AI cannot bear liability for losing trades, responsibility follows delegation: Brickken CEO

AI agents have started executing trades and moving funds without constant human approval, prompting Brickken CEO Edwin Mata to argue that liability must follow the authority granted to the software rather than attach to the AI itself.
Summary

• AI agents cannot assume legal duties because current law does not recognize them as legal persons.
• Mata said principals normally bear the outcome when agents act within an authorized mandate.
• ERC-8226 proposes time limits, financial caps, revocation controls, and verifiable records for AI agents.
• U.S. securities rules already require broker-dealers to control automated systems that access regulated markets.

Sandmark reported on Aug. 6 that existing laws provide no single answer for losses caused by autonomous financial agents, leaving courts to examine the user, developer, platform, and institution involved in each transaction.

The report said contract law, negligence rules, product liability, and fiduciary duties could all apply, depending on who controlled the agent and what caused the loss. A user may bear the result of an authorized trade, while a developer or platform could face claims if faulty design, weak safeguards, or corrupted information pushed the agent outside its intended role.

Commenting on the issue, Edwin Mata, a lawyer and the CEO and co-founder of tokenization platform Brickken, told crypto.news that responsibility should never be assigned directly to the software.

"Under current law, AI is not a legal person capable of assuming duties or bearing liability. It is a technical system acting on behalf of a natural or legal person."

According to Mata, an investigation should instead establish who authorized the agent, whose interests it represented, and what powers it received. Such an inquiry would help distinguish a losing decision made within an approved strategy from a transaction that broke the agent's limits.
You might also like: Crypto research has a joining problem, and AI agents are starting to solve it

AI agent liability follows the granted authority

Mata compared the legal relationship to a power of attorney, under which one party receives permission to act for another within a defined scope. When an issuer, bank, or investor authorizes an agent to transact, he said, the principal would ordinarily bear the consequences of actions that remain within that authority.

Under the same reasoning, an investor could not reject a trade simply because the software produced an unfavorable result. A price loss does not by itself show that the agent acted without permission or that another party failed in its duties.

"An issuer cannot disown an unfavourable but authorised transaction merely because the decision was generated by software," Mata said.

Responsibility may change when an agent exceeds its mandate. Mata said a developer, platform, or financial institution could face exposure if its design or controls caused or allowed the failure, although the final assessment would depend on the facts and applicable law.

Sandmark cited similar legal distinctions in its report. Chanté Eliaszadeh, founder of Astraea Counsel, told the publication that liability would generally follow control. She said users are usually the starting point when agents act on their behalf, but developers could face risk if a system marketed for autonomous trading failed in a foreseeable way.

The question has gained urgency as agents obtain direct access to wallets and payment systems. In May, a Keyrock report found that AI agents had settled $73 million through 176 million transactions during the previous 12 months, with USDC accounting for 98.6% of the payments examined.

Coinbase has also connected agents to trading, portfolio management, and payments under user-set limits. By July, Chainalysis had counted more than 100 million x402-linked payments on Base, although the analytics firm said meme-coin farming and automated activity contributed to the early transaction totals. The figures therefore did not represent only independent agents buying goods or services.

Human approval needs clear and enforceable limits

While a person may formally approve an agent's activity, Mata said consent alone does not provide meaningful control if the person cannot understand the authority being granted.

Effective delegation, in his view, requires a list of permitted actions and eligible assets, along with limits for individual transactions and total spending. A mandate should also specify its duration, the conditions requiring human review, the principal's right to revoke access, and a record of every action taken.

Such controls are already appearing in commercial products. Anchorage Digital introduced agentic banking in May with verified identities, spending limits, and audit controls for autonomous systems accessing crypto and traditional payment rails.

Visa and Wirex have separately tested agent-led stablecoin payments for software subscriptions, marketing budgets, and procurement. According to Wirex, the trials were designed to examine security, reliability, transparency, and consumer control when software initiates payments for a user or business.

A June guide to agentic payments explained how x402 allows autonomous software to pay for data, computing services, and online resources using stablecoins. Because those payments can occur without a person approving each transaction, authorization systems must establish what the agent can buy, how much it can spend, and when its access ends.

ERC-8226 would record AI agent mandates onchain

Mata pointed to ERC-8226, the proposed Regulated Agent Mandate Standard, as one model for making delegated authority verifiable.

Filed as a draft Ethereum standard on April 12, ERC-8226 is designed for AI agents operating with tokenized regulated assets. The proposal was written by Brickken contributors Ludovico Rossi, Dario Lo Buglio, Thamer Dridi, and Nabil El Alami Khalifi.

Known as RAMS, the standard would let a verified principal give an onchain agent permission that is limited by asset, action, duration, and monetary value. A regulated token contract could check the mandate when the agent tries to execute a transaction.

The proposal separates three questions that may arise during an agent-led trade. An identity registry would confirm that the agent exists, a compliance provider would determine whether the principal is eligible to transact in the asset, and the RAMS registry would verify whether the planned action falls within the delegated mandate.

Under the draft specification, a mandate could set a maximum amount for one transaction and a cumulative amount across multiple transactions. It could also include activation and expiry times, allowed assets, approved actions, revocation functions, and records showing how much authority the agent has already used.

Mata said RAMS would not transfer liability to the agent or reimburse a principal for an authorized loss. Instead, the proposed standard would provide evidence showing who granted the authority, what the agent could do, and whether the transaction remained within those limits.

"Its purpose is to make attribution verifiable: who granted the authority, what the agent was permitted to do, whether it remained within those limits and which person or control failed when it did not."

ERC-8226 remains a draft rather than an adopted Ethereum standard or legal requirement. Its discussion page also lists unresolved questions, including whether tokens purchased by an agent should remain in the agent's wallet or settle directly into the principal's wallet.

U.S. rules keep responsibility with regulated firms

For U.S. markets, existing securities rules already place duties on the firms that provide access to exchanges and alternative trading systems.

Under SEC Rule 15c3-5, a broker-dealer providing market access must maintain financial and regulatory risk controls under its direct and exclusive control, subject to limited exceptions. SEC guidance says the broker-dealer remains responsible for the effectiveness of those controls even when it uses technology supplied by an independent third party.

The rule requires automated pre-trade checks designed to stop orders that exceed preset credit or capital thresholds. It also requires controls that restrict trading systems to authorized people, block prohibited securities transactions, and deliver immediate execution reports to surveillance staff.

For consumer payments, Regulation E requires preauthorized electronic fund transfers to carry a written or similarly authenticated authorization from the account holder. CFPB guidance also says the authorization process should demonstrate the consumer's identity and agreement, while allowing the consumer to stop or revoke future payments under specified procedures.

Current CFPB rules do not directly state how a standing instruction such as "manage my portfolio" should apply when an AI agent independently selects and executes individual transfers. Sandmark reported that lawyers remain divided over whether a manipulated agent payment would resemble an unauthorized transfer caused by stolen credentials or an authorized transaction carried out under previously granted access.

Outside the United States, Bank of England Deputy Governor Sarah Breeden said in June that financial oversight frameworks were not designed for autonomous agents and that requiring human approval for every action may be unrealistic. She said regulators were considering stronger safeguards, including circuit breakers or market-wide kill switches if faulty AI models threatened trading systems.
Read more: Digital money needs interoperable settlement rails, Lynq CEO says