Proof of intent: how payments are preparing for agents

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 13 October American Express introduced an "Intent Intelligence Framework" for agent-initiated purchases, TechCrunch reports. A Create Intent API captures the customer's instructions and constraints, such as merchant, maximum budget and product preferences. The customer authenticates and signs off; Amex then creates an Intent ID and a Proof of Intent Token, which is used to obtain a digital payment credential the agent uses at checkout instead of the actual card number. Amex says the intent record can also support fraud prevention and dispute resolution, and that it has not yet seen an increase in disputes from agentic commerce.

Amex's Hannah Bozian gave an example: if a customer asks for white Nike sneakers in a size 9 and the agent tries to buy shoes from Adidas, the transaction can be stopped because it does not match the authorised intent.

In India, India Today, citing Reuters, reported on 1 September that NPCI was expected to develop a Unified Agent Protocol allowing AI agents to make low-value UPI payments within rules and spending limits set in advance, building on UPI Circle and Reserve Pay, with identity checks, audit trails and a liability framework whose details were not yet available.

Why it matters

My essay on agent commerce argued that if agents can collaborate they can trade, and that financial services would need to rebuild its trust plumbing (mandates, limits, audit trails, dispute resolution) for machine counterparties. These designs are that rebuild, and they converge on the same principle: capture the human's intent as a structured, signed object before money moves, and check the transaction against it.

Capture the human's intent as a signed object before money moves. Everything else in agent payments follows from that.

That is a measurement design as much as a security one. A mandate with a merchant, a budget and a category is a set of counter-metrics for the agent: the agent can optimise for price, but it cannot leave the box the customer drew.

The parallel to existing banking practice is close. A standing instruction or an e-mandate already records who authorised what, up to what amount, for how long, and how it can be revoked. Agent payments need the same object, with two additions: which agent is acting, and what the human actually asked for in their own words. The second is new, and it is what lets a dispute be resolved on evidence rather than on memory.

For risk teams the useful shift is from detecting bad transactions to checking transactions against declared intent. Fraud models look for anomalies; an intent check asks a simpler question: is this what the customer said they wanted? Both will be needed. The intent check is cheaper, explainable, and works on the first transaction, before any behavioural history exists.

India's scale makes the design choices consequential. A protocol that sits on a network processing tens of billions of payments a month will set the default for how delegation, limits and liability work for a very large number of people. It is worth getting the limits and the revocation path right before the convenience.

What I'd watch

• Liability when intent is ambiguous. 'Buy paper towels' is easy; 'book me something nice for the weekend' is not.
• Whether intent tokens become interoperable across networks or fragment by provider.
• How banks reflect these patterns in their own agent policies: delegation limits per agent, revocation, and a human approval threshold.

Sources

1. Amex introduces a framework to prevent AI agents from making unauthorized purchases. TechCrunch, 13 Oct 2026.
2. Your AI agent could soon pay through UPI. Here's how it may work. India Today (citing Reuters), 1 Sep 2026.