Researchers from Microsoft Research, Google DeepMind, Columbia University, Virtuals Protocol, and AI startup T54 Labs published an open-source financial settlement framework designed to protect users when AI agents fail at financial tasks. The protocol, called the Agentic Risk Standard (ARS), was detailed in a 30-page paper on arXiv on April 5 and covered by Fortune on April 8.

The Core Problem

The researchers identify what they call a “guarantee gap”: AI safety research can reduce the probability of agent failure, but LLMs are inherently stochastic. No amount of fine-tuning eliminates the chance of hallucination. When an agent sits on top of a brokerage account or executes financial API calls, a single failure produces immediate, realized loss.

“Most trustworthy AI research aims to reduce the probability of failure,” Wenyue Hua, senior researcher at Microsoft Research, told Fortune. “That work is essential, but probability is not a guarantee. ARS takes a complementary approach: instead of trying to make the model perfect, we formalize what happens financially when it isn’t.”

How ARS Works

The framework borrows directly from centuries of financial engineering. It introduces three layers:

  • Escrow vaults that hold service fees and release them only upon verified task delivery.
  • Collateral requirements that AI service providers must post before accessing user funds.
  • Optional underwriting where a risk-bearing third party prices the danger of an AI failure, charges a premium, and reimburses the user if the agent fails.

For standard service tasks (generating a report, writing a slide deck), escrow-based settlement is sufficient. For tasks involving fund transfers, currency trading, or leveraged positions, the agent accesses user capital before outcomes can be verified. That is where underwriting becomes essential. The paper maps this against existing precedents: construction uses performance bonds, e-commerce uses platform escrow, derivatives markets use clearinghouses, and DeFi uses smart contract collateralization. AI agents, they argue, are the next category that needs equivalent infrastructure.

Regulatory Context

Financial regulators are already moving. FINRA’s 2026 regulatory oversight report, released in December, included a first-ever section on generative AI. It warned broker-dealers to develop procedures targeting hallucinations and to scrutinize AI agents that may act “beyond the user’s actual or intended scope and authority.”

ARS is pitched as what regulators have not yet built: not a set of rules, but a protocol. A standardized state machine governing how funds are locked, how claims are filed, and how reimbursements trigger when an agent fails. The implementation is open-source on GitHub.

The Deployment Bottleneck

T54 founder Chandler Fang told Fortune that the financial ecosystem “currently has no way to operate other than to defer all liability back to a human.” The team acknowledges ARS is one layer of a larger trust stack. The real bottleneck: building accurate risk-pricing models for agentic behavior. Nobody yet knows how to underwrite the failure rate of an autonomous agent executing a leveraged currency trade.

For fintech teams and enterprise finance operations evaluating agent deployment: the research community is building the safety infrastructure. The question is whether adoption outpaces it.