Singapore Just Told You Where Agentic Commerce Actually Breaks

The market is treating agentic commerce as an adoption story. It is actually an authorization story, and almost nobody is pricing the difference.

A new consumer survey out of Singapore, part of a seven-country study covering the US, China and the UK, found that 59% of Singaporeans are open to an agent shopping for them while 23% have already used one. That looks like a green light for autonomous retail. It is not. The same survey found that only 15% say they'd let an agent shop, shortlist and pay entirely on its own, with no human check before money moves. That 44-point gap between willingness to browse and willingness to pay is the whole story, and the headline number is the one that will get quoted wrong all week.

What the data actually shows

Zoom out from Singapore and the pattern holds everywhere. A separate global study found a further 41% say they are open to the idea once they understand it better, leaving just 15% who say they would never feel comfortable handing purchase decisions to software. Consumers are not resistant to AI touching commerce. They are resistant to AI touching their money without a checkpoint. The concern is specific: 64% of Singapore consumers surveyed in the Global Payments Agentic Commerce report revealed concern about unauthorised purchases while 57% were worried about incorrect purchases.

Industry research confirms this is an infrastructure gap, not a sentiment gap. Existing payment authorization frameworks, infrastructure readiness and identity verification processes are not yet consistently designed for transactions initiated or completed by autonomous AI agents. That is plumbing, and plumbing problems get solved by whoever controls the pipe.

The mechanism: this is a rails story, not a retail story

Everyone covering agentic commerce is watching the demand side. The transmission into markets runs through the payment networks building the permission layer underneath every one of these transactions.

Visa's answer is the Trusted Agent Protocol, which introduces a Verified Agent ID issued by Visa and a separate consent record signed by the consumer's issuer. Mastercard's answer is Agent Pay, a payments framework that lets verified AI agents transact on a consumer's behalf using Agentic Tokens. Both networks are racing to become the identity layer that tells a merchant an AI agent is who it claims to be and is authorized to spend what it is about to spend.

This is the mechanism. Card networks do not earn from the shopping decision. They earn from the transaction that clears. If agents cannot clear a transaction without a verified identity and a signed consent record, the toll booth for the entire category sits with whoever builds that verification layer, not whoever builds the best shopping assistant. Ant International has now joined Visa and Mastercard because, as Mastercard's own chief digital officer put it, interoperability across Know-Your-Agent frameworks is essential to making agentic commerce work at scale.

The second-order effect nobody is pricing

Here is what almost no one is discussing. The 15% ceiling on full autonomy is not a temporary trust deficit that closes as consumers get comfortable. It is a structural floor created by liability. Someone has to own the loss when an agent buys the wrong item, overspends, or gets spoofed by a merchant pretending to be legitimate. Until that liability question is settled contractually, between card issuer, merchant and agent platform, no amount of consumer comfort raises the ceiling on full autonomy.

That means the real repricing is not in retail multiples. It is in who absorbs fraud and dispute risk in a world where the party clicking checkout is software. Card networks, identity verification vendors and merchant acquirers that solve attribution of intent, proving an agent acted within what a human actually authorized, are building a moat that has nothing to do with whose language model is best. From the delta-neutral volatility book Komey Tetteh runs, this shows up as a slow, structural divergence in how implied volatility gets priced around payments and identity names relative to pure AI application names. The mispricing is not in the AI layer. It sits one level down, in the settlement layer that decides whether the AI layer ever gets to move real money at scale.

The honest counter-case

I could be wrong about the pace. Regulators in Singapore and elsewhere could mandate standardized authorization frameworks faster than card networks can monetize proprietary ones, compressing the toll-booth advantage into a commodity utility. Consumer comfort could also simply outrun my thesis. Separate research found 89 percent of Singapore respondents believe an AI agent could compare more options than they could manually. If trust closes faster than liability frameworks get built, adoption scales before the authorization layer becomes a real differentiator, and the moat I am describing shrinks.

What I am watching next

I am watching whether the Visa, Mastercard and Ant International interoperability push produces one shared standard or three competing walled gardens. A shared standard commoditizes the authorization layer. Three separate ones concentrate the toll. That decision, not adoption curves, determines whether agentic commerce becomes a real revenue line for payments infrastructure or a feature war with margin going to zero.

Where does your own portfolio sit if the transaction layer, not the intelligence layer, turns out to be where agentic commerce actually gets monetized?