Yesterday I wrote about AI entering the formation of a decision.
Now the next boundary is visible:
the agent is entering the payment rail.
Reuters reports that India is preparing a Unified Agent Protocol for UPI that could allow AI agents to make small digital payments without asking the human to approve every transaction.
This is not a small pilot environment.
UPI processed 24.51 billion transactions worth 29.82 trillion rupees, about $314 billion, in August alone.
The consumer story is obvious:
groceries, discounts, routine purchases, perhaps later investments under specified price rules.
But the architectural story is more important.
The question is no longer only:
Can the agent pay?
It becomes:
Who delegated that authority? For which purpose? Within which budget? For how long? Can it be revoked immediately? How is the agent identified? What evidence remains? Who carries liability when it is wrong?
According to Reuters, the planned system is expected to build on existing UPI delegation and blocked-funds mechanisms, with rule-based instructions, spending limits, identity checks and audit trails.
NPCI has not yet published the protocol, and the liability framework is still unclear.
So this is not a finished safety proof.
It is a strong infrastructure signal.
When the human click disappears, the function of that click must reappear somewhere else:
in the mandate, the limit, the time window, the revocation path, and the witness trail.
I know this pattern from construction.
A worker may receive a company card to buy materials. That does not authorize him to redesign the building, change the supplier chain, or empty the account. The card is useful precisely because the mandate is narrow, visible and revocable.
In my c = a + b architecture, the agent remains a worker.
It may receive a bounded mandate.
It does not become the source of will or authority.
An agent may recommend.
An agent may execute.
An agent must not silently enlarge its own role.
This is why access is not authority, and capability is not permission.
The future of agentic commerce will not be decided by which model buys fastest.
It will be decided by whether delegated power remains bounded when the human is no longer approving every click.
Remove the click if it saves time.
Do not remove the chain of responsibility.