial policy. Leave your question or comment in the section below: View the discussion thread.