Schwab already solved this problem. So did Fidelity, Vanguard, and nearly every major retail brokerage in the United States. Intelligent portfolio accounts, available to ordinary customers for years, deliver algorithmic investing inside a structure that includes fiduciary obligation, disclosed methodology, institution-controlled risk parameters, and regulatory oversight. They work. Retail investors use them at scale without understanding the underlying models, and that is precisely the point: the institution absorbs the complexity so the customer does not have to.

What is happening now in Hong Kong, Singapore, and China is structurally different, and the difference is not the presence of artificial intelligence. Machine learning has powered quantitative trading since the early 2000s. High-frequency trading firms run it at microsecond precision. The robo-advisor products at Schwab and Fidelity are ML-driven products sold to mass-market consumers. None of that is new.

What is new is the removal of the institutional layer.

The Customer Is Now the Algorithm Author

Retail investors in Asia are connecting large language models to brokerage application programming interfaces themselves, feeding in strategies written in plain language, and letting the model execute. Futu Holdings, which operates the Futubull and Moomoo platforms in Hong Kong, launched a service that lets users generate quantitative trading strategies without writing code. East Money, one of China's largest retail brokerages, ran a simulated competition in April using an agent called OpenClaw to trade a portfolio of stocks. The simulation framing is notable: brokerages are stress-testing agentic workflows in controlled conditions while simultaneously shipping the underlying connectivity in production.

In the Schwab model, the institution sets the strategy, controls the risk parameters, holds the fiduciary obligation, and answers to regulators when something goes wrong. In the LLM-connected API model, the customer does all of that, with none of the training any of those roles require and none of the backstop any of those roles assume.

Futu's Vincent Yao, head of the AI growth center, described a positive correlation between AI usage and trading activity. That correlation is not evidence of better outcomes. It is evidence of more activity, which generates commission revenue for the brokerage regardless of whether the customer makes money.

Regulators Are Applying 2015 Rules to a 2026 Product

Hong Kong does not have AI-specific rules for financial institutions. The Securities and Futures Commission published a circular in November 2024 classifying AI-driven investment recommendations as high-risk, urging licensed firms to implement mitigation measures. Those rules license the advice-giver. They do not reach the model, and they do not reach the customer who wrote the strategy in plain language and handed execution to an agent.

Singapore moved further. The Monetary Authority of Singapore released AI risk management guidelines in November 2025 requiring institutions to assess risk at the use-case level before deployment, and followed with an implementation toolkit in March 2026. Both documents assume an institutional actor doing the assessing. Neither has a mechanism for the retail customer who has become the institution.

China has the most concrete enforcement record. A Shanghai securities firm was fined two million yuan in 2025 for failing to disclose the limitations of its AI-generated investment recommendations. That fine establishes that disclosure obligations apply to model behavior. It also applies to a licensed securities firm, not to a retail customer running an agent through a personal brokerage account.

David Friedland, Asia-Pacific managing director at Interactive Brokers, named the operational gap without softening it: "We can measure the speed and the duration of the orders coming in, but that can be a computer program too, so it's really hard for us to say what people are using." Order surveillance was built to detect patterns. It was not built to classify the origin of intent. An LLM routing trades through an API is indistinguishable from a rule-based script.

Performance Is the Least of the Problems

At the end of 2025, Nof1 handed eight leading AI models ten thousand dollars each to trade U.S. technology stocks over two weeks, with full latitude on risk-taking and leverage. Only six of thirty-two attempts produced positive returns, according to figures reported by Nikkei Asia. The winning Grok 4.20 instance returned 34.59%. The majority of attempts lost money.

Fidelity Go does not lose money on 26 out of 32 attempts. The institutional product works because the institution controls the model, sets the parameters, and is accountable for the outcome. The retail DIY version carries none of that structure.

Friedland put the risk plainly: "All of a sudden, with AI, everyone can look like the world's greatest options trader, but the reality is they need training and experience to understand the risk behind their trading."

Last week, Anthropic launched ten new AI agents targeting financial services workflows: pitchbooks, know-your-customer screening, closing books. Those are institutional workflows. The pattern is consistent with how agentic AI enters regulated industries. It arrives where the revenue justifies the integration cost, then diffuses to wherever the API is open.

Tiger Brokers reported that TigerAI reached ten million cumulative global interactions by end of March 2026, a 500% increase since launch, with total user base growing 148% over the same period. The scale is real. None of those figures say whether the interactions produced better outcomes than the Schwab product those same customers could have used instead.