What Happens When 10,000 Treasury Agents Make the Same Decision?

Treasury departments have traditionally been centered on helping the treasurer, who is typically a human, make better decisions about corporate cash.

The next generation of treasury technology, however, is proving capable of making those same decisions on its own. Artificial intelligence treasury agents today can decide when to release a payment, where to park excess cash, how much liquidity to hold, which payment rail to use, and eventually how to respond to changes in currencies, interest rates or counterparty risk.

Treasury leaders can give their agentic systems a mandate, such as to maximize yield, preserve liquidity or minimize transaction costs, and the agent can then execute against those objectives without waiting for a treasurer to click “approve.” For an individual chief financial officer, the economics can look compelling. But across the financial system, they become more complicated.

The Deutsche Bundesbank warned in its September Monthly Report published Monday (Sept. 21) that AI agents using homogeneous models, identical data sources or similar objectives could begin making the same decisions simultaneously. Individually sensible optimization could become collective “herd behavior,” potentially increasing liquidity requirements, overwhelming technical infrastructure and amplifying market movements.

That creates a new challenge for treasury teams to answer. What happens when thousands of companies deploy machines designed to reach the same financially rational conclusion at roughly the same time?

Read also: Agentic AI Could Make Net 30 Obsolete

Finance Targets the Gap Between Automating Treasury and Automating Judgment

Corporate treasury has been automating for years. Cash positioning, reconciliation, payment initiation, forecasting and liquidity management increasingly run through software. Agentic treasury changes the location of the decision, something more fundamental.

The PYMNTS Intelligence report “How Agentic AI Went From Zero to CFO Test Runs in 90 Days” found in September 2025 that nearly 7% of enterprise CFOs in the United States had deployed agentic AI in live finance workflows, while an additional 5% were running pilots.

The Bundesbank said in the report that liquidity management is one of the clearest institutional applications. Agents could balance the cost of holding liquidity against faster settlement, prioritize urgent payments, and incorporate expected inflows and outflows into those decisions. Eventually, they could retrieve information from internal and external systems faster than human managers or traditional deterministic treasury systems.

Instead of just showing a treasurer that $20 million is sitting idle, an agent could determine how much of it is actually available, compare investment options, maintain a required liquidity buffer and execute the transaction. Instead of flagging an upcoming foreign exchange exposure, it could determine whether and when to hedge it.

That is where AI’s value proposition becomes most powerful, and where its risk profile changes. The important distinction is between automating a process and automating a response to changing conditions. Traditional treasury automation generally follows predetermined rules. Agentic systems can interpret conditions and decide what action best advances an objective.

See also: AI Agents Push CFOs to Rethink Business Payments

Liquidity Optimization at Scale Creates Liquidity Problems Locally

Financial markets have always contained herd behavior. Humans watch one another, banks react to the same economic data and quantitative trading strategies can amplify movements. Agentic treasury could introduce the mechanism of thousands of autonomous systems independently deriving similar actions from similar inputs.

The Bundesbank’s concern, per the report, is that homogeneous models, shared data sources and comparable target functions can create coordinated behavior even when the underlying companies are different. Individual efficiency gains can therefore become procyclical systemic risks. Speed also magnifies the issue. A human treasury organization may require meetings, approvals and phone calls before moving hundreds of millions of dollars. Software does not. If agents can interpret a signal, select an action and execute through real-time financial infrastructure, reaction times potentially collapse.

Enterprise AI is likely to develop around a small collection of foundation models, cloud providers, protocols, payment networks and software platforms. The Bundesbank already sees signs of concentration around a limited number of protocols, interfaces and platforms, according to the report. The Bundesbank specifically pointed to potential “gridlock” and “free-riding effects,” in which participants delay outgoing payments while waiting for incoming liquidity. Systems optimized independently at the institutional level may fail to account for the consequences they collectively impose on financial infrastructure.

Read also: Agentic B2B Is Here. Are Your Contracts and Invoices Ready?

Still, the PYMNTS Intelligence report “Tech on Tech: How the Technology Sector Is Powering Agentic AI Adoption” found in August 2025 a widening agentic readiness gap between tech companies and firms in goods and services, with 75% of tech firms reporting they were extremely familiar with agentic AI, versus 33% of goods firms and 38% of services firms.

The Bundesbank pointed in the report toward clear mandates, reliable identity and authorization, logging, intervention mechanisms and human escalation for unusual circumstances. Large transactions could require explicit checkpoints. An agent’s permitted autonomy could also vary according to the decision being made.

In other words, the next generation of treasury controls may need to govern not only what a machine is allowed to do, but when it should stop optimizing.

For all PYMNTS B2B and AI coverage, subscribe to the daily B2B and AI newsletters.