There is something genuinely interesting about the idea of a personal robot shopper that can understand what you want and eventually press pay on your behalf.
Still, every time the industry talks about making payments more automated, the conversation somehow finds its way back to the same word. Trust.
We saw a version of this when AI first entered banks and financial institutions, when the early excitement around what the technology could do quickly ran into a more practical question.
How much control are people actually willing to hand over?
Agentic payments seem to be moving through a similar cycle, only this time the stakes feel more personal.
Asking AI to recommend a hotel for your next trip feels harmless enough. Letting an AI shopping agent spend your money is another matter.
Adeline Kim, the newly appointed Group Country Manager, Regional Southeast Asia & SVP, Global Clients & Acquirers, Asia Pacific at Visa, sees that distinction as central to the next stage of AI-driven commerce.
Singapore consumers are clearly not strangers to AI either, with a Visa-commissioned study finding that close to 77% of residents use generative AI tools and 8 in 10 rely on AI assistance when shopping online.
People are already bringing AI into the shopping journey, but I have to really ask, do you trust someone else, or in this case, something else, with your money?
And Adeline seems to agree with my curiosity. She said:
“There is a big difference between using AI to search for a product and allowing an AI agent to act on your behalf when money is involved.”
The Agent Should Not Have a Blank Cheque
Once an AI agent gets close to the wallet, the concern becomes fairly straightforward.
How much freedom should it actually have?
Adeline Kim
“An AI agent should not have open-ended authority to spend,” Adeline answered directly, meaning that consumer consent, in other words, cannot be vague.
The agent needs to have clear boundaries, including how much it can spend and when it needs to check back with the person behind the purchase.
She believes that agentic payments work best when the AI agent has a narrow, well-defined job.
It may be able to handle a routine purchase within a set budget, but anything unusual should trigger a pause, an extra check or a request for confirmation.
“Agents can act at speed, but consumers and issuers should remain in control at key decision points,” Adeline pointed out, adding that faster checkout only works when the agent still knows when to stop and ask.
What If the Agent Recommends What It Was Paid to Recommend?
Agentic payments also raise a different kind of trust issue, and honestly, it is one I did not think about at first.
When an AI agent recommends a product, why did it pick that one?
Did it match what the consumer wanted? Was it genuinely the better option? Or did a commercial arrangement play a role?
Adeline said transparency will matter here because users have the right to know whether a recommendation reflects their preferences or someone else’s incentive.
If money or a partnership influenced the recommendation, the consumer should be able to see that in plain language.
Most of us can live with ads when they look like ads. A recommendation feels different when it comes dressed up as neutral advice, especially if money has shaped the suggestion behind the scenes.
“Once people feel an agent is not acting in their interests, adoption will slow very quickly,” Adeline cautioned.
Mistakes Will Happen, So Resolution Has to Be Clear
Even with controls in place, AI agents may still get things wrong.
They may misunderstand a consumer’s instruction, buy the wrong item or repeat a purchase. They may also choose a merchant that looked right to the system but was not what the consumer intended.
Lionel Grosclaude, CEO of Fime, put this problem in simple terms in a separate interview. An agent could be asked to buy blue shoes and return with a red pair instead. Annoying, yes, but not necessarily serious.
The stakes change when the purchase carries health or safety consequences. A missed allergy warning, for example, may still leave behind a transaction that looks valid in the system.
The harm, however, sits with the person who trusted the agent.
What If AI Makes a Bad Purchase For You? AI can already shop and pay on your behalf. But what happens when it buys something you never wanted, or worse, something that puts you at risk? fintech AI payments
To answer that worry, Adeline said the industry should be realistic about mistakes.
“As commerce becomes more autonomous, it is realistic to expect that AI agents may sometimes make mistakes, just as people do today,” she said. “What matters is that the consumer is not left on their own when that happens.”
Visa sees accountability and traceability as central to agentic payments because a valid-looking transaction may still need to be questioned if the agent acted outside the consumer’s intent.
