What Financial Infrastructure Will AI Agents Need to Transact Autonomously?

AI agents can already search for flights, compare products, generate content and interact with digital services. The limitation is that most agents still stop before the transaction. They can recommend what to buy, but they often cannot complete the purchase on a user's behalf.

A financial system built for agents would extend that capability from recommendation to execution. It would need to handle not only payments, but also the identity of the agent, the authority granted by the user, and the controls governing how that authority can be used.

This article explains why existing payment infrastructure is not fully suited to autonomous agents, how Kite AI approaches identity, permissions and payments, and what agentic transactions could mean for digital commerce.

What is an agentic financial system?

An agentic financial system enables AI agents to make transactions as part of completing a task.

Consider travel. An agent can already search for flights, compare prices and recommend an itinerary. But the user will often still need to visit the booking site, enter personal information and complete the payment.

With agentic payment infrastructure, the same agent could potentially carry the process through to completion: selecting the service, initiating the transaction and paying within the authority given by the user.

Henry Lee, Head of Blockchain Product & Infrastructure at Kite AI, describes this as part of a broader transition toward an “agent-first internet,” where users may rely on a smaller number of personal agents to interact with applications and services on their behalf.

Why can’t AI agents simply use existing payment systems?

They can, but existing systems were not designed around autonomous software.

A conventional payment carries information and trust relationships beyond the transfer of money itself. Banks and payment providers help establish that the person making a transaction is legitimate, while merchants rely on those systems when deciding whether to accept a payment.

An AI agent creates a different situation. It does not naturally have the same identity attributes as a human cardholder, yet it may be attempting to transact on behalf of one.

This creates a problem for merchants. Automated activity is frequently associated with fraud, bots and unauthorized purchases, so a merchant needs a reliable way to distinguish a legitimate agent acting with permission from malicious automation.

For users, there is a parallel issue: giving an agent the ability to pay should not automatically mean giving it unrestricted access to funds.

What does Kite AI do?

Kite AI is building infrastructure for AI agents, including payments, settlement, controls and governance. The company is a Series A business based in San Francisco and is backed by investors including Coinbase Ventures, PayPal Ventures and General Catalyst.

Its approach brings together three components that become closely connected when an agent starts transacting: identity, permissions and payments.

Identity helps establish the agent and the party it represents. Permissions determine what the agent is authorized to do. Payments allow it to complete the transaction once those conditions are satisfied.

The combination matters because moving money is only one part of an agentic transaction. A merchant also needs confidence that the agent is legitimate, while the user needs control over the scope of its authority.

How does Kite Passport work?

Kite Passport is designed to connect these capabilities to AI agents through a relatively simple user experience.

According to Lee, a user can install the Kite Passport skill in an AI assistant such as ChatGPT or Claude, sign in through Kite, and add funds using supported payment methods. The connected agent can then use the available balance when accessing compatible services.

The design reflects Kite AI focus on distribution as well as infrastructure. Rather than requiring users to adopt an entirely separate interface, Kite AI is looking to make its capabilities available through AI assistants people already use.

The company is already using this workflow internally. Kite's San Francisco team uses an agent connected to Slack to order pantry supplies through Amazon. A team member can request items such as milk or snacks, and the agent completes the purchase using the available wallet balance.

What can AI agents pay for today?

Kite Passport supports more than 80 services according to the interview, covering use cases beyond e-commerce.

One example is AI-generated content. A user could ask an agent to create an event poster, and the agent could access a compatible image-generation service, pay for the request and return the generated image.

The difference is in the number of steps the user has to manage.

Normally, using a new digital service might require visiting a website, creating an account, choosing a payment plan, purchasing credits and then completing the task. An agent capable of paying for services directly can handle more of that process in the background.

Kite AI is also exploring developer-focused use cases, particularly around the cost of tokens and compute used by coding agents, although Lee said during the interview that this capability was not yet officially live.

Could agents change how we pay for software?

Agentic payments could support a more granular model for purchasing digital services.

Many software products today are sold through monthly subscriptions or prepaid credits. That model makes sense when users repeatedly interact with the same application, but it can be inefficient when someone needs a service only once.

Lee used image generation as an example. If a user needs only one poster, there may be little reason to subscribe to another platform. An agent could instead purchase that individual generation and pay only for the completed task.

This creates a model closer to pay per job: one image generation, one API request or one specific service when it is required.

For an agent-driven internet, this model is particularly relevant because the agent itself may decide which service is best suited to complete a task. The user does not necessarily need a direct relationship with every underlying provider.

How do users stay in control?

Delegating transactions to AI introduces an obvious trust problem. Users need confidence that an agent will not exceed the authority it has been given.

Kite AI approach includes spending limits and scoped controls intended to define those boundaries. Rather than giving an agent unrestricted access to money, users can constrain what it is allowed to do.

These controls are important because trust in AI agents is still developing. Lee noted that consumers may be hesitant to delegate purchases when agent decisions can be difficult to explain or when models behave unpredictably.

Governance therefore becomes part of the infrastructure required for adoption, alongside identity and payment itself.

How does Kite fit into the wider agentic-payment ecosystem?

The agentic-payment market is still developing, and different companies are taking different approaches.

Lee pointed to protocols including UCP, ACP and AP2 as examples of an ecosystem where standards and philosophies have not yet converged. Some platforms are more closed and work with selected partners, while others are pursuing more open approaches.

Kite AI strategy is to support multiple systems rather than depend on a single protocol. Lee described the company as aiming to remain open and neutral, with support for different protocols as well as fiat and token-based payments.

The approach reflects how early the market remains. Agentic commerce may ultimately develop through several standards and ecosystems rather than one infrastructure provider controlling the entire transaction stack.

What could an agent-first economy look like?

The immediate use cases are relatively practical: purchasing office supplies, paying for an AI-generated image or accessing digital services without creating another account and subscription.

The longer-term implication is a change in how people interact with digital commerce.

Today, users typically choose the application, navigate its interface and complete the transaction themselves. An agent-first model moves more of those decisions and actions into the agent layer. Users specify what they want, while agents coordinate the services required to deliver it.

That model will require reliable ways to identify agents, define their authority, move money and govern their behavior. Payments are therefore only one component of the emerging agentic economy; the larger infrastructure challenge is enabling software to transact across digital services while preserving the user's control.

For Kite AI, the opportunity is to provide part of that underlying infrastructure as AI agents move from assisting with decisions to completing economic tasks on behalf of their users.

Watch the full episode on YouTube: https://youtu.be/p5kX1d8jSos?si=gDMdFqD3O_6iN8KK