Who Pays When Autonomous Finance Goes Wrong?

AI Smart Contracts Need Accountability to Scale

An AI agent moves money after interpreting market data. A smart contract executes the transaction exactly as instructed. The technology works, yet the customer loses funds because the instruction was wrong. Who pays?

That question could help determine whether autonomous finance becomes dependable infrastructure or remains an experiment customers approach cautiously. Faster execution creates value, but institutional adoption also requires clear authority, effective safeguards, and a credible process for handling losses.

A conceptual study by Kwok Tung Cheung and Sridhar Ramamoorti, published in AI and Ethics, examines accountability when agentic AI interacts with smart contracts.¹ Its central argument offers a constructive direction: organizations can establish accountability before deployment, even when identifying individual blame after a failure proves difficult.

Why Agentic AI Changes Smart Contract Risk

A conventional smart contract executes programmed conditions. An AI agent adds interpretation: it can evaluate information, choose tools, and initiate transactions toward a goal. In practice, the model may operate outside the blockchain while calling contracts that execute its decisions.

This creates two distinct questions. Did the contract execute correctly? And was the agent authorized and justified in requesting that action? A technically valid transaction can still represent an operational failure.

The paper illustrates this through a hypothetical decentralized lending scenario. An unexpected interaction with an oracle, a service supplying external price data, causes an AI monitor to trigger an inappropriate collateral liquidation. The resulting loss crosses several boundaries: data provision, model behavior, protocol configuration, and execution.

For readers following AI agents that move money, this is the next governance challenge. Permission to transact establishes what software may do. Accountability determines which participants must answer when that permission produces harm.

What the AI Accountability Study Proposes

The authors distinguish attributability, answerability, and accountability. Broadly, these concern ownership of an action, explaining its reasons, and being subject to an institutional process with consequences.

Their argument is that accountability can survive uncertainty about the other two. A business can accept defined obligations for deploying a system without proving that a particular employee anticipated every outcome. Those obligations must be established fairly, rather than assigned arbitrarily after losses occur.

The following conditions summarize the paper’s proposed basis for legitimate accountability without proof of individual fault.

The distinction matters for retail users. Accepting terms does not give a borrower control over model design or oracle selection. The authors consequently place substantial obligations upstream. This is a proposed framework, not a statement that its liability allocations already apply across jurisdictions.

Human Control Requires Time and Authority

The paper advocates a human-in-the-lead approach: people establish objectives, acceptable behavior, intervention rights, and deployment accountability. A reviewer watching activity is insufficient if transactions complete before that person can respond.

Its illustrative lending safeguards include a $250,000 liquidation review threshold, a price feed averaged over at least 15 minutes, and a challenge window of at least 60 minutes for qualifying liquidations. These are suggested design parameters, not measured safety benchmarks.

That qualification exposes an important commercial tradeoff. Delaying liquidation could protect a borrower from erroneous data, while increasing a lender’s exposure if collateral genuinely collapses. Different assets and markets require different controls. A universal waiting period would be difficult to justify.

A practical implementation would therefore combine human decisions about policy with automated enforcement. Small, permitted actions could proceed immediately; unusual actions could pause; emergencies could trigger predefined restrictions.

• Define which assets, destinations, and amounts an agent can access.
• Monitor cumulative exposure across sequences of transactions.
• Assign intervention authority and test whether it works under pressure.

Accountability Is Part of Financial Infrastructure

The broader commercial implication is that autonomous finance has to manage both transaction costs and failure costs. Cheaper settlement may offer limited savings if customers must absorb expensive disputes, recovery efforts, or unexpected losses.

Consider an illustrative agent permitted to make payments below $100. Hundreds of individually compliant payments could still exceed the customer’s intended budget. The relevant control is therefore cumulative spending across a task or period, rather than transaction size alone.

Controls should also sit outside the model’s discretionary authority. Asking an agent to respect a budget differs from requiring independent software to reject unauthorized spending. This separates the agent’s ability to propose an action from the system’s authority to execute it.

As machine-to-machine payments expand the range of services software can purchase, that separation becomes more valuable. An agent buying data and then trading on it creates a chain of linked decisions. Operators need to understand the complete workflow.

Compensation Needs Funding Behind It

A further implication goes beyond the paper’s role allocation: an accountable entity needs resources to honor its obligations. A named operator with insufficient capital offers limited protection when several customers suffer losses simultaneously.

Reserves, insurance, or contractual guarantees could support a credible recovery arrangement, depending on the activity. Each introduces costs and limitations. Insurance exclusions, for example, could leave a gap precisely where unfamiliar agent behavior creates a loss.

For investors, this suggests evaluating the economics after safeguards and recovery obligations. A platform’s transaction growth tells only part of the story. Sustainable margins depend on what it costs to operate safely and which losses remain on its balance sheet.

Coinbase Connects AI Agents With Financial Execution

Coinbase offers public-market exposure to companies building these capabilities. Its Agentic Wallets launch describes programmable session caps, individual transaction limits, private-key isolation, and transaction screening. These features illustrate how financial access can be bounded at the infrastructure level.

In its September 22, 2026 announcement, Coinbase for Agents added equities and x402 payments, connecting agent-driven trading with payments for data and model services. Coinbase described configurable asset access, spending limits, and per-action approvals.

The same announcement states that users direct agent activity and assume transaction risk. That creates a revealing contrast with the study’s preference for allocating model and protocol risks to participants positioned to influence them. Product guardrails and contractual accountability address different parts of the problem.

Coinbase’s tools do not establish that it implements the paper’s framework. The investment connection is broader: agent adoption could increase demand for wallet, payment, and execution services. Whether that demand produces material earnings depends on usage, monetization, competition, and operating costs. Coinbase also remains exposed to the wider crypto market.

Latest Coinbase (COIN) News and Developments

What Investors Should Watch in Autonomous Finance

The study provides a governance argument, without experimentally validating its proposed safeguards. Enforcement against pseudonymous organizations remains difficult, and its hypothetical liquidation scenario cannot establish how controls perform during a live market crisis.

Nevertheless, it supplies a useful direction for evaluating progress. Investors can look for explicit authority boundaries, tested intervention mechanisms, transparent loss allocation, and evidence that customers value those protections.

The strongest businesses may be those that make autonomy dependable enough to purchase repeatedly. Clear accountability could help turn faster financial execution into a service institutions and individuals are willing to trust.

References:

¹ Cheung, K. T., & Ramamoorti, S. (2026). Accountability in smart contracts and agentic AI: How to address the responsibility gap and retain meaningful human control? AI and Ethics, 6, Article 585. https://doi.org/10.1007/s43681-026-01434-3