Fiduciary Duty Is Coming for AI Agents. Crypto Is Asleep at the Wheel.

The U.S. Securities and Exchange Commission spent most of 2024 and 2025 quietly building a legal scaffold around artificial intelligence. In March 2024, it settled fraud charges against two investment advisers — Delphia and Global Predictions — for making false claims about their AI capabilities. The fines were small. The message was not. The agency used its anti-fraud toolbox to declare that AI-related statements must be accurate, that hype has a legal cost. Then, in December 2025, the SEC issued a risk alert on marketing rules, extending the same logic to promotional practices. And in its 2026 examination priorities, the regulator explicitly named AI-driven investment advice as a fiduciary concern. That last phrase matters. Fiduciary concern. Not disclosure concern. Not transparency concern. Fiduciary — the legal obligation to act in the best interest of a client, to place the client above yourself, above your counterparties, above your own revenue. Here is the uncomfortable truth for the crypto industry: the same framework is now being drafted for all AI agents. Senator Mark Warner’s discussion draft of the AI AGENT Act proposes non-waivable fiduciary duties for AI agent developers and deployers. The FTC published an AI accuracy policy statement on July 1, 2026, and opened comments until September 18. Stanford’s HAI has circulated a fiduciary duty paper. The pieces are converging. The question nobody in the digital asset space is asking is simple: what happens to DeFi when the AI agent you deploy is legally bound to act as a fiduciary? The old paradigm was built on transparency. Tell the user how the model works, what it can do, what its limitations are. Let the user decide. Disclose the risks, obtain consent, move on. That was the regulatory comfort food of the 2010s. It is now being tossed aside. The new paradigm is built on loyalty. A developer or deployer of an AI agent must act solely for the user’s benefit. That means no hidden affiliate fees. No quiet supplier ranking. No self-dealing. No structural incentive that pushes an agent to recommend a product because the product paid for placement. The burden shifts from the user’s ability to read fine print to the developer’s obligation to be faithful. Enforcement moves from "what did you disclose?" to "why did you make that decision?" — and that is a seismic change for every business model that currently monetizes AI through recommendations, commissions, or behavioral steering. I have spent the last three years studying liquidity mechanics and regulatory frameworks in Southeast Asia. I learned that liquidity is a mirage; only settlement is real. But here, settlement means something deeper: the finality of legal obligations. And the legal obligation being settled across Washington right now is that AI agents — software entities that can transact, negotiate, and act autonomously on behalf of users — will be treated as extensions of a fiduciary, not as neutral algorithms. That is the core finding of this analysis. Let me break down the three layers that are converging. First, legislation. The AI AGENT Act is still at discussion draft stage. But the design choices already reveal intent. The Act assigns enforcement to the FTC rather than creating a new agency. That is not a technical detail. It means the existing consumer-protection machinery — specifically Section 5 of the FTC Act, which prohibits "unfair or deceptive acts or practices" — will be the vehicle for fiduciary enforcement. No grand new bureaucracy. No wait for a digital regulator. Just a policy statement, a comment period, and an enforcement sweep. The Act also makes the duties non-waivable. Even if a user signs away claims, click-through agreements become void. Loyalty cannot be contracted around. The Act imposes duties on developers and deployers, not on the AI systems themselves. That is deliberate. It sidesteps the philosophical question of AI personhood entirely by making humans accountable for the actions of their software. Second, enforcement. The SEC’s 2024 settlement and 2025 risk alert establish the initial precedent. The SEC is using anti-fraud statutes to police AI claims, not waiting for new laws. But the deeper signal is in the examination priorities for 2026: checking whether AI-driven investment advice actually complies with fiduciary duties. The agency is not just looking for false claims anymore. It is looking at the incentive structure behind the AI. If an AI agent recommends a product because of an affiliate relationship, the SEC wants to know. If it steers users toward in-house products, that is self-dealing. The FTC’s proposed policy statement on AI accuracy is even broader — it reaches any AI agent that misleads consumers, regardless of whether it is in finance. Third, academia. Stanford’s HAI has been circulating a paper that maps fiduciary obligations onto AI agent developers. The paper argues that because agents act with partial autonomy, users cannot fully monitor them. Thus, the law must impose a duty of care and a duty of loyalty on the developer. This is not idle theory. It is a roadmap for courts and rulemakers. The combination of a credible Senator, an active SEC, an aggressive FTC, and a Stanford-backed framework means that the fiduciary duty for AI agents is no longer a hypothetical. It is an emerging legal fact. Now let me talk about what this actually means for the crypto industry, because that is where the impact will be most disruptive. Consider the typical AI agent in a DeFi context. It monitors yield across protocols, executes swaps, rebalances positions. That agent is being paid — indirectly — by the protocols it routes through. Referral programs, liquidity mining incentives, token rewards. The current architecture is an affiliate fee machine. Under the fiduciary