Who is managing funds for the agent?

Author: Prathik Desai Translation: Shan Ouba, Golden Finance

In March this year, OpenAI shut down a customer AI shopping feature. The feature was launched for five months, with only 30 Shopify merchants onboard. The problem was not with the underlying payment infrastructure, but with the complete lack of supporting rules: what products can the AI purchase, who collects sales tax, how to identify fraudulent transactions, who is responsible for after-sales returns—these processes all lack clear standards.

Configuring wallets for AI agents, building payment channels is not difficult, but ensuring individuals and businesses can confidently entrust AI agents to manage funds, while also achieving regulatory compliance, is challenging. Only through a programmable rule system can a trustworthy usage environment be built. The regulatory rule layer's vacuum has already become a new opportunity in the AI agent economy.

Last year, AI agents completed transactions totaling $73 million, with a cumulative 176 million transactions. This scale may seem insignificant now, but McKinsey predicts that by 2030, AI agents will handle $3 to $5 trillion in global consumer transactions.

Companies vying in this sector are competing for influence over the regulatory rule layer. This layer includes core capabilities such as spending limits, identity verification, rule enforcement, and determines which AI agents can access funds.

This article will outline the current players building financial systems for AI agents and analyze the value that can be gained by leading in this field.

Core Logic of Multi-layered Layout

Profit margins for AI agent payments are very thin. Over the past year, the average transaction amount was only $0.31.

Let's do a quick calculation: a $0.31 transaction, after passing through multiple institutions, leaves little profit. According to Stripe’s standard fee rate (2.9% + $0.30 fixed fee), the merchant ends up with less than a tenth of a cent; Visa’s interchange fee takes another third. In contrast, stablecoin payment channels on crypto layer-2 networks handle similar transactions for just $0.0001.

This cost difference gives crypto solutions a natural advantage at the transaction settlement layer.

Currently, the infrastructure for transaction settlement is largely mature. Of the 176 million AI agent transactions last year, most were completed via Coinbase’s x402 protocol, and now about 3,900 merchants support AI agent payments. Stripe and Tempo jointly launched a competing protocol, Machine Payment Protocol (MPP), in March this year, supporting over 100 services. Meanwhile, Google, Visa, and Mastercard have also launched AI agent payment products. In just one year, five major payment systems are competing simultaneously.

But relying solely on small $0.31 transactions cannot generate high profits. The industry’s true value lies in capital accumulation benefits and payment rule regulation.

Last week, we analyzed that companies can earn by controlling wallets that store stablecoin assets for AI agents, but capital accumulation is only one source of value. The more significant potential lies in a regulatory rule system that governs fund usage.

This set of rules covers spending limits, AI agent identity verification, policy implementation, transaction auditing, and responsibility for transaction failures. Currently, this field remains a blue ocean.

In April this year, American Express launched AI agent shopping protection insurance, specifically compensating for losses caused by AI agent mis-spending. This also indirectly confirms that the current AI agent regulatory system is still imperfect. In this market, which is expected to surpass trillions of dollars within five years, filling regulatory gaps can unlock enormous commercial value.

This is also the core reason why traditional giants are rushing to develop the regulatory rule layer.

As for where to build this rule system—banks, developer APIs, or even digital wallets—are all potential options.

Building a Regulatory System Centered on Wallets

Every expenditure by an AI agent must go through a wallet, making wallets the best checkpoint for implementing spending limits, identity verification, and manual approval. Controlling the wallet is equivalent to holding regulatory authority. Payment infrastructure giant Stripe has long recognized this.

In June 2025, Stripe acquired Privy, a company focused on embedded wallet development. Through this, Stripe gained access to over 75 million wallets across more than a thousand development teams. Before any fund transfer, rules regarding spending limits, manual review, and other requirements can be enforced at this critical node.

Stripe also built a complete AI agent payment tech stack: acquiring the Bridge platform for stablecoin dispatch and fiat conversion; co-developing the public chain Tempo for payment scenarios with Paradigm; and launching the Machine Payment Protocol (MPP), establishing universal standards for initiating, authorizing, and settling AI agent payments.

Today, Stripe’s financial system for AI agents supports balance inquiries, bill payments, fund storage, virtual card issuance, and transfers. AI agents can automatically execute routine payments; if an operation exceeds preset rules, it is automatically handed over for manual review. Its fund accounts are built on Privy’s non-custodial wallets, covering over 150 countries and regions.

Even Amazon, when opening AI payment permissions, chose Privy and Coinbase wallets instead of traditional financial institutions like banks or card networks. The reason is that wallet services, despite being only five years old, are ideal risk control checkpoints—they can be flexibly set with manual intervention points to implement various regulatory checks.

Research firm Keyrock pointed out in the report “Who Pays for AI Agents” that the ultimate form of AI agent business will tend toward a compromise: AI agents will have high autonomous permissions but be bounded by encrypted rules, with humans able to audit and revoke their operations at any time.

Privy plays the role of defining rule boundaries within the Stripe ecosystem.

Privy has designed two operational modes for AI agent wallets: one where the AI agent has full control over the wallet, automatically executing transactions within rule limits without manual approval—suitable for autonomous bots and asset management; the other where the user retains ownership of the wallet, granting only limited signing permissions to the AI, which can be revoked at any time.

Stripe’s MPP protocol also adopts similar regulatory logic.

