As a global hub for technology and innovation, Singapore sits at the heart of AI development. Amidst a growing demand for convenience, agentic payments are the latest evolution in online commerce. Agentic payments are payments that are initiated, authorised and rendered by autonomous AI agents within boundaries set by customers. Singaporeans have expressed their openness to agentic payments, where 44% of Singaporean consumers stated that they would allow an AI agent to shop for them.
The IMF splits agentic payments into 3 layers — intent, policy and settlement.
- Intent: First, customers provide instructions of purchase and establish rules on how money can be spent, including spending limits and payment preferences. AI agents then interpret these instructions, browse to evaluate online options and propose a transaction.
- Policy: Second, the proposed transaction is then evaluated against the rules the customer set. If the evaluation fails at this stage, the transaction never goes through.
- Settlement: If the transaction passes evaluation against the customer’s defined rules , the payment network authorises the transaction and payment is executed.
The shift is already underway. In partnership with local banks, Visa launched its Agentic Ready Programme in Singapore and Mastercard ran a pilot programme with HSBC testing agentic payments. While these developments could improve the customer experience, their novelty brings with them a plethora of legal issues for Payment Service Providers (PSPs) to consider.
A) Liability
If an AI agent authorises a transaction based on its flawed interpretation of a customer’s defined instructions, who should bear liability? Traditionally, liability is attributed to actors who have taken actions that put them at fault. However, the autonomous nature of AI agents complicates the process of allocating responsibility between the AI developer, customer or bank that deployed the system, or if at all. Pending greater regulatory clarity, payment service providers should consider sharing contractual risk between themselves and AI developers. Contracts should outline which party bears responsibility for different issues such as programming errors and system failures. However, the IMDA raises a concern that liability could then disproportionately be shifted to consumers. As businesses typically have more bargaining power than consumers, they are better equipped to allocate risk favourably through disclaimers and other terms of use. This could erode consumer trust in agentic payment services, increasing pressure for regulatory intervention.
B) Contract formation
Agentic payments raise an imperative question on when consumers are legally bound by terms accepted by an AI agent on their behalf. While merchants will argue that an AI agent acting with the consumer’s authorisation amounts to a consumer’s valid acceptance of the contract (and all its associated terms), consumers may contend that authorising an AI to purchase a product does not extend to agreeing to ancillary contractual terms, such as liability disclaimers and refund limits.
Although there have been no case developments surrounding agentic payments yet, Quoine Pte Ltd v B2C2 Ltd [2020] offers useful guidance on how Singapore courts approach automatically generated contracts on electronic platforms. A key consideration is the knowledge and intention of the programmer at the time the code was written. As the programmer did not have knowledge or intention of the mistake at the time of programming, the Court of Appeal upheld that the doctrine of unilateral mistake did not apply and that the contract was still valid.
However, the law is still unsettled on whether this reasoning extends to non-deterministic AI. Unlike the deterministic algorithm in Quoine, which merely executes clearly defined instructions, agentic AI operates with more autonomy as it can interpret a consumer’s instructions in multiple ways. If an AI agent makes a choice its programmer did not foresee, such as booking a flight with a no-refund policy because it was the cheapest option, this could arguably render the programmer’s intention or state of mind an irrelevant factor.
C) Consumer protection implications
Where the AI agent has autonomy to decide which payment methods to use, its chosen method has significant consumer protection implications. If the agent selects a credit card, the consumer benefits from chargeback mechanisms to allow for a reversal of funds. If it opts for PayNow or bank transfers instead, consumers face more limited avenues for recourse. The risk becomes even more apparent if the agent opts for methods not protected by the Consumer Protection (Fair Trading) Act such as cryptocurrency.
In order to mitigate dispute claims, PSPs should:
- Require customers to set their payment method preferences in advance.
- Obtain the consumer’s express consent before selecting a payment method that materially affects their consumer protection rights.
- If left to the agent to decide, PSPs should maintain logs detailing why a specific payment method was chosen and how this decision complied with the consumer’s instructions.
As banks continue to explore agentic payments, we can expect dynamic developments in the area of online commerce. While Singapore does not yet have a standalone law on AI regulation, PSPs should ensure that its AI systems comply with relevant frameworks such as the Personal Data Protection Act and AI Verify. Ultimately, the success of agentic payments will depend not only on its innovative strides but the ability of PSPs to maintain consumer trust through robust safeguards.
