Your AI agent can now trade your money: Should it?
FINRA warns AI trading agents can act beyond their authority, and platforms like Coinbase already let them try
Handing an AI system the ability to inspect your portfolio, interpret your instructions, and actually execute a trade sounds like a future scenario. It's already live.
Coinbase launched Coinbase for Agents in June 2026, letting AI assistants like ChatGPT and Claude connect directly to user accounts to trade, pay, and run workflows within limits the user sets.
German broker Scalable Capital has since opened its own platform to major AI assistants for portfolio analysis and trade initiation.
Conventional robo-advisors function within predefined, limited models: you answer a few questions, receive your allocation of some predefined ETFs, and then experience periodic rebalancing.
In contrast, when an AI trading agent is told "reduce your exposure in case of a dramatic increase in volatility", the agent must interpret what this really means and execute this interpretation.
When a chatbot makes a poor choice of stocks, it is frustrating. If an AI agent with the power to execute trades gets it wrong, it is a costly error.
FINRA's 2026 Regulatory Oversight Report specifically mentions this problem of scope and authority of the agents; they act on behalf of the users in ways that go beyond the users' intentions, and the multistep agent reasoning process becomes hard to audit afterward.
The regulator also warns that general-purpose agents may simply lack the financial-domain knowledge to execute complex, sector-specific tasks reliably.
If a large number of agents are trained on identical information and arrive at identical conclusions at the same time, FINRA has identified risks which could result in herding or correlated movements of the market, different from the risk of any individual agent’s failure.
Machine learning technology is particularly prone to fail precisely where it needs to work best: FINRA points out that unexpected situations such as geopolitical events or volatility occur outside the experience of the model.
Delegated permission structures, capping order sizes, blocking withdrawals, excluding leveraged trades, and requiring approval before selling long-term holdings are already emerging as the standard architecture rather than granting an agent unrestricted account control.