Trusted agent-initiated transactions should still sit within established payment protections, including dispute and chargeback frameworks, so consumers have a clear route to resolution when something goes wrong.
Traceability matters because the transaction journey becomes more complex once an AI agent is involved. The industry needs to know whether the agent followed what the consumer actually asked for.
A wrong purchase cannot simply be brushed aside with “the AI did it.” Someone still has to help the consumer make it right.
Fraudsters May Start Fooling the Agent, Not Just the Shopper
Scammers will not ignore agentic payments, and they will likely adapt quickly.
Many scams today try to fool people through convincing-looking storefronts or checkout flows. Agentic commerce could shift part of that target to the AI shopping agent acting on behalf of the consumer.
Fraudsters may start designing storefronts or checkout flows that look trustworthy to a machine, even if a human shopper might question them.
Adeline warned that fraudsters will always look for the weakest point in any new technology. Agentic commerce, therefore, needs trust on both sides of the transaction.
Merchants need to know whether an AI agent is legitimate and acting on behalf of a real, authenticated consumer. AI agents also need to recognise which merchants and checkout flows they can trust.
Fraud prevention still needs layers, but the people affected by it should not feel shut out of the process. Adeline pointed to tokenisation, authentication checks, behavioural intelligence and real-time risk scoring as part of the broader defence.
The scale of the risk also changes. One shopper may make one bad payment, but a compromised AI shopping agent could trigger multiple transactions quickly if the controls are weak.
Adeline said Visa has used AI to protect the payments ecosystem for more than 30 years, analysing over 200 billion transactions a year and 500 data elements in every transaction to help stop fraud in real time.
Real-time detection becomes even more important once agents can act quickly. Unusual purchase volume, unfamiliar spending patterns or abnormal agent behaviour should prompt additional checks before the problem spreads.
No single company can secure agentic payments alone because the bank, merchant, payment network and AI platform each see different parts of the transaction journey.
Trust, in this case, has to be a team effort.
Personalisation Is Useful, Until It Starts Feeling Creepy
Agentic commerce also raises a privacy question that feels very personal very quickly.
An AI agent can only be useful if it understands enough about the consumer, such as their preferences and usual spending habits. But we cannot ask for personalisation and then pretend there is no privacy trade-off.
Adeline emphasised that agentic commerce will only work if people trust how their data is being used. The more personal the experience becomes, the clearer the limits around access and permission need to be.
The industry also needs to avoid a free-for-all where every party in the transaction gets more information than it needs. The agent may need context, but not everyone needs the full picture.
Adeline captured the balance well.
“The promise of agentic commerce is personalisation, but with boundaries,” she said. “Consumers should feel that AI understands their preferences, not that they are being watched.”
A useful assistant should feel helpful. Once it starts feeling like surveillance with a checkout button, trust becomes much harder to earn.
Not Everyone Cares If AI Goes That Far
All the excitement around agentic payments can make the industry forget who it is ultimately trying to serve.
Yes, some consumers will find it useful if an AI agent can handle a routine purchase without much fuss. But older users or those less comfortable with digital tools may not care for that level of automation, especially when money is involved.
Adeline stressed that innovation only matters if people can use it confidently and safely, which means the experience has to be clear enough for consumers to understand what they are authorising and how to change or revoke permission when they need to.
“The aim should not be to push everyone into automation,” she said. “It should be to give people more confidence, more control and more choice.”
Agentic payments could still become part of everyday commerce, but only if the industry avoids treating every consumer as equally ready for AI to act on their behalf.
Giving people “more confidence, more control and more choice” may sound simple, but it is probably the part that will decide how far agentic payments can really go.
After all, there is no AI when you spell out trust, but there is “us”, the humans who still need to stay in control when AI gets closer to our money.
Featured image: Edited by Fintech News Singapore based on an image by topntp26 via Magnific.
