Commentary · Technology disputes · AI risk

Who pays when an AI trading agent goes wrong?

By V Kumar Sharma · 20 November 2026 · 8 min read

AI chip on a circuit board
Photo by Igor Omilaev on Unsplash
In short
  • Trades made by an AI agent will usually bind the user who deployed it. The law treats the agent as a tool, not a separate person.
  • Unwinding trades for mistake is hard. Quoine v B2C2 tied knowledge to the programmer, and that test fits non-deterministic AI agents poorly.
  • The real fight is over risk allocation between three parties: the user, the platform the agent trades on, and the vendor that built the agent.
  • Contracts decide most of it. Liability caps and exclusions are tested for reasonableness, and governance frameworks increasingly set the standard of care.
  • Define agent permissions, approval checkpoints and logs before anything goes wrong, because they become the evidence when it does.

The scenario

A fund connects an AI trading agent to its exchange account through API keys. The agent is built by a third-party vendor on a large language model. It reads market data and news, decides what to trade and executes without a human approving each order.

One night it misreads a data feed, or is manipulated by a malicious prompt hidden in content it reads, and executes a series of loss-making trades or transfers. By the morning the losses are significant. Who carries them?

Is the user bound by the agent's trades?

Almost certainly. Section 15 of Singapore's Electronic Transactions Act 2010 provides that a contract formed by an automated message system isn't denied validity just because no person reviewed or intervened in it. An AI agent is not a legal person, so its trades are the user's trades.

Courts elsewhere have refused to let businesses disown their automated tools. In Moffatt v Air Canada, 2024 BCCRT 149, a Canadian tribunal rejected the airline's argument that its website chatbot was effectively a separate entity responsible for its own statements, and held the airline liable for what it said.

Can the trades be unwound for mistake?

Only in narrow circumstances. In Quoine v B2C2 [2020] SGCA(I) 02, trades executed by algorithms at about 250 times the market price stood, and the exchange was liable for reversing them. For unilateral mistake, the majority looked to the knowledge of the programmer, from the time of programming up to when the contract was formed. In Chwee Kin Keong v Digilandmall.com [2005] SGCA 2, by contrast, buyers who knew a website's price was an obvious error couldn't enforce their orders.

That test was framed around deterministic algorithms, whose outputs trace back to choices a programmer made. An agent built on a large language model doesn't work that way: the same inputs can produce different actions, and no programmer decided that this trade should happen. Expect arguments that the right question is what a reasonable trader would have understood, which is the approach Lord Mance took in his dissent in Quoine. Until a court decides, anyone relying on mistake to escape an AI agent's trades should assume an uphill battle.

The three relationships that decide who pays

1. User and platform

The exchange's terms govern. Most say that executed orders are final, that users are responsible for activity on their API keys, and that the platform's liability is limited. After Quoine, platforms are also bound by their own irreversibility wording unless they reserve a clear, properly notified right to unwind.

The platform's exposure grows where its own systems contributed: a malfunction, a security failure, a data breach that let an attacker in, or a failure to act on red flags it was obliged to monitor. Exclusion clauses are then tested for reasonableness under the Unfair Contract Terms Act, which I discuss in "You clicked I accept".

2. User and AI vendor

This is where most of the loss will be argued. The user's claim is usually in contract: did the agent perform as described, and did the vendor give warranties about accuracy, guardrails or security? Expect the vendor's terms to disclaim fitness for trading, cap liability at the fees paid and exclude indirect and consequential loss. Those clauses are also subject to the reasonableness test.

A negligence claim is possible, and Singapore is more open than England to claims for pure economic loss under the two-stage test in Spandeck Engineering v Defence Science & Technology Agency [2007] SGCA 37. But where the parties have a contract that allocates risk, the courts are slow to let tort override that bargain.

The questions that usually decide these claims are practical. Did the vendor follow its own documented safeguards? Was the failure a model behaving within its stated limits, or a defect such as a missing permission control or an exploitable prompt injection route? Did the user configure the agent within the vendor's recommended limits?

3. Platform and vendor

If the agent's failure harmed the platform or other users, for example by flooding the order book or draining linked accounts, the platform will look to the vendor under its API terms. As with the Bitget supply-chain exploit, the real target may be the vendor's errors and omissions or cyber insurance rather than its balance sheet.

Governance frameworks are setting the standard of care

No Singapore statute yet allocates liability for AI agents. But published frameworks are shaping what counts as reasonable conduct, and they will be cited in any dispute.

  • IMDA's Model AI Governance Framework for Agentic AI, launched on 22 January 2026 and updated on 20 May 2026, recommends limiting an agent's access to tools and systems, keeping its actions traceable, and requiring human approval at significant checkpoints, such as high-stakes or irreversible actions. It also stresses that humans remain accountable.
  • MAS's AI risk management guidelines for financial institutions expect AI used in trading and risk to be inventoried, tested and overseen by the board, including AI sourced from third parties. See my earlier piece.

A party that ignored these frameworks will find it harder to argue that it acted reasonably, or that its exclusion clause is fair.

The European angle

The EU has withdrawn its proposed AI Liability Directive. Its new Product Liability Directive (EU) 2024/2853 applies to products placed on the market after 9 December 2026 and treats software and AI as products. But its no-fault compensation is aimed at harm to individuals, such as personal injury, property damage and data loss, rather than a fund's trading losses. For commercial losses, contract and negligence will still do most of the work.

A practical checklist

Users and funds deploying agents

  • Restrict API keys: trading only, no withdrawals, with position and loss limits enforced on the exchange side.
  • Require human approval for large or unusual trades, and keep a kill switch someone can reach at 3am.
  • Keep complete logs of the agent's inputs, reasoning and actions. They are your evidence.
  • Read the vendor's liability cap against the capital the agent controls, and negotiate or insure the gap.

AI vendors

  • Document the agent's intended use, limits and safeguards, and make sure the product matches the documentation.
  • Build in permission controls and defences against prompt injection, and test them.
  • Check that your liability caps would survive a reasonableness challenge, and that your insurance responds to AI-caused financial loss.

Platforms

  • Reserve a clear, properly notified right to suspend or unwind trades after malfunctions, and say when it applies.
  • Monitor for abnormal agent behaviour, and decide in advance what you will do when you see it.
  • Make sure your API terms allocate responsibility for third-party agents connected to user accounts.

Frequently asked questions

Am I bound by trades my AI agent makes?
Generally yes. Automated contracts are valid under the Electronic Transactions Act, and mistake is a narrow exception.
Can I sue the AI vendor?
Possibly, in contract or negligence, but expect liability caps and exclusions, which are tested for reasonableness.
Is the platform liable for my agent's trades?
Usually not, unless the platform's own systems, security or conduct contributed to the loss.

Deploying AI agents, or dealing with the fallout from one? Get in touch, or read the first 72 hours after an incident.

This article is general information and is not legal advice. The scenario is hypothetical.

V Kumar Sharma

V Kumar Sharma is a Partner at Eldan Law LLP and a technology disputes lawyer acting on cyber incident response, fraud and asset recovery, data protection and international arbitration. ALB Asia 40 Under 40 (2026) & Rising Star (2025). Co-Chair, Blockchain Association Singapore Web3 Cybersecurity Sub-Committee.
Get in Touch