MPP’s session authorization feature is suitable for high-frequency AI agent transactions: users pre-set spending limits, allowing the AI to perform multiple on-chain payments within the limit without individual authorization requests. The protocol also supports fine-grained billing for large language model inference services and per-transaction data interface charges.

This granular control capability is beyond what traditional card networks can achieve.

Full-stack Vertical Competition

Currently, Coinbase’s x402 protocol leads in AI agent payments, but Privy, backed by Stripe, has formed a unique channel barrier.

Data shows about 3,900 merchants support AI agent payments via Coinbase, while Stripe’s partner merchants number nearly a thousand times more. Privy has stated that once Stripe’s full merchant base adopts machine payment functions, leveraging its wallet ecosystem, the scale of AI agent commerce will explode. Merchants also do not need to build dedicated crypto infrastructure.

The competition between Stripe and Coinbase is intensifying, with other traditional giants also actively expanding along the industry chain.

Keyrock divides the AI agent payment system into six levels: settlement layer, wallet layer, routing layer, protocol layer, regulation layer, and application layer, with a total of 179 related projects across the entire sector.

Among these, Coinbase and Stripe cover five of the six layers; Circle covers four; Google only two; Visa just one.

In the past year, traditional giants have invested over $8 billion to fill industry gaps: First Capital acquired the native AI platform Brex for $5.15 billion; Mastercard bought BVNK for $1.8 billion. These acquisitions mainly target the wallet and AI software layers: Stripe acquired Privy, Fireblocks bought Dynamic, Arbitrum acquired ZeroDev. All these moves point to a clear trend: payment infrastructure companies are actively developing independent wallet services.

This series of actions sends a clear signal: transaction settlement services are becoming homogenized and profit margins thin, while rule-related businesses such as permission control, quota allocation, and responsibility delineation are where the real value lies.

Vertical integration of the industry chain can also generate compound effects.

Controlling the risk management checkpoints allows entities to set fund usage rules, earn from capital deposits, select partner merchants and AI applications, and charge service fees. The channel barrier built by Stripe and Privy is a typical example.

Coinbase’s approach is similar: each x402 protocol payment increases demand for USDC on its Layer 2 network Base, generating capital accumulation benefits. These profits are reinvested into its AI tool suite AgentKit, which includes session limits, per-transaction caps, whitelist contracts, and other controls. The more AI agents connect to AgentKit, the higher the transaction volume on x402, creating a positive feedback loop across business segments.

The investments of traditional giants go far beyond this.

Coinbase Ventures has invested in three leading independent regulation startups: Catena Labs, Skyfire, and Payman. Catena was founded by Circle co-founders; Circle also invested in Skyfire, with VC firm a16z backing both. Visa invested in Payman and partnered with Skyfire.

Leading companies building payment settlement infrastructure are now collectively strengthening the regulation layer. Their strategic approach is clear: if regulatory functions are embedded into existing infrastructure (like Privy’s dual-mode architecture), current giants can maximize revenue; even if the regulation layer develops into an independent sector, they can share the benefits through investments.

The Business Value of Controlling the Regulation Layer

Handling only payment transactions has never been the most valuable part of the industry chain. Once the underlying financial channels become homogenized, profit shifts to transaction approval and constraint-related services.

Many industries have experienced similar evolutions.

In the early days of internet proliferation, wired network infrastructure gradually became homogeneous, causing service providers to lose differentiation. Telecom companies then expanded vertically: India’s Jio and Airtel bundled TV channels, streaming memberships, unlimited calls, set-top boxes, and routers into broadband packages; AT&T spent $85 billion acquiring Time Warner, integrating HBO, Warner Bros., CNN, and other premium content with its vast distribution channels, competing with Netflix and Amazon.

When broadband infrastructure no longer offers profit advantages, value shifts to content services, customer relationships, and comprehensive benefit packages.

The crypto industry has seen similar developments.

Transaction settlement was originally a core protocol function; Ethereum is recognized as a universal settlement blockchain. After Coinbase launched Layer 2 network Base, it earns from on-chain transaction fees, with the network’s validator annual revenue around $60 million.

Companies in the AI agent payment space are also following this path.

In our article “AI Agent Capital Yield,” we mentioned that controlling stablecoin assets during AI agent transactions can open new revenue streams—this is the core value of wallets. The regulation layer can generate even greater or additional revenue.

Visa’s annual payment volume reaches $14.2 trillion, with an overall fee rate of 0.28%. This income includes not only transaction processing fees but also hidden costs for risk control, fraud prevention, dispute resolution, and network rule maintenance.

Applying this logic, the enormous potential of the regulation layer in AI agent business becomes evident. According to McKinsey’s forecast, AI agent transaction volume will reach $3 trillion by 2030. Even with a modest 0.1% regulation fee (about a third of Visa’s rate), annual revenue could hit $3 billion. For comparison, Coinbase’s total subscription and service revenue in 2025 is about $2.8 billion. This means that just the regulation services for AI agent transactions could match Coinbase’s current total revenue from staking, custody, and membership services.

A company that simultaneously develops wallets, settlement, and regulation layers can earn triple benefits: interest income from idle AI agent funds, transaction settlement fees, and compliance service fees.

This also indicates that full vertical integration across the industry chain will become the only viable business model for maintaining competitiveness in the AI agent era.