Gartner forecasts that AI agents will mediate nine in every ten B2B purchases by 2028, moving more than $15 trillion through AI agent exchanges. In finance and procurement, agents already match invoices, audit expenses, screen suppliers and clear routine payments with less human intervention.
Financial infrastructure was built around human workflows, business hours and separate systems for payments, trading, custody and compliance. An agent can make a decision in milliseconds. The infrastructure still has to determine whether that decision is authorised, where the money should move, how it settles and what happens when something goes wrong.
That is the real infrastructure challenge behind autonomous commerce. The agent may be the visible part of the system, but the financial rails underneath will determine how far it can actually scale. Authorisation is only one part of the problem. Settlement, custody and jurisdiction are the other three.
Authorisation has to travel with the transaction
An autonomous system needs more than a way to initiate a payment. It needs a clear mandate: who the agent is acting for, what it can spend, which counterparties it can transact with, and what happens when a transaction falls outside those parameters.
Google’s Agent Payments Protocol, now under the stewardship of the FIDO Alliance, includes mechanisms for transactions made under pre-authorised instructions. Visa and Mastercard are developing similar infrastructure around agent-specific credentials and controls.
The direction is clear: authorisation needs to travel with the transaction rather than sit with a person waiting to approve it. But permission only solves part of the problem. The financial system still has to execute and settle the transaction.
Settlement is where speed gets tested
Research from Visa and Artemis using live on-chain data shows agents already paying for resources at machine speed, often in amounts under one dollar that card rails cannot carry economically. Yet the infrastructure completing those transactions still depends on multiple systems and processes.
The instruction is instant. The payment may still pass through gateways, clearing systems, banks and other intermediaries before the value reaches its destination. A system built around cut-off times, manual reconciliation and regional operating hours will eventually become the bottleneck, regardless of how capable the agent becomes.
Take a simple example: a company in London sends money to Brazil at five o’clock on a Friday. Nobody touches it until a banker logs in after breakfast on Monday, or on Tuesday if there is a bank holiday. On the way, it passes through the sending bank, a correspondent bank and the receiving bank, and each adds a fee or its own FX rate. An agent can send that instruction at once, but on those rails the money still waits for Monday.
This is where stablecoins and other programmable forms of money become relevant. They create new ways to move value across borders and currencies, but the underlying infrastructure still needs to connect liquidity, execution and settlement rather than treating each as separate workflows.
The Bank for International Settlements’ 2026 Annual Economic Report highlights a related problem. The same stablecoin issued across different blockchains can effectively behave as a distinct asset, and moving value between networks can be difficult or costly. For autonomous commerce operating across networks and currencies, that fragmentation becomes a practical infrastructure problem.
Custody cannot sit outside the transaction
The same principle applies to the assets themselves. An agent may know what to buy and have permission to spend, but someone still needs to hold the underlying value and make it available when the transaction happens.
If execution, payment and custody sit across disconnected systems, every transaction creates another handoff and another point where automation can stop. At machine speed, those operational gaps become more consequential.
The infrastructure therefore has to account for the full movement of value, from instruction through execution and settlement to custody. The objective is simple: the money should be able to move through the system without the architecture becoming the constraint.
Jurisdiction becomes part of the architecture
Autonomous commerce will move across borders. The infrastructure supporting it has to account for those borders from the start. A stablecoin cleared for use under one jurisdiction’s framework may need separate authorisation elsewhere before an agent can transact with it the same way.
Cross-border reporting regimes, from the EU’s DAC8 to the OECD’s Crypto-Asset Reporting Framework, are making compliance an increasingly important part of the transaction flow.
In practice, that means jurisdiction is design work, not paperwork. The rules governing payments, custody, reporting and asset management shape which counterparties and markets a financial platform can reach, and how those relationships need to be structured. A licensing decision made at the start can determine how easily a platform expands into new markets later.
The infrastructure has to move with the agent
The friction between legacy clearing chains and fragmented regulatory regimes is a permanent feature of global commerce, not a temporary bug to be patched. Institutions building for autonomous agents have to treat it as a design constraint from day one. That means designing the links between execution, custody and settlement up front, rather than stitching systems together as an afterthought.
The firms that get this right will not be the ones that ship the most capable agent. They will be the ones whose settlement, custody and compliance infrastructure was built for machine-speed transacting from the start, rather than retrofitted onto rails designed for business hours and paper trails. Jurisdiction, licensing and custody design are not back-office details in this world. They are the architecture that determines whether an autonomous transaction actually finishes.
The next phase of autonomous commerce will be decided below the interface. Agents will become better at making decisions. Protocols will make those decisions easier to authorise. Neither matters unless the money can follow, and that requires financial systems designed around how money actually moves, rather than autonomous agents added on top of workflows built for humans.
The agent does not care whether a payment is moving across three systems, two jurisdictions or different settlement windows. It only knows that it has been told to transact. Making sure the money can move is the job of the infrastructure.
