BloxFi Research
The Financial Infrastructure Gap in the AI Economy
AI agents can already pay. The harder question is who lets them, within what limits, and who answers for it.
Payments infrastructure · October 2026 · 6 min read
The trigger: software that starts to act
In April 2026 the IMF published a note on what happens to payments when software acts for people. Its starting point is that AI systems can now take an objective, break it into tasks and use digital services with limited human input. In payments, that moves the decision to pay from a person to an agent (IMF Note 2026/004, 24 April 2026).
We set out to test one question: is the financial infrastructure that agents need to manage and move money safely still underdeveloped? We did not assume the answer. The evidence supports a narrower finding than the headlines. The rails for moving money are further along than the rules for who may use them, and under what limits.
Built for people
Most payment systems assume an identifiable human. Someone opens an account, is verified once, and approves transactions one at a time, usually for a defined purchase or on a schedule.
Agents may need something different: to transact continuously, make many small payments, deal with other agents and stay inside a budget set in advance, without anyone approving each payment.
Fact: Chainalysis describes the friction directly. An agent that needs a data feed or a screening check has to stop for account set-up, API key management and a billing relationship (Chainalysis, 3 June 2026).
Interpretation: this looks like a design mismatch more than a technology gap. The account model works for a business with ten suppliers. It works poorly for software that meets a new service every minute.
Why existing systems struggle
The IMF names the central tension: AI behaviour is probabilistic, while payment authorisation and settlement have to be deterministic. It also lists traceability, opaque decision-making, correlated behaviour across many agents, cyber exposure and unresolved liability as risks (IMF, April 2026).
Our reading is that these turn into two operating costs.
Control. Approval on every payment removes the point of autonomy. No approval at all lets one bad instruction repeat at machine speed.
Accountability. Compliance programmes are built around a known customer. They have no settled answer to a simple question: which agent is this, and who stands behind it?
Testing the thesis: what the data shows
One live attempt to close the gap is x402, a protocol developed by Coinbase. A server answers a request with a payment specification, the agent pays in stablecoin and resubmits the request with a receipt.
Fact: Chainalysis counted well over 100 million cumulative x402 transactions on Base through the first quarter of 2026, from near zero in mid-2025.
Evidence: much of that growth came from meme coin farming, and activity moderated in early 2026. The more telling change is in size (Figure 1). Payments of $1 or more rose from 49% of transfer value in early 2025 to 95% in early 2026, while payments between 10 cents and $1 fell from 46% to 4%.
Interpretation: demand is real but early, and partly speculative. Chainalysis itself concludes that use is moving beyond micro-payment experiments and that mass adoption remains distant. The data covers one protocol on one chain, so it shows direction, not the size of the market.
Payments of $1 or more went from 49% to 95% of x402 transfer value
Share of transfer value by payment size, x402 agent payments on Base, early 2025 versus early 2026.
- Early 2025
- Early 2026
$1 and above
10¢ to $1
Under 10¢ (remainder)
Share of transfer value (%)
Source: Chainalysis, “Agentic Payments Cross the Threshold: Inside x402’s Path to Meaningful Adoption”, 3 June 2026. Under 10¢ is BloxFi’s remainder (100% minus the published shares).
The stack: seven layers between an agent and a settled payment
The IMF frames three layers: intent, authorisation and settlement. We split the middle layer into the parts an operator would have to build. Table 1 sets out the conceptual contrast with today’s systems, and Figure 2 maps the layers to the IMF’s. Both are our analytical framework, not empirical data.
Table 1: Human-designed versus agent-ready finance
| Traditional financial system | AI-native infrastructure | Stack layer |
|---|---|---|
| Human-controlled transactions | Agent-initiated transactions | Intent |
| User identity | Agent identity and KYA | Identity |
| Conventional accounts | Agent wallets | Wallet |
| Manual authorisation | Policy-based authorisation | Policy and budget |
| Human monitoring | Automated monitoring | Compliance and monitoring |
| Fixed payment flows | Programmatic, metered payments | Payment rail and settlement |
Four of the seven layers govern what an agent may do
The agent payment stack, mapped to the IMF’s three layers. Highlighted layers decide whether an agent may pay at all.
Sources: IMF Note 2026/004, “How Agentic AI Will Reshape Payments”, 24 April 2026 (three layers); BloxFi analysis (seven-layer split).
Four layers decide whether an agent may pay at all: who it is, where it pays from, what it may spend and what is checked on the way.
Where BloxFi fits
BloxFi is a technology and connectivity platform. It connects businesses to a multi-provider network of liquidity, payment and settlement infrastructure through one API. It is not a bank, an exchange or a custodian, and it does not move its own money. Clients onboard with the underlying providers themselves.
Through that one API, BloxFi offers:
Smart routing across providers.
Cross-border payments with local currency in and local currency out.
On-ramp (fiat to crypto) and off-ramp (crypto to fiat).
Named accounts.
Compliance and on-chain checks.
30+ local currencies.
In our reading, several of these sit where the argument above puts the pressure. Named accounts and compliance and on-chain checks bear on identity, accountability and monitoring. Smart routing, local payments in and out, and the on-ramp and off-ramp bear on the payment rail and settlement. This describes what BloxFi offers today. It is not evidence that the problem of agent payments is solved, and we cite no usage figures. Availability, rails and limits are confirmed per client following KYB and compliance review.
What we are watching
Signals that would strengthen our view:
Standard-setters or regulators publishing identity and authorisation frameworks for agents.
Agent payment volume holding up without speculative campaigns.
Payment providers offering policy and budget controls as standard.
Signals that would weaken it:
Agent payments staying small and speculative.
Existing card and account rails adding delegated-spend controls that work at agent scale.
Conclusion
Stablecoin rails already let a machine pay. What remains unsettled is permission. If AI agents become economic actors, financial infrastructure has to move from systems built to verify a person at the door towards systems that can govern machine-initiated activity as it happens: who the agent is, what it may do and who is accountable when it acts.
The test is simple. Can a person hand an agent a budget and a purpose, and still answer for every payment it makes?
Sources
IMF, “How Agentic AI Will Reshape Payments”, IMF Note 2026/004, 24 April 2026.
Chainalysis, “Agentic Payments Cross the Threshold: Inside x402’s Path to Meaningful Adoption”, 3 June 2026.