The Race to Build a Bank Account for AI
AI agents need more than digital wallets to pay. Banks, card networks and fintech startups are competing to provide identity checks, spending limits and accountability.
Picture a finance team in Bangkok that lets an AI assistant handle staff travel. An employee types one line: "Book me a roundtrip flight to Singapore from Tuesday to Thursday." The AI picks a flight that fits the company's travel policy, the employee approves the selection, and the ticket is paid. The airline is paid the usual card way and sees nothing unusual, in this illustrative workflow.
The components of this trip are being offered by card companies, payment firms and a wave of startups, although their seamless integration should not be assumed. Seen through an industry analyst's eyes, the pieces fall into simplified three bands. Established platforms have strong positions at the front and in payment rails. The middle, where identity, permissions and spending controls meet, is an especially active competitive zone.
Big Platforms Have a Head Start Where People Ask
Take booking a trip as an illustrative example. The trip starts inside an AI assistant (row 1 in the diagram above). It could be a general assistant such as Meta's Muse, which Stripe says can shop with a customer's saved card, or a business tool such as Amex GBT's travel connector, which supports AI-assisted travel booking within company policy. Established platforms have a distribution advantage, though the interface is not closed to new entrants.
Established Players Run the Payment Rails
When the ticket is paid, familiar names do the work: Visa, Mastercard, Stripe and Coinbase (row 3 in the diagram). Visa and Mastercard are developing agent-specific payment credentials. Behind the scenes, card companies and payment firms are starting to settle some transactions using stablecoins, digital tokens designed to track the value of a currency and transferable outside traditional banking hours. In some arrangements, the merchant still receives conventional card settlement while stablecoins are used elsewhere in the payment chain.
Startups Are Racing to Control How AI Spends
Between the request and the payment sits the hardest question. The money belongs to the company. But how does the AI get permission to spend it? Who sets the limits, and what stops the AI from exceeding them? How does the airline know the payment is authorized?
Startups are building the infrastructure to answer these questions, often combining identity, spending controls, and payments in one product.
Some start from the card. Reap in Hong Kong and Rain in the US offer agent-focused card and spending controls, though their funding arrangements and product features differ. For our Singapore trip, Reap can create a card that works only for that one flight, with merchant, amount and time restrictions, according to Reap's product description.
Others start from the account. Catena Labs and Natural are building account, wallet and payment infrastructure for agents, with varying degrees of owner control. This generally means delegated access to funds, not an independent legal bank account owned by the AI. In our example, some spending restrictions could live there, alongside the employer's existing travel policy.
A third group helps companies that build AI assistants. Skyfire develops agent identity and payment credentials intended to help a service verify the agent and its authority. Crossmint and Sapiom offer infrastructure intended to simplify how developers add wallets, payments and spending controls.
Kite AI approaches the problem from the underlying infrastructure. Its Agent Passport is designed to give agents verifiable identities, delegated spending limits and payment records, while its network supports settlement. I previously interviewed Henry Lee, Kite AI’s Head of Blockchain Product & Infrastructure, for SCB 10X’s AI-VOLUTION series. Our conversation highlighted why an AI payment needs more than a wallet: the system must also establish who authorized the agent and what it is allowed to spend.
Who Pays When the AI Gets It Wrong?
Now suppose the AI books the wrong week. A wrong-date booking is not necessarily an unauthorized card transaction. Responsibility may depend on the travel supplier's terms, the company's approval workflow, and agreements with its AI and payment providers. American Express has announced Agent Purchase Protection for eligible registered-agent purchases, but its coverage should not be assumed to extend to every agent mistake. Startups such as t54 are building identity and risk-scoring tools. Its founder has also discussed accountability when agent payments go wrong, but this is not the same as a published contractual guarantee.
The Middle Will Go to Whoever Earns Trust
Look again at the middle band. Holding money, setting limits, issuing cards, checking who a customer is and absorbing mistakes are all things banks already do. The startups are moving faster, and that is why they are starting to look like small banks for software. The wrong-week ticket is a useful test of the business model: who can prevent the error, prove what the agent was authorized to do, resolve a dispute and, where appropriate, price the risk? The winner may be the provider that combines distribution, seamless integration, and trusted financial controls into infrastructure that businesses and individuals come to depend on.
Disclaimer: The views expressed in this article are my own and do not necessarily reflect those of my employer or any organization I represent.
Oravee Smithiphol is Tech Intelligence & Insights Manager at SCB 10X, where she tracks emerging AI developments across startups, enterprises, and financial services organizations worldwide.