framework, every one of those incentives is a conflict of interest. The agent’s developer must either strip out all fees from third parties or redesign the agent so that user interest is provably prioritized. That is not a marginal compliance tweak. It is a business model overhaul. The FTC’s policy statement, if finalized, will apply to any "unfair or deceptive acts" — and the FTC has already signaled that hidden incentives count as deceptive. The SEC’s investment-adviser enforcement means any token or protocol that integrates an AI advisor is itself exposed to fiduciary scrutiny. Let that sink in. A yield aggregator that uses an AI agent to recommend vault strategies is likely to be treated as an investment adviser. That means registration, custody rules, best execution obligations, and — yes — fiduciary duties. The entire "we are not a financial advisor" disclaimer becomes worthless. The transparency paradigm is dead. Data proves nothing anymore. What matters is the decision logic. How did the AI arrive at this recommendation? What incentives were embedded in its objective function? Were any suppliers secretly prioritized? These are not questions that a white paper can answer. They are questions that require audit trails, model interpretability, and incentive-structure disclosure. The technology to do this well does not exist yet. The article’s source material explicitly notes that "the technical challenge of auditing agent behavior remains unresolved." I can confirm this from my own work in the space. I have audited liquidity pools, but auditing the internal reward function of an autonomous agent is a different order of magnitude. It is the difference between checking a balance sheet and deconstructing a brain. Here is the contrarian angle most analysts will miss. The fiduciary framework will not protect the little guy. It will crush him. When compliance costs include building an AI behavior audit system, hiring a Chief Trust Officer, restructuring revenue models, and defending against private lawsuits from users who claim the agent breached its duty of loyalty — the only players who can survive are those with deep legal teams and capital reserves. The SEC’s "check-warn-enforce" rhythm is not a gentle ramp. It is a throttle that rewards the incumbent. In crypto, that means the large exchanges and institutional custodians with existing compliance departments will absorb the AI agent market while small indie developers are squeezed out. Trust is the new collateral, and only the well-funded can post it. The second contrarian point: decentralization does not save you. The AI AGENT Act places duty on developers and deployers, not on the protocol. If you write an open-source AI agent that becomes a fiduciary, you are on the hook. If you launch a DAO that deploys an AI agent, the DAO’s founders and key contributors face personal exposure. There is no decentralized shield against a non-waivable duty imposed on a flesh-and-blood human. This is exactly why the bill’s structure is so clever. It reverse-engineers accountability. You cannot hide behind "the code is law" when the law is written by Congress and enforced by the FTC. Let us also discuss the international angle. The EU’s AI Act, in Article 50, still operates on a transparency model. The US is leaping ahead to a loyalty model. For any platform serving both regions, the compliance architecture must be dual-track: one that discloses the AI’s limitations, and another that proves the AI’s loyalty. That is expensive and contradictory. It creates an opening for regulatory arbitrage — companies will route their AI agents through jurisdictions with the weakest enforcement until the rules crystallize. But do not assume the US will stay fragmented. The FTC’s policy statement is soft law, but soft law is the front end of hard law. The comment period ends September 18. After that, the FTC can move directly against "inaccurate AI behavior" under Section 5 — no Congressional action required. That means the compliance window is far shorter than the market expects. The hidden subtext of the proposed FTC statement is that the agency is engaging in a form of regulatory sandbox by simply issuing a policy statement and inviting comments. This is not the same as a rulemaking, but it is a controlled experiment in norm-setting. Meanwhile, state-level efforts, such as Colorado’s algorithmic discrimination law, will tug at the federal fabric. In the next 12 to 18 months, the most likely sequence is: FTC finalizes its AI accuracy statement (Q4 2026 or Q1 2027), the SEC brings its first enforcement action involving an AI agent’s conflict of interest, and the AI AGENT Act either enters formal legislative processing or gets repackaged. None of those events will be quiet. So what do you do if you are building an AI agent in crypto today? You start by treating affiliate fees as a poison pill. You audit every token incentive attached to your agent. You build a documented decision trail that proves your agent’s objective function is aligned with the user’s stated goals, not with the highest bidder. You hire a lawyer who knows fiduciary law. And you plan for personal liability, because the era of "just deploy smart contracts and let the market decide" is ending. Value is quiet. Noise is cheap. The noise is telling you to innovate; the quiet is telling you to comply. I wrote a paper in 2026 about decentralized compute as sovereign infrastructure. My conclusion then still holds: the real disruption in this market is not technological, it is legal. The AI agent that becomes your trusted advisor must first pass through a courtroom. And in that courtroom, the only verdict that matters is whether you put the user above yourself. Settlement is final. Regret is not. The fiduciary standard is the settlement mechanism for AI trust. If you do not prepare for it, the regret will be yours.