You can build an AI system that can analyze cash positions across 40 subsidiaries, predict FX movements, optimize working capital in real time, and recommend the perfect moment to execute a currency hedge. And then it stops, because you can’t give it a bank account.
AI can think about money. But the financial system still assumes a human is on the other end of every transaction.
But really, the challenge is broader than permissions. The opportunity set has exploded while the infrastructure remains stuck.
Faced with global markets, dozens of asset classes, and FX spreads that exist for seconds across fragmented venues, we can build AI to see all of it, to analyze positions across subsidiaries, track real-time rates across counterparties, and model the optimal move. The real reason it can’t act is because acting would require coordinating across systems that refuse to talk to each other, during windows when half of them are offline.
Fedwire operates weekdays, from 9 p.m. the prior calendar day to 7 p.m. Eastern Time, and is closed during weekends and federal holidays. The global financial system shuts down for roughly 40% of every week. That’s fine for us, because humans need to sleep. But it’s absurd when the systems managing your treasury don’t.
This is the infrastructure gap: intelligence that operates globally and continuously, connected to plumbing that’s local and intermittent.
To compete in this environment, you need interoperability across fragmented systems, and the ability to act continuously, not just during banking hours, and not just when a human clicks approve.
The Permission Problem
The traditional financial system is built for humans. Bank accounts require a human identity and a mountain of “Know-Your-Customer” checks. Credit cards, wire transfers, PayPal, they all assume a human is on the other end. AI systems are left out in the cold, no matter how smart they are.
But requiring human approval for every transaction defeats the purpose of autonomous systems. The breakthrough is a new model for authority: giving agents the ability to act within defined boundaries, without giving them the keys to everything.
Many pieces are already in place. Account abstraction can provide hierarchical access controls where enterprises can set granular permissions, defining who can approve transactions, and at what thresholds. Smart wallets can be programmed to reject transactions that violate certain criteria.
Smart wallets have become particularly important for enterprise applications and AI agent platforms because enterprises need custody solutions that go beyond a single private key. Features like signer rotation allow companies to change controlling keys when employees leave or keys are compromised without changing the wallet address or moving assets.
This solves the permission problem by making trust unnecessary. The constraints are predefined.
The Dead Time Problem
Wire transfers process through interbank clearing and settlement networks that settle USD payments within a day and international payments within two days. Miss the cutoff window, and your transfer rolls to the next business day. Hit a weekend, and it waits until Monday.
For most of the 20th century, this was fine. Humans initiated transactions during business hours, and the system processed them overnight. The rhythm of banking matched the rhythm of work.
But autonomous finance agents are already making their way into real-world treasury workflows. AI agents gather balances from banks and ERP systems, calculate daily positions, and factor in incoming and outgoing payments. They’re optimizing continuously, except when they hit infrastructure that isn’t.
When there’s idle cash sitting around, agents step in. They suggest investment options based on real-time market conditions and company risk profiles. But a human still has to approve and execute. This all changes as companies take advantage of better infrastructure that gives AI 24/7 autonomy.
When Machines Transact With Machines
The market for agent-to-agent transactions barely exists today. Within a few years, it will be enormous, and it will route around any infrastructure that requires human approval for every exchange. To take advantage of this, systems need to be integrated.
Treasury operations sprawl across ERP systems, banking portals, FX platforms, money market funds, and internal accounting tools. Each system has its own login, its own data format, its own API (if it has an API at all). Building a unified view requires either expensive middleware or armies of analysts copying data between spreadsheets.
Agentic AI is an emerging class of AI that can independently pursue goals, make decisions, and take actions with limited human input. The value unlock, first, is smarter analysis. Later, it’s connecting intelligence to action through infrastructure that agents can actually use.
This creates markets so efficient that humans can’t participate. HFT systems can already calculate market pricing ahead of official feeds. They create arbitrage opportunities that exist for only single-digit milliseconds. AI will take this further.
Call them machine-only payment windows: settlement periods where the only participants are agents transacting with other agents. The enterprises that capture this value will do so via AI connected to infrastructure that can execute at 3 a.m. on a Saturday, against a counterparty’s agent, settling in seconds. Everyone else will be waiting for Monday, watching the opportunity in the rearview mirror.
The intelligence layer is maturing fast. The execution layer (infrastructure that lets agents actually transact) is the gap that remains.
What Comes Next
Picture a treasury office where an AI avatar greets the CFO with a briefing of overnight currency swings, projected cash balances, and alerts about supplier default risk. The AI proposes mitigation strategies, complete with financial impacts and probabilities. The CFO asks follow-up questions, challenges assumptions, and ultimately greenlights one path. The agent then executes, documenting rationale for auditors.
That’s the near-term: AI as strategic partner, human as approver of significant decisions.
The further horizon is different. The future of corporate treasury points toward self-driving treasury functions that automate investment decisions, FX hedging, and working capital optimization. I believe we’ll see 20% of enterprise financial operations, ranging from accounts payable, treasury management, vendor payments, procurement, and liquidity optimization executed by autonomous agents within the next 18 months.
In the future machines will manage money. The competitive edge comes from leveraging the best infrastructure to do it.
